feat: provider modules — ProviderProfile ABC, 30 providers, fetch_models, transport single-path

feat: provider modules — ProviderProfile ABC, 29 providers, fetch_models, transport single-path

Introduces providers/ as the single source of truth for every inference
provider. All 29 providers declared with correct data cross-checked against
auth.py, runtime_provider.py and auxiliary_client.py.

Rebased onto main (30307a980). Incorporates post-salvage fixes from
56724147e (gmi aux model google/gemini-3.1-flash-lite-preview, already set in providers/gmi.py).

.gitignore

@@ -69,3 +69,4 @@ mini-swe-agent/
 .nix-stamps/
 result
 website/static/api/skills-index.json
+models-dev-upstream/

acp_adapter/copilot_client.py

@@ -0,0 +1,632 @@
+"""OpenAI-compatible shim that forwards Hermes requests to `copilot --acp`.
+
+This adapter lets Hermes treat the GitHub Copilot ACP server as a chat-style
+backend. Each request starts a short-lived ACP session, sends the formatted
+conversation as a single prompt, collects text chunks, and converts the result
+back into the minimal shape Hermes expects from an OpenAI client.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import queue
+import re
+import shlex
+import subprocess
+import threading
+import time
+from collections import deque
+from pathlib import Path
+from types import SimpleNamespace
+from typing import Any
+
+from agent.file_safety import get_read_block_error, is_write_denied
+from agent.redact import redact_sensitive_text
+
+ACP_MARKER_BASE_URL = "acp://copilot"
+_DEFAULT_TIMEOUT_SECONDS = 900.0
+
+_TOOL_CALL_BLOCK_RE = re.compile(r"<tool_call>\s*(\{.*?\})\s*</tool_call>", re.DOTALL)
+_TOOL_CALL_JSON_RE = re.compile(
+    r"\{\s*\"id\"\s*:\s*\"[^\"]+\"\s*,\s*\"type\"\s*:\s*\"function\"\s*,\s*\"function\"\s*:\s*\{.*?\}\s*\}",
+    re.DOTALL,
+)
+
+
+def _resolve_command() -> str:
+    return (
+        os.getenv("HERMES_COPILOT_ACP_COMMAND", "").strip()
+        or os.getenv("COPILOT_CLI_PATH", "").strip()
+        or "copilot"
+    )
+
+
+def _resolve_args() -> list[str]:
+    raw = os.getenv("HERMES_COPILOT_ACP_ARGS", "").strip()
+    if not raw:
+        return ["--acp", "--stdio"]
+    return shlex.split(raw)
+
+
+def _jsonrpc_error(message_id: Any, code: int, message: str) -> dict[str, Any]:
+    return {
+        "jsonrpc": "2.0",
+        "id": message_id,
+        "error": {
+            "code": code,
+            "message": message,
+        },
+    }
+
+
+def _permission_denied(message_id: Any) -> dict[str, Any]:
+    return {
+        "jsonrpc": "2.0",
+        "id": message_id,
+        "result": {
+            "outcome": {
+                "outcome": "cancelled",
+            }
+        },
+    }
+
+
+def _format_messages_as_prompt(
+    messages: list[dict[str, Any]],
+    model: str | None = None,
+    tools: list[dict[str, Any]] | None = None,
+    tool_choice: Any = None,
+) -> str:
+    sections: list[str] = [
+        "You are being used as the active ACP agent backend for Hermes.",
+        "Use ACP capabilities to complete tasks.",
+        "IMPORTANT: If you take an action with a tool, you MUST output tool calls using <tool_call>{...}</tool_call> blocks with JSON exactly in OpenAI function-call shape.",
+        "If no tool is needed, answer normally.",
+    ]
+    if model:
+        sections.append(f"Hermes requested model hint: {model}")
+
+    if isinstance(tools, list) and tools:
+        tool_specs: list[dict[str, Any]] = []
+        for t in tools:
+            if not isinstance(t, dict):
+                continue
+            fn = t.get("function") or {}
+            if not isinstance(fn, dict):
+                continue
+            name = fn.get("name")
+            if not isinstance(name, str) or not name.strip():
+                continue
+            tool_specs.append(
+                {
+                    "name": name.strip(),
+                    "description": fn.get("description", ""),
+                    "parameters": fn.get("parameters", {}),
+                }
+            )
+        if tool_specs:
+            sections.append(
+                "Available tools (OpenAI function schema). "
+                "When using a tool, emit ONLY <tool_call>{...}</tool_call> with one JSON object "
+                "containing id/type/function{name,arguments}. arguments must be a JSON string.\n"
+                + json.dumps(tool_specs, ensure_ascii=False)
+            )
+
+    if tool_choice is not None:
+        sections.append(
+            f"Tool choice hint: {json.dumps(tool_choice, ensure_ascii=False)}"
+        )
+
+    transcript: list[str] = []
+    for message in messages:
+        if not isinstance(message, dict):
+            continue
+        role = str(message.get("role") or "unknown").strip().lower()
+        if role == "tool":
+            role = "tool"
+        elif role not in {"system", "user", "assistant"}:
+            role = "context"
+
+        content = message.get("content")
+        rendered = _render_message_content(content)
+        if not rendered:
+            continue
+
+        label = {
+            "system": "System",
+            "user": "User",
+            "assistant": "Assistant",
+            "tool": "Tool",
+            "context": "Context",
+        }.get(role, role.title())
+        transcript.append(f"{label}:\n{rendered}")
+
+    if transcript:
+        sections.append("Conversation transcript:\n\n" + "\n\n".join(transcript))
+
+    sections.append("Continue the conversation from the latest user request.")
+    return "\n\n".join(
+        section.strip() for section in sections if section and section.strip()
+    )
+
+
+def _render_message_content(content: Any) -> str:
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content.strip()
+    if isinstance(content, dict):
+        if "text" in content:
+            return str(content.get("text") or "").strip()
+        if "content" in content and isinstance(content.get("content"), str):
+            return str(content.get("content") or "").strip()
+        return json.dumps(content, ensure_ascii=True)
+    if isinstance(content, list):
+        parts: list[str] = []
+        for item in content:
+            if isinstance(item, str):
+                parts.append(item)
+            elif isinstance(item, dict):
+                text = item.get("text")
+                if isinstance(text, str) and text.strip():
+                    parts.append(text.strip())
+        return "\n".join(parts).strip()
+    return str(content).strip()
+
+
+def _extract_tool_calls_from_text(text: str) -> tuple[list[SimpleNamespace], str]:
+    if not isinstance(text, str) or not text.strip():
+        return [], ""
+
+    extracted: list[SimpleNamespace] = []
+    consumed_spans: list[tuple[int, int]] = []
+
+    def _try_add_tool_call(raw_json: str) -> None:
+        try:
+            obj = json.loads(raw_json)
+        except Exception:
+            return
+        if not isinstance(obj, dict):
+            return
+        fn = obj.get("function")
+        if not isinstance(fn, dict):
+            return
+        fn_name = fn.get("name")
+        if not isinstance(fn_name, str) or not fn_name.strip():
+            return
+        fn_args = fn.get("arguments", "{}")
+        if not isinstance(fn_args, str):
+            fn_args = json.dumps(fn_args, ensure_ascii=False)
+        call_id = obj.get("id")
+        if not isinstance(call_id, str) or not call_id.strip():
+            call_id = f"acp_call_{len(extracted) + 1}"
+
+        extracted.append(
+            SimpleNamespace(
+                id=call_id,
+                call_id=call_id,
+                response_item_id=None,
+                type="function",
+                function=SimpleNamespace(name=fn_name.strip(), arguments=fn_args),
+            )
+        )
+
+    for m in _TOOL_CALL_BLOCK_RE.finditer(text):
+        raw = m.group(1)
+        _try_add_tool_call(raw)
+        consumed_spans.append((m.start(), m.end()))
+
+    # Only try bare-JSON fallback when no XML blocks were found.
+    if not extracted:
+        for m in _TOOL_CALL_JSON_RE.finditer(text):
+            raw = m.group(0)
+            _try_add_tool_call(raw)
+            consumed_spans.append((m.start(), m.end()))
+
+    if not consumed_spans:
+        return extracted, text.strip()
+
+    consumed_spans.sort()
+    merged: list[tuple[int, int]] = []
+    for start, end in consumed_spans:
+        if not merged or start > merged[-1][1]:
+            merged.append((start, end))
+        else:
+            merged[-1] = (merged[-1][0], max(merged[-1][1], end))
+
+    parts: list[str] = []
+    cursor = 0
+    for start, end in merged:
+        if cursor < start:
+            parts.append(text[cursor:start])
+        cursor = max(cursor, end)
+    if cursor < len(text):
+        parts.append(text[cursor:])
+
+    cleaned = "\n".join(p.strip() for p in parts if p and p.strip()).strip()
+    return extracted, cleaned
+
+
+def _ensure_path_within_cwd(path_text: str, cwd: str) -> Path:
+    candidate = Path(path_text)
+    if not candidate.is_absolute():
+        raise PermissionError("ACP file-system paths must be absolute.")
+    resolved = candidate.resolve()
+    root = Path(cwd).resolve()
+    try:
+        resolved.relative_to(root)
+    except ValueError as exc:
+        raise PermissionError(
+            f"Path '{resolved}' is outside the session cwd '{root}'."
+        ) from exc
+    return resolved
+
+
+class _ACPChatCompletions:
+    def __init__(self, client: CopilotACPClient):
+        self._client = client
+
+    def create(self, **kwargs: Any) -> Any:
+        return self._client._create_chat_completion(**kwargs)
+
+
+class _ACPChatNamespace:
+    def __init__(self, client: CopilotACPClient):
+        self.completions = _ACPChatCompletions(client)
+
+
+class CopilotACPClient:
+    """Minimal OpenAI-client-compatible facade for Copilot ACP."""
+
+    def __init__(
+        self,
+        *,
+        api_key: str | None = None,
+        base_url: str | None = None,
+        default_headers: dict[str, str] | None = None,
+        acp_command: str | None = None,
+        acp_args: list[str] | None = None,
+        acp_cwd: str | None = None,
+        command: str | None = None,
+        args: list[str] | None = None,
+        **_: Any,
+    ):
+        self.api_key = api_key or "copilot-acp"
+        self.base_url = base_url or ACP_MARKER_BASE_URL
+        self._default_headers = dict(default_headers or {})
+        self._acp_command = acp_command or command or _resolve_command()
+        self._acp_args = list(acp_args or args or _resolve_args())
+        self._acp_cwd = str(Path(acp_cwd or os.getcwd()).resolve())
+        self.chat = _ACPChatNamespace(self)
+        self.is_closed = False
+        self._active_process: subprocess.Popen[str] | None = None
+        self._active_process_lock = threading.Lock()
+
+    def close(self) -> None:
+        proc: subprocess.Popen[str] | None
+        with self._active_process_lock:
+            proc = self._active_process
+            self._active_process = None
+        self.is_closed = True
+        if proc is None:
+            return
+        try:
+            proc.terminate()
+            proc.wait(timeout=2)
+        except Exception:
+            try:
+                proc.kill()
+            except Exception:
+                pass
+
+    def _create_chat_completion(
+        self,
+        *,
+        model: str | None = None,
+        messages: list[dict[str, Any]] | None = None,
+        timeout: float | None = None,
+        tools: list[dict[str, Any]] | None = None,
+        tool_choice: Any = None,
+        **_: Any,
+    ) -> Any:
+        prompt_text = _format_messages_as_prompt(
+            messages or [],
+            model=model,
+            tools=tools,
+            tool_choice=tool_choice,
+        )
+        # Normalise timeout: run_agent.py may pass an httpx.Timeout object
+        # (used natively by the OpenAI SDK) rather than a plain float.
+        if timeout is None:
+            _effective_timeout = _DEFAULT_TIMEOUT_SECONDS
+        elif isinstance(timeout, (int, float)):
+            _effective_timeout = float(timeout)
+        else:
+            # httpx.Timeout or similar — pick the largest component so the
+            # subprocess has enough wall-clock time for the full response.
+            _candidates = [
+                getattr(timeout, attr, None)
+                for attr in ("read", "write", "connect", "pool", "timeout")
+            ]
+            _numeric = [float(v) for v in _candidates if isinstance(v, (int, float))]
+            _effective_timeout = max(_numeric) if _numeric else _DEFAULT_TIMEOUT_SECONDS
+
+        response_text, reasoning_text = self._run_prompt(
+            prompt_text,
+            timeout_seconds=_effective_timeout,
+        )
+
+        tool_calls, cleaned_text = _extract_tool_calls_from_text(response_text)
+
+        usage = SimpleNamespace(
+            prompt_tokens=0,
+            completion_tokens=0,
+            total_tokens=0,
+            prompt_tokens_details=SimpleNamespace(cached_tokens=0),
+        )
+        assistant_message = SimpleNamespace(
+            content=cleaned_text,
+            tool_calls=tool_calls,
+            reasoning=reasoning_text or None,
+            reasoning_content=reasoning_text or None,
+            reasoning_details=None,
+        )
+        finish_reason = "tool_calls" if tool_calls else "stop"
+        choice = SimpleNamespace(message=assistant_message, finish_reason=finish_reason)
+        return SimpleNamespace(
+            choices=[choice],
+            usage=usage,
+            model=model or "copilot-acp",
+        )
+
+    def _run_prompt(
+        self, prompt_text: str, *, timeout_seconds: float
+    ) -> tuple[str, str]:
+        try:
+            proc = subprocess.Popen(
+                [self._acp_command] + self._acp_args,
+                stdin=subprocess.PIPE,
+                stdout=subprocess.PIPE,
+                stderr=subprocess.PIPE,
+                text=True,
+                bufsize=1,
+                cwd=self._acp_cwd,
+            )
+        except FileNotFoundError as exc:
+            raise RuntimeError(
+                f"Could not start Copilot ACP command '{self._acp_command}'. "
+                "Install GitHub Copilot CLI or set HERMES_COPILOT_ACP_COMMAND/COPILOT_CLI_PATH."
+            ) from exc
+
+        if proc.stdin is None or proc.stdout is None:
+            proc.kill()
+            raise RuntimeError("Copilot ACP process did not expose stdin/stdout pipes.")
+
+        self.is_closed = False
+        with self._active_process_lock:
+            self._active_process = proc
+
+        inbox: queue.Queue[dict[str, Any]] = queue.Queue()
+        stderr_tail: deque[str] = deque(maxlen=40)
+
+        def _stdout_reader() -> None:
+            if proc.stdout is None:
+                return
+            for line in proc.stdout:
+                try:
+                    inbox.put(json.loads(line))
+                except Exception:
+                    inbox.put({"raw": line.rstrip("\n")})
+
+        def _stderr_reader() -> None:
+            if proc.stderr is None:
+                return
+            for line in proc.stderr:
+                stderr_tail.append(line.rstrip("\n"))
+
+        out_thread = threading.Thread(target=_stdout_reader, daemon=True)
+        err_thread = threading.Thread(target=_stderr_reader, daemon=True)
+        out_thread.start()
+        err_thread.start()
+
+        next_id = 0
+
+        def _request(
+            method: str,
+            params: dict[str, Any],
+            *,
+            text_parts: list[str] | None = None,
+            reasoning_parts: list[str] | None = None,
+        ) -> Any:
+            nonlocal next_id
+            next_id += 1
+            request_id = next_id
+            payload = {
+                "jsonrpc": "2.0",
+                "id": request_id,
+                "method": method,
+                "params": params,
+            }
+            assert proc.stdin is not None  # always set: Popen(stdin=PIPE)
+            proc.stdin.write(json.dumps(payload) + "\n")
+            proc.stdin.flush()
+
+            deadline = time.time() + timeout_seconds
+            while time.time() < deadline:
+                if proc.poll() is not None:
+                    break
+                try:
+                    msg = inbox.get(timeout=0.1)
+                except queue.Empty:
+                    continue
+
+                if self._handle_server_message(
+                    msg,
+                    process=proc,
+                    cwd=self._acp_cwd,
+                    text_parts=text_parts,
+                    reasoning_parts=reasoning_parts,
+                ):
+                    continue
+
+                if msg.get("id") != request_id:
+                    continue
+                if "error" in msg:
+                    err = msg.get("error") or {}
+                    raise RuntimeError(
+                        f"Copilot ACP {method} failed: {err.get('message') or err}"
+                    )
+                return msg.get("result")
+
+            stderr_text = "\n".join(stderr_tail).strip()
+            if proc.poll() is not None and stderr_text:
+                raise RuntimeError(f"Copilot ACP process exited early: {stderr_text}")
+            raise TimeoutError(
+                f"Timed out waiting for Copilot ACP response to {method}."
+            )
+
+        try:
+            _request(
+                "initialize",
+                {
+                    "protocolVersion": 1,
+                    "clientCapabilities": {
+                        "fs": {
+                            "readTextFile": True,
+                            "writeTextFile": True,
+                        }
+                    },
+                    "clientInfo": {
+                        "name": "hermes-agent",
+                        "title": "Hermes Agent",
+                        "version": "0.0.0",
+                    },
+                },
+            )
+            session = (
+                _request(
+                    "session/new",
+                    {
+                        "cwd": self._acp_cwd,
+                        "mcpServers": [],
+                    },
+                )
+                or {}
+            )
+            session_id = str(session.get("sessionId") or "").strip()
+            if not session_id:
+                raise RuntimeError("Copilot ACP did not return a sessionId.")
+
+            text_parts: list[str] = []
+            reasoning_parts: list[str] = []
+            _request(
+                "session/prompt",
+                {
+                    "sessionId": session_id,
+                    "prompt": [
+                        {
+                            "type": "text",
+                            "text": prompt_text,
+                        }
+                    ],
+                },
+                text_parts=text_parts,
+                reasoning_parts=reasoning_parts,
+            )
+            return "".join(text_parts), "".join(reasoning_parts)
+        finally:
+            self.close()
+
+    def _handle_server_message(
+        self,
+        msg: dict[str, Any],
+        *,
+        process: subprocess.Popen[str],
+        cwd: str,
+        text_parts: list[str] | None,
+        reasoning_parts: list[str] | None,
+    ) -> bool:
+        method = msg.get("method")
+        if not isinstance(method, str):
+            return False
+
+        if method == "session/update":
+            params = msg.get("params") or {}
+            update = params.get("update") or {}
+            kind = str(update.get("sessionUpdate") or "").strip()
+            content = update.get("content") or {}
+            chunk_text = ""
+            if isinstance(content, dict):
+                chunk_text = str(content.get("text") or "")
+            if kind == "agent_message_chunk" and chunk_text and text_parts is not None:
+                text_parts.append(chunk_text)
+            elif (
+                kind == "agent_thought_chunk"
+                and chunk_text
+                and reasoning_parts is not None
+            ):
+                reasoning_parts.append(chunk_text)
+            return True
+
+        if process.stdin is None:
+            return True
+
+        message_id = msg.get("id")
+        params = msg.get("params") or {}
+
+        if method == "session/request_permission":
+            response = _permission_denied(message_id)
+        elif method == "fs/read_text_file":
+            try:
+                path = _ensure_path_within_cwd(str(params.get("path") or ""), cwd)
+                block_error = get_read_block_error(str(path))
+                if block_error:
+                    raise PermissionError(block_error)
+                content = path.read_text() if path.exists() else ""
+                line = params.get("line")
+                limit = params.get("limit")
+                if isinstance(line, int) and line > 1:
+                    lines = content.splitlines(keepends=True)
+                    start = line - 1
+                    end = (
+                        start + limit if isinstance(limit, int) and limit > 0 else None
+                    )
+                    content = "".join(lines[start:end])
+                if content:
+                    content = redact_sensitive_text(content)
+                response = {
+                    "jsonrpc": "2.0",
+                    "id": message_id,
+                    "result": {
+                        "content": content,
+                    },
+                }
+            except Exception as exc:
+                response = _jsonrpc_error(message_id, -32602, str(exc))
+        elif method == "fs/write_text_file":
+            try:
+                path = _ensure_path_within_cwd(str(params.get("path") or ""), cwd)
+                if is_write_denied(str(path)):
+                    raise PermissionError(
+                        f"Write denied: '{path}' is a protected system/credential file."
+                    )
+                path.parent.mkdir(parents=True, exist_ok=True)
+                path.write_text(str(params.get("content") or ""))
+                response = {
+                    "jsonrpc": "2.0",
+                    "id": message_id,
+                    "result": None,
+                }
+            except Exception as exc:
+                response = _jsonrpc_error(message_id, -32602, str(exc))
+        else:
+            response = _jsonrpc_error(
+                message_id,
+                -32601,
+                f"ACP client method '{method}' is not supported by Hermes yet.",
+            )
+
+        process.stdin.write(json.dumps(response) + "\n")
+        process.stdin.flush()
+        return True

agent/auxiliary_client.py

@@ -82,6 +82,8 @@ _PROVIDER_ALIASES = {
     "moonshot": "kimi-coding",
     "kimi-cn": "kimi-coding-cn",
     "moonshot-cn": "kimi-coding-cn",
+    "gmi-cloud": "gmi",
+    "gmicloud": "gmi",
     "minimax-china": "minimax-cn",
     "minimax_cn": "minimax-cn",
     "claude": "anthropic",
@@ -149,22 +151,31 @@ def _fixed_temperature_for_model(
     return None
 
 # Default auxiliary models for direct API-key providers (cheap/fast for side tasks)
-_API_KEY_PROVIDER_AUX_MODELS: Dict[str, str] = {
-    "gemini": "gemini-3-flash-preview",
-    "zai": "glm-4.5-flash",
-    "kimi-coding": "kimi-k2-turbo-preview",
-    "stepfun": "step-3.5-flash",
-    "kimi-coding-cn": "kimi-k2-turbo-preview",
-    "minimax": "MiniMax-M2.7",
-    "minimax-cn": "MiniMax-M2.7",
+def _get_aux_model_for_provider(provider_id: str) -> str:
+    """Return the cheap auxiliary model for a provider.
+
+    Reads from ProviderProfile.default_aux_model first, falling back to the
+    legacy hardcoded dict for providers that predate the profiles system.
+    """
+    try:
+        from providers import get_provider_profile
+        _p = get_provider_profile(provider_id)
+        if _p and _p.default_aux_model:
+            return _p.default_aux_model
+    except Exception:
+        pass
+    return _API_KEY_PROVIDER_AUX_MODELS_FALLBACK.get(provider_id, "")
+
+
+# Fallback for providers not yet migrated to ProviderProfile.default_aux_model.
+# New providers should set default_aux_model on their profile instead.
+_API_KEY_PROVIDER_AUX_MODELS_FALLBACK: Dict[str, str] = {
     "anthropic": "claude-haiku-4-5-20251001",
-    "ai-gateway": "google/gemini-3-flash",
-    "opencode-zen": "gemini-3-flash",
-    "opencode-go": "glm-5",
-    "kilocode": "google/gemini-3-flash-preview",
-    "ollama-cloud": "nemotron-3-nano:30b",
 }
 
+# Legacy alias — callers that haven't been updated yet can still use this.
+_API_KEY_PROVIDER_AUX_MODELS: Dict[str, str] = _API_KEY_PROVIDER_AUX_MODELS_FALLBACK
+
 # Vision-specific model overrides for direct providers.
 # When the user's main provider has a dedicated vision/multimodal model that
 # differs from their main chat model, map it here.  The vision auto-detect
@@ -865,7 +876,7 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]:
             base_url = _to_openai_base_url(
                 _pool_runtime_base_url(entry, pconfig.inference_base_url) or pconfig.inference_base_url
             )
-            model = _API_KEY_PROVIDER_AUX_MODELS.get(provider_id)
+            model = _get_aux_model_for_provider(provider_id) or None
             if model is None:
                 continue  # skip provider if we don't know a valid aux model
             logger.debug("Auxiliary text client: %s (%s) via pool", pconfig.name, model)
@@ -874,14 +885,22 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]:
 
                 if is_native_gemini_base_url(base_url):
                     return GeminiNativeClient(api_key=api_key, base_url=base_url), model
-            extra = {}
-            if base_url_host_matches(base_url, "api.kimi.com"):
-                extra["default_headers"] = {"User-Agent": "claude-code/0.1.0"}
-            elif base_url_host_matches(base_url, "api.githubcopilot.com"):
-                from hermes_cli.models import copilot_default_headers
+        extra = {}
+        if base_url_host_matches(base_url, "api.kimi.com"):
+            extra["default_headers"] = {"User-Agent": "claude-code/0.1.0"}
+        elif base_url_host_matches(base_url, "api.githubcopilot.com"):
+            from hermes_cli.models import copilot_default_headers
 
-                extra["default_headers"] = copilot_default_headers()
-            return OpenAI(api_key=api_key, base_url=base_url, **extra), model
+            extra["default_headers"] = copilot_default_headers()
+        else:
+            try:
+                from providers import get_provider_profile as _gpf_aux
+                _ph_aux = _gpf_aux(provider_id)
+                if _ph_aux and _ph_aux.default_headers:
+                    extra["default_headers"] = dict(_ph_aux.default_headers)
+            except Exception:
+                pass
+        return OpenAI(api_key=api_key, base_url=base_url, **extra), model
 
         creds = resolve_api_key_provider_credentials(provider_id)
         api_key = str(creds.get("api_key", "")).strip()
@@ -891,7 +910,7 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]:
         base_url = _to_openai_base_url(
             str(creds.get("base_url", "")).strip().rstrip("/") or pconfig.inference_base_url
         )
-        model = _API_KEY_PROVIDER_AUX_MODELS.get(provider_id)
+        model = _get_aux_model_for_provider(provider_id) or None
         if model is None:
             continue  # skip provider if we don't know a valid aux model
         logger.debug("Auxiliary text client: %s (%s)", pconfig.name, model)
@@ -907,6 +926,14 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]:
             from hermes_cli.models import copilot_default_headers
 
             extra["default_headers"] = copilot_default_headers()
+        else:
+            try:
+                from providers import get_provider_profile as _gpf_aux2
+                _ph_aux2 = _gpf_aux2(provider_id)
+                if _ph_aux2 and _ph_aux2.default_headers:
+                    extra["default_headers"] = dict(_ph_aux2.default_headers)
+            except Exception:
+                pass
         return OpenAI(api_key=api_key, base_url=base_url, **extra), model
 
     return None, None
@@ -1255,7 +1282,7 @@ def _try_anthropic() -> Tuple[Optional[Any], Optional[str]]:
 
     from agent.anthropic_adapter import _is_oauth_token
     is_oauth = _is_oauth_token(token)
-    model = _API_KEY_PROVIDER_AUX_MODELS.get("anthropic", "claude-haiku-4-5-20251001")
+    model = _get_aux_model_for_provider("anthropic") or "claude-haiku-4-5-20251001"
     logger.debug("Auxiliary client: Anthropic native (%s) at %s (oauth=%s)", model, base_url, is_oauth)
     try:
         real_client = build_anthropic_client(token, base_url)
@@ -1639,7 +1666,7 @@ def _to_async_client(sync_client, model: str, is_vision: bool = False):
     except ImportError:
         pass
     try:
-        from agent.copilot_acp_client import CopilotACPClient
+        from acp_adapter.copilot_client import CopilotACPClient
         if isinstance(sync_client, CopilotACPClient):
             return sync_client, model
     except ImportError:
@@ -1983,7 +2010,7 @@ def resolve_provider_client(
             str(creds.get("base_url", "")).strip().rstrip("/") or pconfig.inference_base_url
         )
 
-        default_model = _API_KEY_PROVIDER_AUX_MODELS.get(provider, "")
+        default_model = _get_aux_model_for_provider(provider)
         final_model = _normalize_resolved_model(model or default_model, provider)
 
         if provider == "gemini":
@@ -2053,7 +2080,7 @@ def resolve_provider_client(
                     "process credentials are incomplete"
                 )
                 return None, None
-            from agent.copilot_acp_client import CopilotACPClient
+            from acp_adapter.copilot_client import CopilotACPClient
 
             client = CopilotACPClient(
                 api_key=api_key,
@@ -2558,12 +2585,19 @@ def _is_openrouter_client(client: Any) -> bool:
     return False
 
 
+def _cached_client_accepts_slash_models(client: Any, cached_default: Optional[str]) -> bool:
+    """Best-effort check for cached clients that accept ``vendor/model`` IDs."""
+    if _is_openrouter_client(client):
+        return True
+    return bool(cached_default and "/" in cached_default)
+
+
 def _compat_model(client: Any, model: Optional[str], cached_default: Optional[str]) -> Optional[str]:
-    """Drop OpenRouter-format model slugs (with '/') for non-OpenRouter clients.
+    """Keep slash-bearing model IDs only for cached clients that support them.
 
     Mirrors the guard in resolve_provider_client() which is skipped on cache hits.
     """
-    if model and "/" in model and not _is_openrouter_client(client):
+    if model and "/" in model and not _cached_client_accepts_slash_models(client, cached_default):
         return cached_default
     return model or cached_default
 

agent/context_compressor.py

@@ -338,6 +338,10 @@ class ContextCompressor(ContextEngine):
         self._context_probe_persistable = False
         self._previous_summary = None
         self._last_summary_error = None
+        self._last_summary_dropped_count = 0
+        self._last_summary_fallback_used = False
+        self._last_aux_model_failure_error = None
+        self._last_aux_model_failure_model = None
         self._last_compression_savings_pct = 100.0
         self._ineffective_compression_count = 0
 
@@ -441,6 +445,17 @@ class ContextCompressor(ContextEngine):
         self._ineffective_compression_count: int = 0
         self._summary_failure_cooldown_until: float = 0.0
         self._last_summary_error: Optional[str] = None
+        # When summary generation fails and a static fallback is inserted,
+        # record how many turns were unrecoverably dropped so callers
+        # (gateway hygiene, /compress) can surface a visible warning.
+        self._last_summary_dropped_count: int = 0
+        self._last_summary_fallback_used: bool = False
+        # When a user-configured summary model fails and we recover by
+        # retrying on the main model, record the failure so gateway /
+        # CLI callers can still warn the user even though compression
+        # succeeded.  Silent recovery would hide the broken config.
+        self._last_aux_model_failure_error: Optional[str] = None
+        self._last_aux_model_failure_model: Optional[str] = None
 
     def update_from_response(self, usage: Dict[str, Any]):
         """Update tracked token usage from API response."""
@@ -900,10 +915,50 @@ The user has requested that this compaction PRIORITISE preserving all informatio
                     "Falling back to main model '%s' for compression.",
                     self.summary_model, e, self.model,
                 )
+                # Record the aux-model failure so callers can warn the user
+                # even if the retry-on-main succeeds — a misconfigured aux
+                # model is something the user needs to fix.
+                _err_text = str(e).strip() or e.__class__.__name__
+                if len(_err_text) > 220:
+                    _err_text = _err_text[:217].rstrip() + "..."
+                self._last_aux_model_failure_error = _err_text
+                self._last_aux_model_failure_model = self.summary_model
                 self.summary_model = ""  # empty = use main model
                 self._summary_failure_cooldown_until = 0.0  # no cooldown
                 return self._generate_summary(turns_to_summarize, focus_topic=focus_topic)  # retry immediately
 
+            # Unknown-error best-effort retry on main model.  Losing N turns of
+            # context is almost always worse than one extra summary attempt, so
+            # if we haven't already fallen back and the summary model differs
+            # from the main model, try once more on main before entering
+            # cooldown.  Errors that DID match _is_model_not_found above are
+            # already handled by the fast-path retry; this branch catches
+            # everything else (400s, provider-specific "no route" strings,
+            # aggregator rejections, etc.) where auto-retry is still safer
+            # than dropping the turns.
+            if (
+                self.summary_model
+                and self.summary_model != self.model
+                and not getattr(self, "_summary_model_fallen_back", False)
+            ):
+                self._summary_model_fallen_back = True
+                logging.warning(
+                    "Summary model '%s' failed (%s). "
+                    "Retrying on main model '%s' before giving up.",
+                    self.summary_model, e, self.model,
+                )
+                # Record the aux-model failure (see 404 branch above) — user
+                # should know their configured model is broken even if main
+                # recovers the call.
+                _err_text = str(e).strip() or e.__class__.__name__
+                if len(_err_text) > 220:
+                    _err_text = _err_text[:217].rstrip() + "..."
+                self._last_aux_model_failure_error = _err_text
+                self._last_aux_model_failure_model = self.summary_model
+                self.summary_model = ""  # empty = use main model
+                self._summary_failure_cooldown_until = 0.0
+                return self._generate_summary(turns_to_summarize, focus_topic=focus_topic)
+
             # Transient errors (timeout, rate limit, network) — shorter cooldown
             _transient_cooldown = 60
             self._summary_failure_cooldown_until = time.monotonic() + _transient_cooldown
@@ -1196,6 +1251,13 @@ The user has requested that this compaction PRIORITISE preserving all informatio
                 related to this topic and be more aggressive about compressing
                 everything else.  Inspired by Claude Code's ``/compact``.
         """
+        # Reset per-call summary failure state — callers inspect these fields
+        # after compress() returns to decide whether to surface a warning.
+        self._last_summary_dropped_count = 0
+        self._last_summary_fallback_used = False
+        self._last_summary_error = None
+        self._last_aux_model_failure_error = None
+        self._last_aux_model_failure_model = None
         n_messages = len(messages)
         # Only need head + 3 tail messages minimum (token budget decides the real tail size)
         _min_for_compress = self.protect_first_n + 3 + 1
@@ -1274,11 +1336,13 @@ The user has requested that this compaction PRIORITISE preserving all informatio
             if not self.quiet_mode:
                 logger.warning("Summary generation failed — inserting static fallback context marker")
             n_dropped = compress_end - compress_start
+            self._last_summary_dropped_count = n_dropped
+            self._last_summary_fallback_used = True
             summary = (
                 f"{SUMMARY_PREFIX}\n"
-                f"Summary generation was unavailable. {n_dropped} conversation turns were "
+                f"Summary generation was unavailable. {n_dropped} message(s) were "
                 f"removed to free context space but could not be summarized. The removed "
-                f"turns contained earlier work in this session. Continue based on the "
+                f"messages contained earlier work in this session. Continue based on the "
                 f"recent messages below and the current state of any files or resources."
             )
 

agent/copilot_acp_client.py

@@ -1,646 +1,8 @@
-"""OpenAI-compatible shim that forwards Hermes requests to `copilot --acp`.
+"""Backward-compatibility shim.
 
-This adapter lets Hermes treat the GitHub Copilot ACP server as a chat-style
-backend. Each request starts a short-lived ACP session, sends the formatted
-conversation as a single prompt, collects text chunks, and converts the result
-back into the minimal shape Hermes expects from an OpenAI client.
+CopilotACPClient has moved to acp_adapter/copilot_client.py.
+This module re-exports it so existing callers continue to work.
 """
+from acp_adapter.copilot_client import CopilotACPClient  # noqa: F401
 
-from __future__ import annotations
-
-import json
-import os
-import queue
-import re
-import shlex
-import subprocess
-import threading
-import time
-from collections import deque
-from pathlib import Path
-from types import SimpleNamespace
-from typing import Any
-
-from agent.file_safety import get_read_block_error, is_write_denied
-from agent.redact import redact_sensitive_text
-
-ACP_MARKER_BASE_URL = "acp://copilot"
-_DEFAULT_TIMEOUT_SECONDS = 900.0
-
-_TOOL_CALL_BLOCK_RE = re.compile(r"<tool_call>\s*(\{.*?\})\s*</tool_call>", re.DOTALL)
-_TOOL_CALL_JSON_RE = re.compile(r"\{\s*\"id\"\s*:\s*\"[^\"]+\"\s*,\s*\"type\"\s*:\s*\"function\"\s*,\s*\"function\"\s*:\s*\{.*?\}\s*\}", re.DOTALL)
-
-
-def _resolve_command() -> str:
-    return (
-        os.getenv("HERMES_COPILOT_ACP_COMMAND", "").strip()
-        or os.getenv("COPILOT_CLI_PATH", "").strip()
-        or "copilot"
-    )
-
-
-def _resolve_args() -> list[str]:
-    raw = os.getenv("HERMES_COPILOT_ACP_ARGS", "").strip()
-    if not raw:
-        return ["--acp", "--stdio"]
-    return shlex.split(raw)
-
-
-def _resolve_home_dir() -> str:
-    """Return a stable HOME for child ACP processes."""
-
-    try:
-        from hermes_constants import get_subprocess_home
-
-        profile_home = get_subprocess_home()
-        if profile_home:
-            return profile_home
-    except Exception:
-        pass
-
-    home = os.environ.get("HOME", "").strip()
-    if home:
-        return home
-
-    expanded = os.path.expanduser("~")
-    if expanded and expanded != "~":
-        return expanded
-
-    try:
-        import pwd
-
-        resolved = pwd.getpwuid(os.getuid()).pw_dir.strip()
-        if resolved:
-            return resolved
-    except Exception:
-        pass
-
-    # Last resort: /tmp (writable on any POSIX system). Avoids crashing the
-    # subprocess with no HOME; callers can set HERMES_HOME explicitly if they
-    # need a different writable dir.
-    return "/tmp"
-
-
-def _build_subprocess_env() -> dict[str, str]:
-    env = os.environ.copy()
-    env["HOME"] = _resolve_home_dir()
-    return env
-
-
-def _jsonrpc_error(message_id: Any, code: int, message: str) -> dict[str, Any]:
-    return {
-        "jsonrpc": "2.0",
-        "id": message_id,
-        "error": {
-            "code": code,
-            "message": message,
-        },
-    }
-
-
-def _permission_denied(message_id: Any) -> dict[str, Any]:
-    return {
-        "jsonrpc": "2.0",
-        "id": message_id,
-        "result": {
-            "outcome": {
-                "outcome": "cancelled",
-            }
-        },
-    }
-
-
-def _format_messages_as_prompt(
-    messages: list[dict[str, Any]],
-    model: str | None = None,
-    tools: list[dict[str, Any]] | None = None,
-    tool_choice: Any = None,
-) -> str:
-    sections: list[str] = [
-        "You are being used as the active ACP agent backend for Hermes.",
-        "Use ACP capabilities to complete tasks.",
-        "IMPORTANT: If you take an action with a tool, you MUST output tool calls using <tool_call>{...}</tool_call> blocks with JSON exactly in OpenAI function-call shape.",
-        "If no tool is needed, answer normally.",
-    ]
-    if model:
-        sections.append(f"Hermes requested model hint: {model}")
-
-    if isinstance(tools, list) and tools:
-        tool_specs: list[dict[str, Any]] = []
-        for t in tools:
-            if not isinstance(t, dict):
-                continue
-            fn = t.get("function") or {}
-            if not isinstance(fn, dict):
-                continue
-            name = fn.get("name")
-            if not isinstance(name, str) or not name.strip():
-                continue
-            tool_specs.append(
-                {
-                    "name": name.strip(),
-                    "description": fn.get("description", ""),
-                    "parameters": fn.get("parameters", {}),
-                }
-            )
-        if tool_specs:
-            sections.append(
-                "Available tools (OpenAI function schema). "
-                "When using a tool, emit ONLY <tool_call>{...}</tool_call> with one JSON object "
-                "containing id/type/function{name,arguments}. arguments must be a JSON string.\n"
-                + json.dumps(tool_specs, ensure_ascii=False)
-            )
-
-    if tool_choice is not None:
-        sections.append(f"Tool choice hint: {json.dumps(tool_choice, ensure_ascii=False)}")
-
-    transcript: list[str] = []
-    for message in messages:
-        if not isinstance(message, dict):
-            continue
-        role = str(message.get("role") or "unknown").strip().lower()
-        if role == "tool":
-            role = "tool"
-        elif role not in {"system", "user", "assistant"}:
-            role = "context"
-
-        content = message.get("content")
-        rendered = _render_message_content(content)
-        if not rendered:
-            continue
-
-        label = {
-            "system": "System",
-            "user": "User",
-            "assistant": "Assistant",
-            "tool": "Tool",
-            "context": "Context",
-        }.get(role, role.title())
-        transcript.append(f"{label}:\n{rendered}")
-
-    if transcript:
-        sections.append("Conversation transcript:\n\n" + "\n\n".join(transcript))
-
-    sections.append("Continue the conversation from the latest user request.")
-    return "\n\n".join(section.strip() for section in sections if section and section.strip())
-
-
-def _render_message_content(content: Any) -> str:
-    if content is None:
-        return ""
-    if isinstance(content, str):
-        return content.strip()
-    if isinstance(content, dict):
-        if "text" in content:
-            return str(content.get("text") or "").strip()
-        if "content" in content and isinstance(content.get("content"), str):
-            return str(content.get("content") or "").strip()
-        return json.dumps(content, ensure_ascii=True)
-    if isinstance(content, list):
-        parts: list[str] = []
-        for item in content:
-            if isinstance(item, str):
-                parts.append(item)
-            elif isinstance(item, dict):
-                text = item.get("text")
-                if isinstance(text, str) and text.strip():
-                    parts.append(text.strip())
-        return "\n".join(parts).strip()
-    return str(content).strip()
-
-
-def _extract_tool_calls_from_text(text: str) -> tuple[list[SimpleNamespace], str]:
-    if not isinstance(text, str) or not text.strip():
-        return [], ""
-
-    extracted: list[SimpleNamespace] = []
-    consumed_spans: list[tuple[int, int]] = []
-
-    def _try_add_tool_call(raw_json: str) -> None:
-        try:
-            obj = json.loads(raw_json)
-        except Exception:
-            return
-        if not isinstance(obj, dict):
-            return
-        fn = obj.get("function")
-        if not isinstance(fn, dict):
-            return
-        fn_name = fn.get("name")
-        if not isinstance(fn_name, str) or not fn_name.strip():
-            return
-        fn_args = fn.get("arguments", "{}")
-        if not isinstance(fn_args, str):
-            fn_args = json.dumps(fn_args, ensure_ascii=False)
-        call_id = obj.get("id")
-        if not isinstance(call_id, str) or not call_id.strip():
-            call_id = f"acp_call_{len(extracted)+1}"
-
-        extracted.append(
-            SimpleNamespace(
-                id=call_id,
-                call_id=call_id,
-                response_item_id=None,
-                type="function",
-                function=SimpleNamespace(name=fn_name.strip(), arguments=fn_args),
-            )
-        )
-
-    for m in _TOOL_CALL_BLOCK_RE.finditer(text):
-        raw = m.group(1)
-        _try_add_tool_call(raw)
-        consumed_spans.append((m.start(), m.end()))
-
-    # Only try bare-JSON fallback when no XML blocks were found.
-    if not extracted:
-        for m in _TOOL_CALL_JSON_RE.finditer(text):
-            raw = m.group(0)
-            _try_add_tool_call(raw)
-            consumed_spans.append((m.start(), m.end()))
-
-    if not consumed_spans:
-        return extracted, text.strip()
-
-    consumed_spans.sort()
-    merged: list[tuple[int, int]] = []
-    for start, end in consumed_spans:
-        if not merged or start > merged[-1][1]:
-            merged.append((start, end))
-        else:
-            merged[-1] = (merged[-1][0], max(merged[-1][1], end))
-
-    parts: list[str] = []
-    cursor = 0
-    for start, end in merged:
-        if cursor < start:
-            parts.append(text[cursor:start])
-        cursor = max(cursor, end)
-    if cursor < len(text):
-        parts.append(text[cursor:])
-
-    cleaned = "\n".join(p.strip() for p in parts if p and p.strip()).strip()
-    return extracted, cleaned
-
-
-
-def _ensure_path_within_cwd(path_text: str, cwd: str) -> Path:
-    candidate = Path(path_text)
-    if not candidate.is_absolute():
-        raise PermissionError("ACP file-system paths must be absolute.")
-    resolved = candidate.resolve()
-    root = Path(cwd).resolve()
-    try:
-        resolved.relative_to(root)
-    except ValueError as exc:
-        raise PermissionError(f"Path '{resolved}' is outside the session cwd '{root}'.") from exc
-    return resolved
-
-
-class _ACPChatCompletions:
-    def __init__(self, client: "CopilotACPClient"):
-        self._client = client
-
-    def create(self, **kwargs: Any) -> Any:
-        return self._client._create_chat_completion(**kwargs)
-
-
-class _ACPChatNamespace:
-    def __init__(self, client: "CopilotACPClient"):
-        self.completions = _ACPChatCompletions(client)
-
-
-class CopilotACPClient:
-    """Minimal OpenAI-client-compatible facade for Copilot ACP."""
-
-    def __init__(
-        self,
-        *,
-        api_key: str | None = None,
-        base_url: str | None = None,
-        default_headers: dict[str, str] | None = None,
-        acp_command: str | None = None,
-        acp_args: list[str] | None = None,
-        acp_cwd: str | None = None,
-        command: str | None = None,
-        args: list[str] | None = None,
-        **_: Any,
-    ):
-        self.api_key = api_key or "copilot-acp"
-        self.base_url = base_url or ACP_MARKER_BASE_URL
-        self._default_headers = dict(default_headers or {})
-        self._acp_command = acp_command or command or _resolve_command()
-        self._acp_args = list(acp_args or args or _resolve_args())
-        self._acp_cwd = str(Path(acp_cwd or os.getcwd()).resolve())
-        self.chat = _ACPChatNamespace(self)
-        self.is_closed = False
-        self._active_process: subprocess.Popen[str] | None = None
-        self._active_process_lock = threading.Lock()
-
-    def close(self) -> None:
-        proc: subprocess.Popen[str] | None
-        with self._active_process_lock:
-            proc = self._active_process
-            self._active_process = None
-        self.is_closed = True
-        if proc is None:
-            return
-        try:
-            proc.terminate()
-            proc.wait(timeout=2)
-        except Exception:
-            try:
-                proc.kill()
-            except Exception:
-                pass
-
-    def _create_chat_completion(
-        self,
-        *,
-        model: str | None = None,
-        messages: list[dict[str, Any]] | None = None,
-        timeout: float | None = None,
-        tools: list[dict[str, Any]] | None = None,
-        tool_choice: Any = None,
-        **_: Any,
-    ) -> Any:
-        prompt_text = _format_messages_as_prompt(
-            messages or [],
-            model=model,
-            tools=tools,
-            tool_choice=tool_choice,
-        )
-        # Normalise timeout: run_agent.py may pass an httpx.Timeout object
-        # (used natively by the OpenAI SDK) rather than a plain float.
-        if timeout is None:
-            _effective_timeout = _DEFAULT_TIMEOUT_SECONDS
-        elif isinstance(timeout, (int, float)):
-            _effective_timeout = float(timeout)
-        else:
-            # httpx.Timeout or similar — pick the largest component so the
-            # subprocess has enough wall-clock time for the full response.
-            _candidates = [
-                getattr(timeout, attr, None)
-                for attr in ("read", "write", "connect", "pool", "timeout")
-            ]
-            _numeric = [float(v) for v in _candidates if isinstance(v, (int, float))]
-            _effective_timeout = max(_numeric) if _numeric else _DEFAULT_TIMEOUT_SECONDS
-
-        response_text, reasoning_text = self._run_prompt(
-            prompt_text,
-            timeout_seconds=_effective_timeout,
-        )
-
-        tool_calls, cleaned_text = _extract_tool_calls_from_text(response_text)
-
-        usage = SimpleNamespace(
-            prompt_tokens=0,
-            completion_tokens=0,
-            total_tokens=0,
-            prompt_tokens_details=SimpleNamespace(cached_tokens=0),
-        )
-        assistant_message = SimpleNamespace(
-            content=cleaned_text,
-            tool_calls=tool_calls,
-            reasoning=reasoning_text or None,
-            reasoning_content=reasoning_text or None,
-            reasoning_details=None,
-        )
-        finish_reason = "tool_calls" if tool_calls else "stop"
-        choice = SimpleNamespace(message=assistant_message, finish_reason=finish_reason)
-        return SimpleNamespace(
-            choices=[choice],
-            usage=usage,
-            model=model or "copilot-acp",
-        )
-
-    def _run_prompt(self, prompt_text: str, *, timeout_seconds: float) -> tuple[str, str]:
-        try:
-            proc = subprocess.Popen(
-                [self._acp_command] + self._acp_args,
-                stdin=subprocess.PIPE,
-                stdout=subprocess.PIPE,
-                stderr=subprocess.PIPE,
-                text=True,
-                bufsize=1,
-                cwd=self._acp_cwd,
-                env=_build_subprocess_env(),
-            )
-        except FileNotFoundError as exc:
-            raise RuntimeError(
-                f"Could not start Copilot ACP command '{self._acp_command}'. "
-                "Install GitHub Copilot CLI or set HERMES_COPILOT_ACP_COMMAND/COPILOT_CLI_PATH."
-            ) from exc
-
-        if proc.stdin is None or proc.stdout is None:
-            proc.kill()
-            raise RuntimeError("Copilot ACP process did not expose stdin/stdout pipes.")
-
-        self.is_closed = False
-        with self._active_process_lock:
-            self._active_process = proc
-
-        inbox: queue.Queue[dict[str, Any]] = queue.Queue()
-        stderr_tail: deque[str] = deque(maxlen=40)
-
-        def _stdout_reader() -> None:
-            if proc.stdout is None:
-                return
-            for line in proc.stdout:
-                try:
-                    inbox.put(json.loads(line))
-                except Exception:
-                    inbox.put({"raw": line.rstrip("\n")})
-
-        def _stderr_reader() -> None:
-            if proc.stderr is None:
-                return
-            for line in proc.stderr:
-                stderr_tail.append(line.rstrip("\n"))
-
-        out_thread = threading.Thread(target=_stdout_reader, daemon=True)
-        err_thread = threading.Thread(target=_stderr_reader, daemon=True)
-        out_thread.start()
-        err_thread.start()
-
-        next_id = 0
-
-        def _request(method: str, params: dict[str, Any], *, text_parts: list[str] | None = None, reasoning_parts: list[str] | None = None) -> Any:
-            nonlocal next_id
-            next_id += 1
-            request_id = next_id
-            payload = {
-                "jsonrpc": "2.0",
-                "id": request_id,
-                "method": method,
-                "params": params,
-            }
-            proc.stdin.write(json.dumps(payload) + "\n")
-            proc.stdin.flush()
-
-            deadline = time.time() + timeout_seconds
-            while time.time() < deadline:
-                if proc.poll() is not None:
-                    break
-                try:
-                    msg = inbox.get(timeout=0.1)
-                except queue.Empty:
-                    continue
-
-                if self._handle_server_message(
-                    msg,
-                    process=proc,
-                    cwd=self._acp_cwd,
-                    text_parts=text_parts,
-                    reasoning_parts=reasoning_parts,
-                ):
-                    continue
-
-                if msg.get("id") != request_id:
-                    continue
-                if "error" in msg:
-                    err = msg.get("error") or {}
-                    raise RuntimeError(
-                        f"Copilot ACP {method} failed: {err.get('message') or err}"
-                    )
-                return msg.get("result")
-
-            stderr_text = "\n".join(stderr_tail).strip()
-            if proc.poll() is not None and stderr_text:
-                raise RuntimeError(f"Copilot ACP process exited early: {stderr_text}")
-            raise TimeoutError(f"Timed out waiting for Copilot ACP response to {method}.")
-
-        try:
-            _request(
-                "initialize",
-                {
-                    "protocolVersion": 1,
-                    "clientCapabilities": {
-                        "fs": {
-                            "readTextFile": True,
-                            "writeTextFile": True,
-                        }
-                    },
-                    "clientInfo": {
-                        "name": "hermes-agent",
-                        "title": "Hermes Agent",
-                        "version": "0.0.0",
-                    },
-                },
-            )
-            session = _request(
-                "session/new",
-                {
-                    "cwd": self._acp_cwd,
-                    "mcpServers": [],
-                },
-            ) or {}
-            session_id = str(session.get("sessionId") or "").strip()
-            if not session_id:
-                raise RuntimeError("Copilot ACP did not return a sessionId.")
-
-            text_parts: list[str] = []
-            reasoning_parts: list[str] = []
-            _request(
-                "session/prompt",
-                {
-                    "sessionId": session_id,
-                    "prompt": [
-                        {
-                            "type": "text",
-                            "text": prompt_text,
-                        }
-                    ],
-                },
-                text_parts=text_parts,
-                reasoning_parts=reasoning_parts,
-            )
-            return "".join(text_parts), "".join(reasoning_parts)
-        finally:
-            self.close()
-
-    def _handle_server_message(
-        self,
-        msg: dict[str, Any],
-        *,
-        process: subprocess.Popen[str],
-        cwd: str,
-        text_parts: list[str] | None,
-        reasoning_parts: list[str] | None,
-    ) -> bool:
-        method = msg.get("method")
-        if not isinstance(method, str):
-            return False
-
-        if method == "session/update":
-            params = msg.get("params") or {}
-            update = params.get("update") or {}
-            kind = str(update.get("sessionUpdate") or "").strip()
-            content = update.get("content") or {}
-            chunk_text = ""
-            if isinstance(content, dict):
-                chunk_text = str(content.get("text") or "")
-            if kind == "agent_message_chunk" and chunk_text and text_parts is not None:
-                text_parts.append(chunk_text)
-            elif kind == "agent_thought_chunk" and chunk_text and reasoning_parts is not None:
-                reasoning_parts.append(chunk_text)
-            return True
-
-        if process.stdin is None:
-            return True
-
-        message_id = msg.get("id")
-        params = msg.get("params") or {}
-
-        if method == "session/request_permission":
-            response = _permission_denied(message_id)
-        elif method == "fs/read_text_file":
-            try:
-                path = _ensure_path_within_cwd(str(params.get("path") or ""), cwd)
-                block_error = get_read_block_error(str(path))
-                if block_error:
-                    raise PermissionError(block_error)
-                content = path.read_text() if path.exists() else ""
-                line = params.get("line")
-                limit = params.get("limit")
-                if isinstance(line, int) and line > 1:
-                    lines = content.splitlines(keepends=True)
-                    start = line - 1
-                    end = start + limit if isinstance(limit, int) and limit > 0 else None
-                    content = "".join(lines[start:end])
-                if content:
-                    content = redact_sensitive_text(content)
-                response = {
-                    "jsonrpc": "2.0",
-                    "id": message_id,
-                    "result": {
-                        "content": content,
-                    },
-                }
-            except Exception as exc:
-                response = _jsonrpc_error(message_id, -32602, str(exc))
-        elif method == "fs/write_text_file":
-            try:
-                path = _ensure_path_within_cwd(str(params.get("path") or ""), cwd)
-                if is_write_denied(str(path)):
-                    raise PermissionError(
-                        f"Write denied: '{path}' is a protected system/credential file."
-                    )
-                path.parent.mkdir(parents=True, exist_ok=True)
-                path.write_text(str(params.get("content") or ""))
-                response = {
-                    "jsonrpc": "2.0",
-                    "id": message_id,
-                    "result": None,
-                }
-            except Exception as exc:
-                response = _jsonrpc_error(message_id, -32602, str(exc))
-        else:
-            response = _jsonrpc_error(
-                message_id,
-                -32601,
-                f"ACP client method '{method}' is not supported by Hermes yet.",
-            )
-
-        process.stdin.write(json.dumps(response) + "\n")
-        process.stdin.flush()
-        return True
+__all__ = ["CopilotACPClient"]

agent/memory_manager.py

@@ -63,15 +63,124 @@ def sanitize_context(text: str) -> str:
     return text
 
 
-def build_memory_context_block(raw_context: str) -> str:
-    """Wrap prefetched memory in a fenced block with system note.
+class StreamingContextScrubber:
+    """Stateful scrubber for streaming text that may contain split memory-context spans.
 
-    The fence prevents the model from treating recalled context as user
-    discourse.  Injected at API-call time only — never persisted.
+    The one-shot ``sanitize_context`` regex cannot survive chunk boundaries:
+    a ``<memory-context>`` opened in one delta and closed in a later delta
+    leaks its payload to the UI because the non-greedy block regex needs
+    both tags in one string.  This scrubber runs a small state machine
+    across deltas, holding back partial-tag tails and discarding
+    everything inside a span (including the system-note line).
+
+    Usage::
+
+        scrubber = StreamingContextScrubber()
+        for delta in stream:
+            visible = scrubber.feed(delta)
+            if visible:
+                emit(visible)
+        trailing = scrubber.flush()  # at end of stream
+        if trailing:
+            emit(trailing)
+
+    The scrubber is re-entrant per agent instance.  Callers building new
+    top-level responses (new turn) should create a fresh scrubber or call
+    ``reset()``.
     """
+
+    _OPEN_TAG = "<memory-context>"
+    _CLOSE_TAG = "</memory-context>"
+
+    def __init__(self) -> None:
+        self._in_span: bool = False
+        self._buf: str = ""
+
+    def reset(self) -> None:
+        self._in_span = False
+        self._buf = ""
+
+    def feed(self, text: str) -> str:
+        """Return the visible portion of ``text`` after scrubbing.
+
+        Any trailing fragment that could be the start of an open/close tag
+        is held back in the internal buffer and surfaced on the next
+        ``feed()`` call or discarded/emitted by ``flush()``.
+        """
+        if not text:
+            return ""
+        buf = self._buf + text
+        self._buf = ""
+        out: list[str] = []
+
+        while buf:
+            if self._in_span:
+                idx = buf.lower().find(self._CLOSE_TAG)
+                if idx == -1:
+                    # Hold back a potential partial close tag; drop the rest
+                    held = self._max_partial_suffix(buf, self._CLOSE_TAG)
+                    self._buf = buf[-held:] if held else ""
+                    return "".join(out)
+                # Found close — skip span content + tag, continue
+                buf = buf[idx + len(self._CLOSE_TAG):]
+                self._in_span = False
+            else:
+                idx = buf.lower().find(self._OPEN_TAG)
+                if idx == -1:
+                    # No open tag — hold back a potential partial open tag
+                    held = self._max_partial_suffix(buf, self._OPEN_TAG)
+                    if held:
+                        out.append(buf[:-held])
+                        self._buf = buf[-held:]
+                    else:
+                        out.append(buf)
+                    return "".join(out)
+                # Emit text before the tag, enter span
+                if idx > 0:
+                    out.append(buf[:idx])
+                buf = buf[idx + len(self._OPEN_TAG):]
+                self._in_span = True
+
+        return "".join(out)
+
+    def flush(self) -> str:
+        """Emit any held-back buffer at end-of-stream.
+
+        If we're still inside an unterminated span the remaining content is
+        discarded (safer: leaking partial memory context is worse than a
+        truncated answer).  Otherwise the held-back partial-tag tail is
+        emitted verbatim (it turned out not to be a real tag).
+        """
+        if self._in_span:
+            self._buf = ""
+            self._in_span = False
+            return ""
+        tail = self._buf
+        self._buf = ""
+        return tail
+
+    @staticmethod
+    def _max_partial_suffix(buf: str, tag: str) -> int:
+        """Return the length of the longest buf-suffix that is a tag-prefix.
+
+        Case-insensitive.  Returns 0 if no suffix could start the tag.
+        """
+        tag_lower = tag.lower()
+        buf_lower = buf.lower()
+        max_check = min(len(buf_lower), len(tag_lower) - 1)
+        for i in range(max_check, 0, -1):
+            if tag_lower.startswith(buf_lower[-i:]):
+                return i
+        return 0
+
+
+def build_memory_context_block(raw_context: str) -> str:
+    """Wrap prefetched memory in a fenced block with system note."""
     if not raw_context or not raw_context.strip():
         return ""
     clean = sanitize_context(raw_context)
+    if clean != raw_context:
+        logger.warning("memory provider returned pre-wrapped context; stripped")
     return (
         "<memory-context>\n"
         "[System note: The following is recalled memory context, "

agent/model_metadata.py

@@ -51,6 +51,7 @@ _PROVIDER_PREFIXES: frozenset[str] = frozenset({
     "qwen-oauth",
     "xiaomi",
     "arcee",
+    "gmi",
     "custom", "local",
     # Common aliases
     "google", "google-gemini", "google-ai-studio",
@@ -60,6 +61,7 @@ _PROVIDER_PREFIXES: frozenset[str] = frozenset({
     "stepfun", "opencode", "zen", "go", "vercel", "kilo", "dashscope", "aliyun", "qwen",
     "mimo", "xiaomi-mimo",
     "arcee-ai", "arceeai",
+    "gmi-cloud", "gmicloud",
     "xai", "x-ai", "x.ai", "grok",
     "nvidia", "nim", "nvidia-nim", "nemotron",
     "qwen-portal",
@@ -307,9 +309,21 @@ _URL_TO_PROVIDER: Dict[str, str] = {
     "integrate.api.nvidia.com": "nvidia",
     "api.xiaomimimo.com": "xiaomi",
     "xiaomimimo.com": "xiaomi",
+    "api.gmi-serving.com": "gmi",
     "ollama.com": "ollama-cloud",
 }
 
+# Auto-extend with hostnames derived from provider profiles.
+# Any provider with a base_url not already in the map gets added automatically.
+try:
+    from providers import list_providers as _list_providers
+    for _pp in _list_providers():
+        _host = _pp.get_hostname()
+        if _host and _host not in _URL_TO_PROVIDER:
+            _URL_TO_PROVIDER[_host] = _pp.name
+except Exception:
+    pass
+
 
 def _infer_provider_from_url(base_url: str) -> Optional[str]:
     """Infer the models.dev provider name from a base URL.
@@ -702,6 +716,29 @@ def fetch_endpoint_model_metadata(
     return {}
 
 
+def _resolve_endpoint_context_length(
+    model: str,
+    base_url: str,
+    api_key: str = "",
+) -> Optional[int]:
+    """Resolve context length from an endpoint's live ``/models`` metadata."""
+    endpoint_metadata = fetch_endpoint_model_metadata(base_url, api_key=api_key)
+    matched = endpoint_metadata.get(model)
+    if not matched:
+        if len(endpoint_metadata) == 1:
+            matched = next(iter(endpoint_metadata.values()))
+        else:
+            for key, entry in endpoint_metadata.items():
+                if model in key or key in model:
+                    matched = entry
+                    break
+    if matched:
+        context_length = matched.get("context_length")
+        if isinstance(context_length, int):
+            return context_length
+    return None
+
+
 def _get_context_cache_path() -> Path:
     """Return path to the persistent context length cache file."""
     from hermes_constants import get_hermes_home
@@ -1295,22 +1332,9 @@ def get_model_context_length(
     # returns 128k) instead of the model's full context (400k).  models.dev
     # has the correct per-provider values and is checked at step 5+.
     if _is_custom_endpoint(base_url) and not _is_known_provider_base_url(base_url):
-        endpoint_metadata = fetch_endpoint_model_metadata(base_url, api_key=api_key)
-        matched = endpoint_metadata.get(model)
-        if not matched:
-            # Single-model servers: if only one model is loaded, use it
-            if len(endpoint_metadata) == 1:
-                matched = next(iter(endpoint_metadata.values()))
-            else:
-                # Fuzzy match: substring in either direction
-                for key, entry in endpoint_metadata.items():
-                    if model in key or key in model:
-                        matched = entry
-                        break
-        if matched:
-            context_length = matched.get("context_length")
-            if isinstance(context_length, int):
-                return context_length
+        context_length = _resolve_endpoint_context_length(model, base_url, api_key=api_key)
+        if context_length is not None:
+            return context_length
         if not _is_known_provider_base_url(base_url):
             # 3. Try querying local server directly
             if is_local_endpoint(base_url):
@@ -1374,6 +1398,12 @@ def get_model_context_length(
             if base_url:
                 save_context_length(model, base_url, codex_ctx)
             return codex_ctx
+    if effective_provider == "gmi" and base_url:
+        # GMI exposes authoritative context_length via /models, but it is not
+        # in models.dev yet. Preserve that higher-fidelity endpoint lookup.
+        ctx = _resolve_endpoint_context_length(model, base_url, api_key=api_key)
+        if ctx is not None:
+            return ctx
     if effective_provider:
         from agent.models_dev import lookup_models_dev_context
         ctx = lookup_models_dev_context(effective_provider, model)

agent/transports/__init__.py

@@ -6,9 +6,16 @@ Usage:
     result = transport.normalize_response(raw_response)
 """
 
-from agent.transports.types import NormalizedResponse, ToolCall, Usage, build_tool_call, map_finish_reason  # noqa: F401
+from agent.transports.types import (
+    NormalizedResponse,
+    ToolCall,
+    Usage,
+    build_tool_call,
+    map_finish_reason,
+)  # noqa: F401
 
 _REGISTRY: dict = {}
+_discovered: bool = False
 
 
 def register_transport(api_mode: str, transport_cls: type) -> None:
@@ -23,6 +30,9 @@ def get_transport(api_mode: str):
     This allows gradual migration — call sites can check for None
     and fall back to the legacy code path.
     """
+    global _discovered
+    if not _discovered:
+        _discover_transports()
     cls = _REGISTRY.get(api_mode)
     if cls is None:
         # The registry can be partially populated when a specific transport
@@ -38,6 +48,8 @@ def get_transport(api_mode: str):
 
 def _discover_transports() -> None:
     """Import all transport modules to trigger auto-registration."""
+    global _discovered
+    _discovered = True
     try:
         import agent.transports.anthropic  # noqa: F401
     except ImportError:

agent/transports/chat_completions.py

@@ -10,7 +10,7 @@ reasoning configuration, temperature handling, and extra_body assembly.
 """
 
 import copy
-from typing import Any, Dict, List, Optional
+from typing import Any
 
 from agent.moonshot_schema import is_moonshot_model, sanitize_moonshot_tools
 from agent.prompt_builder import DEVELOPER_ROLE_MODELS
@@ -28,7 +28,9 @@ class ChatCompletionsTransport(ProviderTransport):
     def api_mode(self) -> str:
         return "chat_completions"
 
-    def convert_messages(self, messages: List[Dict[str, Any]], **kwargs) -> List[Dict[str, Any]]:
+    def convert_messages(
+        self, messages: list[dict[str, Any]], **kwargs
+    ) -> list[dict[str, Any]]:
         """Messages are already in OpenAI format — sanitize Codex leaks only.
 
         Strips Codex Responses API fields (``codex_reasoning_items`` /
@@ -45,7 +47,9 @@ class ChatCompletionsTransport(ProviderTransport):
             tool_calls = msg.get("tool_calls")
             if isinstance(tool_calls, list):
                 for tc in tool_calls:
-                    if isinstance(tc, dict) and ("call_id" in tc or "response_item_id" in tc):
+                    if isinstance(tc, dict) and (
+                        "call_id" in tc or "response_item_id" in tc
+                    ):
                         needs_sanitize = True
                         break
                 if needs_sanitize:
@@ -68,76 +72,52 @@ class ChatCompletionsTransport(ProviderTransport):
                         tc.pop("response_item_id", None)
         return sanitized
 
-    def convert_tools(self, tools: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+    def convert_tools(self, tools: list[dict[str, Any]]) -> list[dict[str, Any]]:
         """Tools are already in OpenAI format — identity."""
         return tools
 
     def build_kwargs(
         self,
         model: str,
-        messages: List[Dict[str, Any]],
-        tools: Optional[List[Dict[str, Any]]] = None,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]] | None = None,
         **params,
-    ) -> Dict[str, Any]:
+    ) -> dict[str, Any]:
         """Build chat.completions.create() kwargs.
 
-        This is the most complex transport method — it handles ~16 providers
-        via params rather than subclasses.
-
-        params:
+        params (all optional):
             timeout: float — API call timeout
             max_tokens: int | None — user-configured max tokens
-            ephemeral_max_output_tokens: int | None — one-shot override (error recovery)
+            ephemeral_max_output_tokens: int | None — one-shot override
             max_tokens_param_fn: callable — returns {max_tokens: N} or {max_completion_tokens: N}
             reasoning_config: dict | None
             request_overrides: dict | None
             session_id: str | None
-            qwen_session_metadata: dict | None — {sessionId, promptId} precomputed
             model_lower: str — lowercase model name for pattern matching
-            # Provider detection flags (all optional, default False)
-            is_openrouter: bool
-            is_nous: bool
-            is_qwen_portal: bool
-            is_github_models: bool
-            is_nvidia_nim: bool
-            is_kimi: bool
-            is_custom_provider: bool
-            ollama_num_ctx: int | None
-            # Provider routing
-            provider_preferences: dict | None
-            # Qwen-specific
-            qwen_prepare_fn: callable | None — runs AFTER codex sanitization
-            qwen_prepare_inplace_fn: callable | None — in-place variant for deepcopied lists
-            # Temperature
-            fixed_temperature: Any — from _fixed_temperature_for_model()
-            omit_temperature: bool
-            # Reasoning
+            # Provider profile path (all per-provider quirks live in providers/)
+            provider_profile: ProviderProfile | None — when present, delegates to
+                _build_kwargs_from_profile(); all flag params below are bypassed.
+            # Remaining flags — only used by the legacy fallback for unregistered
+            # providers (i.e. get_provider_profile() returned None).  Known
+            # providers all go through provider_profile.
+            qwen_session_metadata: dict | None
             supports_reasoning: bool
-            github_reasoning_extra: dict | None
-            # Claude on OpenRouter/Nous max output
             anthropic_max_output: int | None
-            # Extra
-            extra_body_additions: dict | None — pre-built extra_body entries
+            extra_body_additions: dict | None
         """
         # Codex sanitization: drop reasoning_items / call_id / response_item_id
         sanitized = self.convert_messages(messages)
 
-        # Qwen portal prep AFTER codex sanitization.  If sanitize already
-        # deepcopied, reuse that copy via the in-place variant to avoid a
-        # second deepcopy.
-        is_qwen = params.get("is_qwen_portal", False)
-        if is_qwen:
-            qwen_prep = params.get("qwen_prepare_fn")
-            qwen_prep_inplace = params.get("qwen_prepare_inplace_fn")
-            if sanitized is messages:
-                if qwen_prep is not None:
-                    sanitized = qwen_prep(sanitized)
-            else:
-                # Already deepcopied — transform in place
-                if qwen_prep_inplace is not None:
-                    qwen_prep_inplace(sanitized)
-                elif qwen_prep is not None:
-                    sanitized = qwen_prep(sanitized)
+        # ── Provider profile: single-path when present ──────────────────
+        _profile = params.get("provider_profile")
+        if _profile:
+            return self._build_kwargs_from_profile(
+                _profile, model, sanitized, tools, params
+            )
+
+        # ── Legacy fallback (unregistered / unknown provider) ───────────
+        # Reached only when get_provider_profile() returned None.
+        # Known providers always go through the profile path above.
 
         # Developer role swap for GPT-5/Codex models
         model_lower = params.get("model_lower", (model or "").lower())
@@ -150,7 +130,7 @@ class ChatCompletionsTransport(ProviderTransport):
             sanitized = list(sanitized)
             sanitized[0] = {**sanitized[0], "role": "developer"}
 
-        api_kwargs: Dict[str, Any] = {
+        api_kwargs: dict[str, Any] = {
             "model": model,
             "messages": sanitized,
         }
@@ -159,19 +139,6 @@ class ChatCompletionsTransport(ProviderTransport):
         if timeout is not None:
             api_kwargs["timeout"] = timeout
 
-        # Temperature
-        fixed_temp = params.get("fixed_temperature")
-        omit_temp = params.get("omit_temperature", False)
-        if omit_temp:
-            api_kwargs.pop("temperature", None)
-        elif fixed_temp is not None:
-            api_kwargs["temperature"] = fixed_temp
-
-        # Qwen metadata (caller precomputes {sessionId, promptId})
-        qwen_meta = params.get("qwen_session_metadata")
-        if qwen_meta and is_qwen:
-            api_kwargs["metadata"] = qwen_meta
-
         # Tools
         if tools:
             # Moonshot/Kimi uses a stricter flavored JSON Schema.  Rewriting
@@ -186,96 +153,24 @@ class ChatCompletionsTransport(ProviderTransport):
         ephemeral = params.get("ephemeral_max_output_tokens")
         max_tokens = params.get("max_tokens")
         anthropic_max_out = params.get("anthropic_max_output")
-        is_nvidia_nim = params.get("is_nvidia_nim", False)
-        is_kimi = params.get("is_kimi", False)
-        reasoning_config = params.get("reasoning_config")
 
         if ephemeral is not None and max_tokens_fn:
             api_kwargs.update(max_tokens_fn(ephemeral))
         elif max_tokens is not None and max_tokens_fn:
             api_kwargs.update(max_tokens_fn(max_tokens))
-        elif is_nvidia_nim and max_tokens_fn:
-            api_kwargs.update(max_tokens_fn(16384))
-        elif is_qwen and max_tokens_fn:
-            api_kwargs.update(max_tokens_fn(65536))
-        elif is_kimi and max_tokens_fn:
-            # Kimi/Moonshot: 32000 matches Kimi CLI's default
-            api_kwargs.update(max_tokens_fn(32000))
         elif anthropic_max_out is not None:
             api_kwargs["max_tokens"] = anthropic_max_out
 
-        # Kimi: top-level reasoning_effort (unless thinking disabled)
-        if is_kimi:
-            _kimi_thinking_off = bool(
-                reasoning_config
-                and isinstance(reasoning_config, dict)
-                and reasoning_config.get("enabled") is False
-            )
-            if not _kimi_thinking_off:
-                _kimi_effort = "medium"
-                if reasoning_config and isinstance(reasoning_config, dict):
-                    _e = (reasoning_config.get("effort") or "").strip().lower()
-                    if _e in ("low", "medium", "high"):
-                        _kimi_effort = _e
-                api_kwargs["reasoning_effort"] = _kimi_effort
-
         # extra_body assembly
-        extra_body: Dict[str, Any] = {}
+        extra_body: dict[str, Any] = {}
 
-        is_openrouter = params.get("is_openrouter", False)
-        is_nous = params.get("is_nous", False)
-        is_github_models = params.get("is_github_models", False)
-
-        provider_prefs = params.get("provider_preferences")
-        if provider_prefs and is_openrouter:
-            extra_body["provider"] = provider_prefs
-
-        # Kimi extra_body.thinking
-        if is_kimi:
-            _kimi_thinking_enabled = True
-            if reasoning_config and isinstance(reasoning_config, dict):
-                if reasoning_config.get("enabled") is False:
-                    _kimi_thinking_enabled = False
-            extra_body["thinking"] = {
-                "type": "enabled" if _kimi_thinking_enabled else "disabled",
-            }
-
-        # Reasoning
+        # Generic reasoning passthrough for unknown providers
         if params.get("supports_reasoning", False):
-            if is_github_models:
-                gh_reasoning = params.get("github_reasoning_extra")
-                if gh_reasoning is not None:
-                    extra_body["reasoning"] = gh_reasoning
+            reasoning_config = params.get("reasoning_config")
+            if reasoning_config is not None:
+                extra_body["reasoning"] = dict(reasoning_config)
             else:
-                if reasoning_config is not None:
-                    rc = dict(reasoning_config)
-                    if is_nous and rc.get("enabled") is False:
-                        pass  # omit for Nous when disabled
-                    else:
-                        extra_body["reasoning"] = rc
-                else:
-                    extra_body["reasoning"] = {"enabled": True, "effort": "medium"}
-
-        if is_nous:
-            extra_body["tags"] = ["product=hermes-agent"]
-
-        # Ollama num_ctx
-        ollama_ctx = params.get("ollama_num_ctx")
-        if ollama_ctx:
-            options = extra_body.get("options", {})
-            options["num_ctx"] = ollama_ctx
-            extra_body["options"] = options
-
-        # Ollama/custom think=false
-        if params.get("is_custom_provider", False):
-            if reasoning_config and isinstance(reasoning_config, dict):
-                _effort = (reasoning_config.get("effort") or "").strip().lower()
-                _enabled = reasoning_config.get("enabled", True)
-                if _effort == "none" or _enabled is False:
-                    extra_body["think"] = False
-
-        if is_qwen:
-            extra_body["vl_high_resolution_images"] = True
+                extra_body["reasoning"] = {"enabled": True, "effort": "medium"}
 
         # Merge any pre-built extra_body additions
         additions = params.get("extra_body_additions")
@@ -292,6 +187,117 @@ class ChatCompletionsTransport(ProviderTransport):
 
         return api_kwargs
 
+    def _build_kwargs_from_profile(self, profile, model, sanitized, tools, params):
+        """Build API kwargs using a ProviderProfile — single path, no legacy flags.
+
+        This method replaces the entire flag-based kwargs assembly when a
+        provider_profile is passed. Every quirk comes from the profile object.
+        """
+        from providers.base import OMIT_TEMPERATURE
+
+        # Message preprocessing
+        sanitized = profile.prepare_messages(sanitized)
+
+        # Developer role swap — model-name-based, applies to all providers
+        _model_lower = (model or "").lower()
+        if (
+            sanitized
+            and isinstance(sanitized[0], dict)
+            and sanitized[0].get("role") == "system"
+            and any(p in _model_lower for p in DEVELOPER_ROLE_MODELS)
+        ):
+            sanitized = list(sanitized)
+            sanitized[0] = {**sanitized[0], "role": "developer"}
+
+        api_kwargs: dict[str, Any] = {
+            "model": model,
+            "messages": sanitized,
+        }
+
+        # Temperature
+        if profile.fixed_temperature is OMIT_TEMPERATURE:
+            pass  # Don't include temperature at all
+        elif profile.fixed_temperature is not None:
+            api_kwargs["temperature"] = profile.fixed_temperature
+        else:
+            # Use caller's temperature if provided
+            temp = params.get("temperature")
+            if temp is not None:
+                api_kwargs["temperature"] = temp
+
+        # Timeout
+        timeout = params.get("timeout")
+        if timeout is not None:
+            api_kwargs["timeout"] = timeout
+
+        # Tools — apply Moonshot/Kimi schema sanitization regardless of path
+        if tools:
+            if is_moonshot_model(model):
+                tools = sanitize_moonshot_tools(tools)
+            api_kwargs["tools"] = tools
+
+        # max_tokens resolution — priority: ephemeral > user > profile default
+        max_tokens_fn = params.get("max_tokens_param_fn")
+        ephemeral = params.get("ephemeral_max_output_tokens")
+        user_max = params.get("max_tokens")
+        anthropic_max = params.get("anthropic_max_output")
+
+        if ephemeral is not None and max_tokens_fn:
+            api_kwargs.update(max_tokens_fn(ephemeral))
+        elif user_max is not None and max_tokens_fn:
+            api_kwargs.update(max_tokens_fn(user_max))
+        elif profile.default_max_tokens and max_tokens_fn:
+            api_kwargs.update(max_tokens_fn(profile.default_max_tokens))
+        elif anthropic_max is not None:
+            api_kwargs["max_tokens"] = anthropic_max
+
+        # Provider-specific api_kwargs extras (reasoning_effort, metadata, etc.)
+        reasoning_config = params.get("reasoning_config")
+        extra_body_from_profile, top_level_from_profile = (
+            profile.build_api_kwargs_extras(
+                reasoning_config=reasoning_config,
+                supports_reasoning=params.get("supports_reasoning", False),
+                qwen_session_metadata=params.get("qwen_session_metadata"),
+                model=model,
+                ollama_num_ctx=params.get("ollama_num_ctx"),
+            )
+        )
+        api_kwargs.update(top_level_from_profile)
+
+        # extra_body assembly
+        extra_body: dict[str, Any] = {}
+
+        # Profile's extra_body (tags, provider prefs, vl_high_resolution, etc.)
+        profile_body = profile.build_extra_body(
+            session_id=params.get("session_id"),
+            provider_preferences=params.get("provider_preferences"),
+        )
+        if profile_body:
+            extra_body.update(profile_body)
+
+        # Profile's reasoning/thinking extra_body entries
+        if extra_body_from_profile:
+            extra_body.update(extra_body_from_profile)
+
+        # Merge any pre-built extra_body additions from the caller
+        additions = params.get("extra_body_additions")
+        if additions:
+            extra_body.update(additions)
+
+        # Request overrides (user config)
+        overrides = params.get("request_overrides")
+        if overrides:
+            for k, v in overrides.items():
+                if k == "extra_body" and isinstance(v, dict):
+                    extra_body.update(v)
+                else:
+                    api_kwargs[k] = v
+
+        if extra_body:
+            api_kwargs["extra_body"] = extra_body
+
+        return api_kwargs
+
     def normalize_response(self, response: Any, **kwargs) -> NormalizedResponse:
         """Normalize OpenAI ChatCompletion to NormalizedResponse.
 
@@ -313,7 +319,7 @@ class ChatCompletionsTransport(ProviderTransport):
                 # Gemini 3 thinking models attach extra_content with
                 # thought_signature — without replay on the next turn the API
                 # rejects the request with 400.
-                tc_provider_data: Dict[str, Any] = {}
+                tc_provider_data: dict[str, Any] = {}
                 extra = getattr(tc, "extra_content", None)
                 if extra is None and hasattr(tc, "model_extra"):
                     extra = (tc.model_extra or {}).get("extra_content")
@@ -324,12 +330,14 @@ class ChatCompletionsTransport(ProviderTransport):
                         except Exception:
                             pass
                     tc_provider_data["extra_content"] = extra
-                tool_calls.append(ToolCall(
-                    id=tc.id,
-                    name=tc.function.name,
-                    arguments=tc.function.arguments,
-                    provider_data=tc_provider_data or None,
-                ))
+                tool_calls.append(
+                    ToolCall(
+                        id=tc.id,
+                        name=tc.function.name,
+                        arguments=tc.function.arguments,
+                        provider_data=tc_provider_data or None,
+                    )
+                )
 
         usage = None
         if hasattr(response, "usage") and response.usage:
@@ -347,7 +355,7 @@ class ChatCompletionsTransport(ProviderTransport):
         reasoning = getattr(msg, "reasoning", None)
         reasoning_content = getattr(msg, "reasoning_content", None)
 
-        provider_data: Dict[str, Any] = {}
+        provider_data: dict[str, Any] = {}
         if reasoning_content:
             provider_data["reasoning_content"] = reasoning_content
         rd = getattr(msg, "reasoning_details", None)
@@ -373,7 +381,7 @@ class ChatCompletionsTransport(ProviderTransport):
             return False
         return True
 
-    def extract_cache_stats(self, response: Any) -> Optional[Dict[str, int]]:
+    def extract_cache_stats(self, response: Any) -> dict[str, int] | None:
         """Extract OpenRouter/OpenAI cache stats from prompt_tokens_details."""
         usage = getattr(response, "usage", None)
         if usage is None:

agent/transports/types.py

@@ -12,7 +12,7 @@ from __future__ import annotations
 
 import json
 from dataclasses import dataclass, field
-from typing import Any, Dict, List, Optional
+from typing import Any
 
 
 @dataclass
@@ -32,10 +32,10 @@ class ToolCall:
     * Others: ``None``
     """
 
-    id: Optional[str]
+    id: str | None
     name: str
     arguments: str  # JSON string
-    provider_data: Optional[Dict[str, Any]] = field(default=None, repr=False)
+    provider_data: dict[str, Any] | None = field(default=None, repr=False)
 
     # ── Backward compatibility ──────────────────────────────────
     # The agent loop reads tc.function.name / tc.function.arguments
@@ -47,17 +47,17 @@ class ToolCall:
         return "function"
 
     @property
-    def function(self) -> "ToolCall":
+    def function(self) -> ToolCall:
         """Return self so tc.function.name / tc.function.arguments work."""
         return self
 
     @property
-    def call_id(self) -> Optional[str]:
+    def call_id(self) -> str | None:
         """Codex call_id from provider_data, accessed via getattr by _build_assistant_message."""
         return (self.provider_data or {}).get("call_id")
 
     @property
-    def response_item_id(self) -> Optional[str]:
+    def response_item_id(self) -> str | None:
         """Codex response_item_id from provider_data."""
         return (self.provider_data or {}).get("response_item_id")
 
@@ -101,18 +101,18 @@ class NormalizedResponse:
     * Others: ``None``
     """
 
-    content: Optional[str]
-    tool_calls: Optional[List[ToolCall]]
+    content: str | None
+    tool_calls: list[ToolCall] | None
     finish_reason: str  # "stop", "tool_calls", "length", "content_filter"
-    reasoning: Optional[str] = None
-    usage: Optional[Usage] = None
-    provider_data: Optional[Dict[str, Any]] = field(default=None, repr=False)
+    reasoning: str | None = None
+    usage: Usage | None = None
+    provider_data: dict[str, Any] | None = field(default=None, repr=False)
 
     # ── Backward compatibility ──────────────────────────────────
     # The shim _nr_to_assistant_message() mapped these from provider_data.
     # These properties let NormalizedResponse pass through directly.
     @property
-    def reasoning_content(self) -> Optional[str]:
+    def reasoning_content(self) -> str | None:
         pd = self.provider_data or {}
         return pd.get("reasoning_content")
 
@@ -136,8 +136,9 @@ class NormalizedResponse:
 # Factory helpers
 # ---------------------------------------------------------------------------
 
+
 def build_tool_call(
-    id: Optional[str],
+    id: str | None,
     name: str,
     arguments: Any,
     **provider_fields: Any,
@@ -151,7 +152,7 @@ def build_tool_call(
     return ToolCall(id=id, name=name, arguments=args_str, provider_data=pd)
 
 
-def map_finish_reason(reason: Optional[str], mapping: Dict[str, str]) -> str:
+def map_finish_reason(reason: str | None, mapping: dict[str, str]) -> str:
     """Translate a provider-specific stop reason to the normalised set.
 
     Falls back to ``"stop"`` for unknown or ``None`` reasons.

cli.py

@@ -6000,6 +6000,7 @@ class HermesCLI:
             platform_status = {
                 Platform.TELEGRAM: ("Telegram", "TELEGRAM_BOT_TOKEN"),
                 Platform.DISCORD: ("Discord", "DISCORD_BOT_TOKEN"),
+                Platform.SLACK: ("Slack", "SLACK_BOT_TOKEN"),
                 Platform.WHATSAPP: ("WhatsApp", "WHATSAPP_ENABLED"),
             }
             

flake.nix

@@ -36,6 +36,7 @@
 
       imports = [
         ./nix/packages.nix
+        ./nix/overlays.nix
         ./nix/nixosModules.nix
         ./nix/checks.nix
         ./nix/devShell.nix

gateway/config.py

@@ -566,6 +566,8 @@ def load_gateway_config() -> GatewayConfig:
                         existing = {}
                     # Deep-merge extra dicts so gateway.json defaults survive
                     merged_extra = {**existing.get("extra", {}), **plat_block.get("extra", {})}
+                    if plat_name == Platform.SLACK.value and "enabled" in plat_block:
+                        merged_extra["_enabled_explicit"] = True
                     merged = {**existing, **plat_block}
                     if merged_extra:
                         merged["extra"] = merged_extra
@@ -610,16 +612,21 @@ def load_gateway_config() -> GatewayConfig:
                         bridged["channel_prompts"] = {str(k): v for k, v in channel_prompts.items()}
                     else:
                         bridged["channel_prompts"] = channel_prompts
-                if not bridged:
+                enabled_was_explicit = "enabled" in platform_cfg
+                if not bridged and not enabled_was_explicit:
                     continue
                 plat_data = platforms_data.setdefault(plat.value, {})
                 if not isinstance(plat_data, dict):
                     plat_data = {}
                     platforms_data[plat.value] = plat_data
+                if enabled_was_explicit:
+                    plat_data["enabled"] = platform_cfg["enabled"]
                 extra = plat_data.setdefault("extra", {})
                 if not isinstance(extra, dict):
                     extra = {}
                     plat_data["extra"] = extra
+                if plat == Platform.SLACK and enabled_was_explicit:
+                    extra["_enabled_explicit"] = True
                 extra.update(bridged)
 
             # Slack settings → env vars (env vars take precedence)
@@ -941,6 +948,14 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             # No yaml config for Slack — env-only setup, enable it
             config.platforms[Platform.SLACK] = PlatformConfig()
             config.platforms[Platform.SLACK].enabled = True
+        else:
+            slack_config = config.platforms[Platform.SLACK]
+            enabled_was_explicit = bool(slack_config.extra.pop("_enabled_explicit", False))
+            if not slack_config.enabled and not enabled_was_explicit:
+                # Top-level Slack settings such as channel prompts should not
+                # turn an env-token setup into a disabled platform. Only an
+                # explicit slack.enabled/platforms.slack.enabled false should.
+                slack_config.enabled = True
         # If yaml config exists, respect its enabled flag (don't override
         # explicit enabled: false). Token is still stored so skills that
         # send Slack messages can use it without activating the gateway adapter.

gateway/platforms/base.py

@@ -1702,13 +1702,41 @@ class BasePlatformAdapter(ABC):
         the agent is waiting for dangerous-command approval).  This is critical
         for Slack's Assistant API where ``assistant_threads_setStatus`` disables
         the compose box — pausing lets the user type ``/approve`` or ``/deny``.
+
+        Each ``send_typing`` call is bounded by a ~1.5s timeout so a slow
+        network round-trip can't stall the refresh cadence.  Telegram- and
+        Discord-side typing expire after ~5s; if any individual send_typing
+        takes longer than the refresh interval, the bubble would die and
+        stay dead until that call returns.  Abandoning the slow call lets
+        the next tick fire a fresh send_typing on schedule — as long as
+        one of them succeeds within the 5s platform-side window, the bubble
+        stays visible across provider stalls / upstream API timeouts.
         """
+        # Bound each send_typing round-trip so the refresh cadence isn't
+        # gated on network health.  Must stay below ``interval`` so a slow
+        # call gets abandoned before the next scheduled tick.
+        _send_typing_timeout = max(0.25, min(1.5, interval - 0.25))
         try:
             while True:
                 if stop_event is not None and stop_event.is_set():
                     return
                 if chat_id not in self._typing_paused:
-                    await self.send_typing(chat_id, metadata=metadata)
+                    try:
+                        await asyncio.wait_for(
+                            self.send_typing(chat_id, metadata=metadata),
+                            timeout=_send_typing_timeout,
+                        )
+                    except asyncio.TimeoutError:
+                        # Slow network — abandon this tick, keep the loop
+                        # on schedule so the next send_typing fires fresh.
+                        pass
+                    except asyncio.CancelledError:
+                        raise
+                    except Exception as typing_err:
+                        logger.debug(
+                            "[%s] send_typing error (non-fatal): %s",
+                            self.name, typing_err,
+                        )
                 if stop_event is None:
                     await asyncio.sleep(interval)
                     continue

gateway/run.py

@@ -4800,6 +4800,58 @@ class GatewayRunner:
                                             "compression",
                                             f"{_new_tokens:,}",
                                         )
+
+                                    # If summary generation failed, the
+                                    # compressor inserted a static fallback
+                                    # placeholder and the dropped turns are
+                                    # gone for good.  Surface a visible
+                                    # warning to the gateway user — agent.log
+                                    # alone is invisible on TG/Discord/etc.
+                                    _comp = getattr(_hyg_agent, "context_compressor", None)
+                                    if _comp is not None and getattr(_comp, "_last_summary_fallback_used", False):
+                                        _dropped = getattr(_comp, "_last_summary_dropped_count", 0)
+                                        _err = getattr(_comp, "_last_summary_error", None) or "unknown error"
+                                        _warn_msg = (
+                                            "⚠️ Context compression summary failed "
+                                            f"({_err}). {_dropped} historical message(s) "
+                                            "were removed and replaced with a placeholder. "
+                                            "Earlier context is no longer recoverable. "
+                                            "Consider /reset for a clean session, or check "
+                                            "your auxiliary.compression model configuration."
+                                        )
+                                        try:
+                                            _adapter = self.adapters.get(source.platform)
+                                            if _adapter and source.chat_id:
+                                                await _adapter.send(source.chat_id, _warn_msg, metadata=_hyg_meta)
+                                        except Exception as _werr:
+                                            logger.warning(
+                                                "Failed to deliver compression-failure warning to user: %s",
+                                                _werr,
+                                            )
+                                    # Separately: if the user's CONFIGURED aux
+                                    # model failed and we recovered by falling
+                                    # back to the main model, tell them — a
+                                    # misconfigured auxiliary.compression.model
+                                    # is something only they can fix, and
+                                    # silent recovery would hide it.
+                                    elif _comp is not None and getattr(_comp, "_last_aux_model_failure_model", None):
+                                        _aux_model = getattr(_comp, "_last_aux_model_failure_model", "")
+                                        _aux_err = getattr(_comp, "_last_aux_model_failure_error", None) or "unknown error"
+                                        _aux_msg = (
+                                            f"ℹ️ Configured compression model `{_aux_model}` "
+                                            f"failed ({_aux_err}). Recovered using your main "
+                                            "model — context is intact — but you may want to "
+                                            "check `auxiliary.compression.model` in config.yaml."
+                                        )
+                                        try:
+                                            _adapter = self.adapters.get(source.platform)
+                                            if _adapter and source.chat_id:
+                                                await _adapter.send(source.chat_id, _aux_msg, metadata=_hyg_meta)
+                                        except Exception as _werr:
+                                            logger.warning(
+                                                "Failed to deliver aux-model-fallback notice to user: %s",
+                                                _werr,
+                                            )
                                 finally:
                                     self._cleanup_agent_resources(_hyg_agent)
 
@@ -7343,6 +7395,17 @@ class GatewayRunner:
                     approx_tokens,
                     new_tokens,
                 )
+                # Detect summary-generation failure so we can surface a
+                # visible warning to the user even on the manual /compress
+                # path (otherwise the failure is silently logged).
+                _summary_failed = bool(getattr(compressor, "_last_summary_fallback_used", False))
+                _dropped_count = int(getattr(compressor, "_last_summary_dropped_count", 0) or 0)
+                _summary_err = getattr(compressor, "_last_summary_error", None)
+                # Separately: did the user's CONFIGURED aux model fail
+                # and we recovered via main?  Surface that as an info
+                # note so they can fix their config.
+                _aux_fail_model = getattr(compressor, "_last_aux_model_failure_model", None)
+                _aux_fail_err = getattr(compressor, "_last_aux_model_failure_error", None)
             finally:
                 self._cleanup_agent_resources(tmp_agent)
             lines = [f"🗜️ {summary['headline']}"]
@@ -7351,6 +7414,20 @@ class GatewayRunner:
             lines.append(summary["token_line"])
             if summary["note"]:
                 lines.append(summary["note"])
+            if _summary_failed:
+                lines.append(
+                    f"⚠️ Summary generation failed ({_summary_err or 'unknown error'}). "
+                    f"{_dropped_count} historical message(s) were removed and replaced "
+                    "with a placeholder; earlier context is no longer recoverable. "
+                    "Consider checking your auxiliary.compression model configuration."
+                )
+            elif _aux_fail_model:
+                lines.append(
+                    f"ℹ️ Configured compression model `{_aux_fail_model}` failed "
+                    f"({_aux_fail_err or 'unknown error'}). Recovered using your main "
+                    "model — context is intact — but you may want to check "
+                    "`auxiliary.compression.model` in config.yaml."
+                )
             return "\n".join(lines)
         except Exception as e:
             logger.warning("Manual compress failed: %s", e)
@@ -8483,6 +8560,7 @@ class GatewayRunner:
             The enriched message string with vision descriptions prepended.
         """
         from tools.vision_tools import vision_analyze_tool
+        from agent.memory_manager import sanitize_context
 
         analysis_prompt = (
             "Describe everything visible in this image in thorough detail. "
@@ -8501,6 +8579,7 @@ class GatewayRunner:
                 result = json.loads(result_json)
                 if result.get("success"):
                     description = result.get("analysis", "")
+                    description = sanitize_context(description)
                     enriched_parts.append(
                         f"[The user sent an image~ Here's what I can see:\n{description}]\n"
                         f"[If you need a closer look, use vision_analyze with "

hermes_cli/auth.py

@@ -224,6 +224,14 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
         api_key_env_vars=("ARCEEAI_API_KEY",),
         base_url_env_var="ARCEE_BASE_URL",
     ),
+    "gmi": ProviderConfig(
+        id="gmi",
+        name="GMI Cloud",
+        auth_type="api_key",
+        inference_base_url="https://api.gmi-serving.com/v1",
+        api_key_env_vars=("GMI_API_KEY",),
+        base_url_env_var="GMI_BASE_URL",
+    ),
     "minimax": ProviderConfig(
         id="minimax",
         name="MiniMax",
@@ -366,6 +374,37 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
     ),
 }
 
+# Auto-extend PROVIDER_REGISTRY with any api-key provider registered in
+# providers/ that is not already declared above.  New providers only need a
+# providers/*.py file — no edits to this file required.
+try:
+    from providers import list_providers as _list_providers_for_registry
+    for _pp in _list_providers_for_registry():
+        if _pp.name in PROVIDER_REGISTRY:
+            continue
+        if _pp.auth_type != "api_key" or not _pp.env_vars:
+            continue
+        # Skip providers that need custom token resolution (copilot, kimi, zai)
+        # — those are already fully declared above.
+        if _pp.name in {"copilot", "kimi-coding", "kimi-coding-cn", "zai"}:
+            continue
+        _api_key_vars = tuple(v for v in _pp.env_vars if not v.endswith("_BASE_URL") and not v.endswith("_URL"))
+        _base_url_var = next((v for v in _pp.env_vars if v.endswith("_BASE_URL") or v.endswith("_URL")), None)
+        PROVIDER_REGISTRY[_pp.name] = ProviderConfig(
+            id=_pp.name,
+            name=_pp.display_name or _pp.name,
+            auth_type="api_key",
+            inference_base_url=_pp.base_url,
+            api_key_env_vars=_api_key_vars or _pp.env_vars,
+            base_url_env_var=_base_url_var or "",
+        )
+        # Also register aliases so resolve_provider() resolves them
+        for _alias in _pp.aliases:
+            if _alias not in PROVIDER_REGISTRY:
+                PROVIDER_REGISTRY[_alias] = PROVIDER_REGISTRY[_pp.name]
+except Exception:
+    pass
+
 
 # =============================================================================
 # Anthropic Key Helper
@@ -1120,6 +1159,7 @@ def resolve_provider(
         "kimi-cn": "kimi-coding-cn", "moonshot-cn": "kimi-coding-cn",
         "step": "stepfun", "stepfun-coding-plan": "stepfun",
         "arcee-ai": "arcee", "arceeai": "arcee",
+        "gmi-cloud": "gmi", "gmicloud": "gmi",
         "minimax-china": "minimax-cn", "minimax_cn": "minimax-cn",
         "alibaba_coding": "alibaba-coding-plan", "alibaba-coding": "alibaba-coding-plan",
         "alibaba_coding_plan": "alibaba-coding-plan",
@@ -1141,6 +1181,17 @@ def resolve_provider(
         "vllm": "custom", "llamacpp": "custom",
         "llama.cpp": "custom", "llama-cpp": "custom",
     }
+    # Extend with aliases declared in providers/*.py that aren't already mapped.
+    # This keeps providers/ as the single source for new aliases while the
+    # hardcoded dict above remains authoritative for existing ones.
+    try:
+        from providers import list_providers as _lp
+        for _pp in _lp():
+            for _alias in _pp.aliases:
+                if _alias not in _PROVIDER_ALIASES:
+                    _PROVIDER_ALIASES[_alias] = _pp.name
+    except Exception:
+        pass
     normalized = _PROVIDER_ALIASES.get(normalized, normalized)
 
     if normalized == "openrouter":

hermes_cli/config.py

@@ -1254,6 +1254,22 @@ OPTIONAL_ENV_VARS = {
         "category": "provider",
         "advanced": True,
     },
+    "GMI_API_KEY": {
+        "description": "GMI Cloud API key",
+        "prompt": "GMI Cloud API key",
+        "url": "https://www.gmicloud.ai/",
+        "password": True,
+        "category": "provider",
+        "advanced": True,
+    },
+    "GMI_BASE_URL": {
+        "description": "GMI Cloud base URL override",
+        "prompt": "GMI Cloud base URL (leave empty for default)",
+        "url": None,
+        "password": False,
+        "category": "provider",
+        "advanced": True,
+    },
     "MINIMAX_API_KEY": {
         "description": "MiniMax API key (international)",
         "prompt": "MiniMax API key",
@@ -4236,3 +4252,45 @@ def config_command(args):
         print("  hermes config path      Show config file path")
         print("  hermes config env-path  Show .env file path")
         sys.exit(1)
+
+
+# ── Profile-driven env var injection ─────────────────────────────────────────
+# Any provider registered in providers/ with auth_type="api_key" automatically
+# gets its env_vars exposed in OPTIONAL_ENV_VARS without editing this file.
+# Runs once at import time.
+
+_profile_env_vars_injected = False
+
+
+def _inject_profile_env_vars() -> None:
+    """Populate OPTIONAL_ENV_VARS from provider profiles not already listed.
+
+    Called once at module load time. Idempotent — repeated calls are no-ops.
+    """
+    global _profile_env_vars_injected
+    if _profile_env_vars_injected:
+        return
+    _profile_env_vars_injected = True
+    try:
+        from providers import list_providers
+        for _pp in list_providers():
+            if _pp.auth_type not in ("api_key",):
+                continue
+            for _var in _pp.env_vars:
+                if _var in OPTIONAL_ENV_VARS:
+                    continue
+                _is_key = not _var.endswith("_BASE_URL") and not _var.endswith("_URL")
+                OPTIONAL_ENV_VARS[_var] = {
+                    "description": f"{_pp.display_name or _pp.name} {'API key' if _is_key else 'base URL override'}",
+                    "prompt": f"{_pp.display_name or _pp.name} {'API key' if _is_key else 'base URL (leave empty for default)'}",
+                    "url": _pp.signup_url or None,
+                    "password": _is_key,
+                    "category": "provider",
+                    "advanced": True,
+                }
+    except Exception:
+        pass
+
+
+# Eagerly inject so that OPTIONAL_ENV_VARS is fully populated at import time.
+_inject_profile_env_vars()

hermes_cli/doctor.py

@@ -46,6 +46,7 @@ _PROVIDER_ENV_HINTS = (
     "Z_AI_API_KEY",
     "KIMI_API_KEY",
     "KIMI_CN_API_KEY",
+    "GMI_API_KEY",
     "MINIMAX_API_KEY",
     "MINIMAX_CN_API_KEY",
     "KILOCODE_API_KEY",
@@ -163,6 +164,84 @@ def _check_gateway_service_linger(issues: list[str]) -> None:
         check_warn("Could not verify systemd linger", f"({linger_detail})")
 
 
+_APIKEY_PROVIDERS_CACHE: list | None = None
+
+
+def _build_apikey_providers_list() -> list:
+    """Build the API-key provider health-check list once and cache it.
+
+    Tuple format: (name, env_vars, default_url, base_env, supports_models_endpoint)
+    Base list augmented with any ProviderProfile with auth_type="api_key" not
+    already present — adding providers/*.py is sufficient to get into doctor.
+    """
+    _static = [
+        ("Z.AI / GLM",      ("GLM_API_KEY", "ZAI_API_KEY", "Z_AI_API_KEY"), "https://api.z.ai/api/paas/v4/models", "GLM_BASE_URL", True),
+        ("Kimi / Moonshot",  ("KIMI_API_KEY",),                              "https://api.moonshot.ai/v1/models",   "KIMI_BASE_URL", True),
+        ("StepFun Step Plan", ("STEPFUN_API_KEY",),                          "https://api.stepfun.ai/step_plan/v1/models", "STEPFUN_BASE_URL", True),
+        ("Kimi / Moonshot (China)", ("KIMI_CN_API_KEY",),                    "https://api.moonshot.cn/v1/models",   None, True),
+        ("Arcee AI",         ("ARCEEAI_API_KEY",),                           "https://api.arcee.ai/api/v1/models",  "ARCEE_BASE_URL", True),
+        ("GMI Cloud",        ("GMI_API_KEY",),                               "https://api.gmi-serving.com/v1/models", "GMI_BASE_URL", True),
+        ("DeepSeek",         ("DEEPSEEK_API_KEY",),                          "https://api.deepseek.com/v1/models",  "DEEPSEEK_BASE_URL", True),
+        ("Hugging Face",     ("HF_TOKEN",),                                  "https://router.huggingface.co/v1/models", "HF_BASE_URL", True),
+        ("NVIDIA NIM",       ("NVIDIA_API_KEY",),                            "https://integrate.api.nvidia.com/v1/models", "NVIDIA_BASE_URL", True),
+        ("Alibaba/DashScope", ("DASHSCOPE_API_KEY",),                        "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/models", "DASHSCOPE_BASE_URL", True),
+        # MiniMax: the /anthropic endpoint doesn't support /models; use the /v1 surface.
+        ("MiniMax",          ("MINIMAX_API_KEY",),                           "https://api.minimax.io/v1/models",    "MINIMAX_BASE_URL", True),
+        ("MiniMax (China)",  ("MINIMAX_CN_API_KEY",),                        "https://api.minimaxi.com/v1/models",  "MINIMAX_CN_BASE_URL", True),
+        ("Vercel AI Gateway", ("AI_GATEWAY_API_KEY",),                       "https://ai-gateway.vercel.sh/v1/models", "AI_GATEWAY_BASE_URL", True),
+        ("Kilo Code",        ("KILOCODE_API_KEY",),                          "https://api.kilo.ai/api/gateway/models", "KILOCODE_BASE_URL", True),
+        ("OpenCode Zen",     ("OPENCODE_ZEN_API_KEY",),                      "https://opencode.ai/zen/v1/models",  "OPENCODE_ZEN_BASE_URL", True),
+        # OpenCode Go has no shared /models endpoint; skip the health check.
+        ("OpenCode Go",      ("OPENCODE_GO_API_KEY",),                       None,                                  "OPENCODE_GO_BASE_URL", False),
+    ]
+    _known_names = {t[0] for t in _static}
+    # Also index by profile canonical name so profiles without display_name
+    # don't create duplicate entries for providers already in the static list.
+    _known_canonical: set[str] = set()
+    _name_to_canonical = {
+        "Z.AI / GLM": "zai", "Kimi / Moonshot": "kimi-coding",
+        "StepFun Step Plan": "stepfun", "Kimi / Moonshot (China)": "kimi-coding-cn",
+        "Arcee AI": "arcee", "GMI Cloud": "gmi", "DeepSeek": "deepseek",
+        "Hugging Face": "huggingface", "NVIDIA NIM": "nvidia",
+        "Alibaba/DashScope": "alibaba", "MiniMax": "minimax",
+        "MiniMax (China)": "minimax-cn", "Vercel AI Gateway": "ai-gateway",
+        "Kilo Code": "kilocode", "OpenCode Zen": "opencode-zen",
+        "OpenCode Go": "opencode-go",
+    }
+    for _label, _canonical in _name_to_canonical.items():
+        _known_canonical.add(_canonical)
+    try:
+        from providers import list_providers
+        from providers.base import ProviderProfile as _PP
+        for _pp in list_providers():
+            if not isinstance(_pp, _PP) or _pp.auth_type != "api_key" or not _pp.env_vars:
+                continue
+            _label = _pp.display_name or _pp.name
+            if _label in _known_names or _pp.name in _known_canonical:
+                continue
+            # Separate API-key vars from base-URL override vars — the health-check
+            # loop sends the first found value as Authorization: Bearer, so a URL
+            # string must never be picked.
+            _key_vars = tuple(
+                v for v in _pp.env_vars
+                if not v.endswith("_BASE_URL") and not v.endswith("_URL")
+            )
+            _base_var = next(
+                (v for v in _pp.env_vars if v.endswith("_BASE_URL") or v.endswith("_URL")),
+                None,
+            )
+            if not _key_vars:
+                continue
+            _models_url = (
+                (_pp.models_url or (_pp.base_url.rstrip("/") + "/models"))
+                if _pp.base_url else None
+            )
+            _static.append((_label, _key_vars, _models_url, _base_var, True))
+    except Exception:
+        pass
+    return _static
+
+
 def run_doctor(args):
     """Run diagnostic checks."""
     should_fix = getattr(args, 'fix', False)
@@ -930,26 +1009,11 @@ def run_doctor(args):
 
     # -- API-key providers --
     # Tuple: (name, env_vars, default_url, base_env, supports_models_endpoint)
-    # If supports_models_endpoint is False, we skip the health check and just show "configured"
-    _apikey_providers = [
-        ("Z.AI / GLM",      ("GLM_API_KEY", "ZAI_API_KEY", "Z_AI_API_KEY"), "https://api.z.ai/api/paas/v4/models", "GLM_BASE_URL", True),
-        ("Kimi / Moonshot",  ("KIMI_API_KEY",),                              "https://api.moonshot.ai/v1/models",   "KIMI_BASE_URL", True),
-        ("StepFun Step Plan",   ("STEPFUN_API_KEY",),                           "https://api.stepfun.ai/step_plan/v1/models", "STEPFUN_BASE_URL", True),
-        ("Kimi / Moonshot (China)", ("KIMI_CN_API_KEY",),                    "https://api.moonshot.cn/v1/models",   None, True),
-        ("Arcee AI",         ("ARCEEAI_API_KEY",),                            "https://api.arcee.ai/api/v1/models",  "ARCEE_BASE_URL", True),
-        ("DeepSeek",         ("DEEPSEEK_API_KEY",),                           "https://api.deepseek.com/v1/models",  "DEEPSEEK_BASE_URL", True),
-        ("Hugging Face",     ("HF_TOKEN",),                                   "https://router.huggingface.co/v1/models", "HF_BASE_URL", True),
-        ("NVIDIA NIM",       ("NVIDIA_API_KEY",),                             "https://integrate.api.nvidia.com/v1/models", "NVIDIA_BASE_URL", True),
-        ("Alibaba/DashScope", ("DASHSCOPE_API_KEY",),                         "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/models", "DASHSCOPE_BASE_URL", True),
-        # MiniMax: the /anthropic endpoint doesn't support /models, but the /v1 endpoint does.
-        ("MiniMax",          ("MINIMAX_API_KEY",),                            "https://api.minimax.io/v1/models",    "MINIMAX_BASE_URL", True),
-        ("MiniMax (China)",  ("MINIMAX_CN_API_KEY",),                         "https://api.minimaxi.com/v1/models",  "MINIMAX_CN_BASE_URL", True),
-        ("Vercel AI Gateway",       ("AI_GATEWAY_API_KEY",),                          "https://ai-gateway.vercel.sh/v1/models", "AI_GATEWAY_BASE_URL", True),
-        ("Kilo Code",        ("KILOCODE_API_KEY",),                            "https://api.kilo.ai/api/gateway/models",  "KILOCODE_BASE_URL", True),
-        ("OpenCode Zen",     ("OPENCODE_ZEN_API_KEY",),                        "https://opencode.ai/zen/v1/models",  "OPENCODE_ZEN_BASE_URL", True),
-        # OpenCode Go has no shared /models endpoint; skip the health check.
-        ("OpenCode Go",      ("OPENCODE_GO_API_KEY",),                         None,                                  "OPENCODE_GO_BASE_URL", False),
-    ]
+    # Cached at module level after first build — profiles auto-extend it.
+    global _APIKEY_PROVIDERS_CACHE
+    if _APIKEY_PROVIDERS_CACHE is None:
+        _APIKEY_PROVIDERS_CACHE = _build_apikey_providers_list()
+    _apikey_providers = _APIKEY_PROVIDERS_CACHE
     for _pname, _env_vars, _default_url, _base_env, _supports_health_check in _apikey_providers:
         _key = ""
         for _ev in _env_vars:

hermes_cli/main.py

@@ -829,8 +829,29 @@ def _print_tui_exit_summary(session_id: Optional[str], active_session_file: Opti
     )
 
 
+_NPM_LOCK_RUNTIME_KEYS = frozenset({"ideallyInert"})
+
+
 def _tui_need_npm_install(root: Path) -> bool:
-    """True when @hermes/ink is missing or node_modules is behind package-lock.json (post-pull)."""
+    """True when @hermes/ink is missing or node_modules is behind package-lock.json.
+
+    Compares ``package-lock.json`` against ``node_modules/.package-lock.json``
+    (npm's hidden lockfile) by **content**, not mtime: git checkouts and npm
+    rewrites can bump the root lockfile's timestamp even when installed deps
+    already match, which used to trigger a spurious "Installing TUI
+    dependencies" on every launch.
+
+    For each entry in the root lock's ``packages`` map:
+      - missing from hidden lock → reinstall (unless the entry is marked
+        ``optional`` or ``peer``, which npm may intentionally skip per platform)
+      - present but with differing fields (excluding npm-written runtime
+        annotations like ``ideallyInert``) → reinstall
+
+    Extra entries that exist only in the hidden lock are ignored — stale
+    transitives left over from a removed dependency don't break runtime and
+    we'd rather not force a reinstall for them. Falls back to mtime
+    comparison if either lockfile is unparseable.
+    """
     ink = root / "node_modules" / "@hermes" / "ink" / "package.json"
     if not ink.is_file():
         return True
@@ -840,7 +861,35 @@ def _tui_need_npm_install(root: Path) -> bool:
     marker = root / "node_modules" / ".package-lock.json"
     if not marker.is_file():
         return True
-    return lock.stat().st_mtime > marker.stat().st_mtime
+
+    # Compare lockfile contents, not mtimes: git checkouts and npm rewrites
+    # can bump the root lockfile timestamp even when installed deps already
+    # match. Fall back to mtime when either file is unparseable.
+    try:
+        wanted = json.loads(lock.read_text(encoding="utf-8")).get("packages") or {}
+        installed = json.loads(marker.read_text(encoding="utf-8")).get("packages") or {}
+    except (OSError, UnicodeDecodeError, json.JSONDecodeError):
+        return lock.stat().st_mtime > marker.stat().st_mtime
+
+    def comparable(pkg: dict) -> dict:
+        return {k: v for k, v in pkg.items() if k not in _NPM_LOCK_RUNTIME_KEYS}
+
+    for name, pkg in wanted.items():
+        if not name:
+            continue
+
+        if not isinstance(pkg, dict):
+            continue
+
+        if name not in installed:
+            if pkg.get("optional") or pkg.get("peer"):
+                continue
+            return True
+
+        if isinstance(installed[name], dict) and comparable(pkg) != comparable(installed[name]):
+            return True
+
+    return False
 
 
 def _find_bundled_tui(tui_dir: Path) -> Optional[Path]:
@@ -1479,6 +1528,21 @@ def cmd_model(args):
     select_provider_and_model(args=args)
 
 
+def _is_profile_api_key_provider(provider_id: str) -> bool:
+    """Return True when provider_id maps to a profile with auth_type='api_key'.
+
+    Used as a catch-all in select_provider_and_model() so that new providers
+    declared in providers/*.py automatically dispatch to _model_flow_api_key_provider
+    without requiring an explicit elif branch here.
+    """
+    try:
+        from providers import get_provider_profile
+        _p = get_provider_profile(provider_id)
+        return _p is not None and _p.auth_type == "api_key"
+    except Exception:
+        return False
+
+
 def select_provider_and_model(args=None):
     """Core provider selection + model picking logic.
 
@@ -1768,9 +1832,10 @@ def select_provider_and_model(args=None):
         "huggingface",
         "xiaomi",
         "arcee",
+        "gmi",
         "nvidia",
         "ollama-cloud",
-    ):
+    ) or _is_profile_api_key_provider(selected_provider):
         _model_flow_api_key_provider(config, selected_provider, current_model)
 
     # ── Post-switch cleanup: clear stale OPENAI_BASE_URL ──────────────
@@ -7568,6 +7633,22 @@ def cmd_logs(args):
     )
 
 
+def _build_provider_choices() -> list[str]:
+    """Build the --provider choices list from CANONICAL_PROVIDERS + 'auto'."""
+    try:
+        from hermes_cli.models import CANONICAL_PROVIDERS as _cp
+        return ["auto"] + [p.slug for p in _cp]
+    except Exception:
+        # Fallback: static list guarantees the CLI always works
+        return [
+            "auto", "openrouter", "nous", "openai-codex", "copilot-acp", "copilot",
+            "anthropic", "gemini", "google-gemini-cli", "xai", "bedrock", "azure-foundry",
+            "ollama-cloud", "huggingface", "zai", "kimi-coding", "kimi-coding-cn",
+            "stepfun", "minimax", "minimax-cn", "kilocode", "xiaomi", "arcee",
+            "nvidia", "deepseek", "alibaba", "qwen-oauth", "opencode-zen", "opencode-go",
+        ]
+
+
 def main():
     """Main entry point for hermes CLI."""
     parser = argparse.ArgumentParser(
@@ -7761,29 +7842,7 @@ For more help on a command:
     )
     chat_parser.add_argument(
         "--provider",
-        choices=[
-            "auto",
-            "openrouter",
-            "nous",
-            "openai-codex",
-            "copilot-acp",
-            "copilot",
-            "anthropic",
-            "gemini",
-            "xai",
-            "ollama-cloud",
-            "huggingface",
-            "zai",
-            "kimi-coding",
-            "kimi-coding-cn",
-            "stepfun",
-            "minimax",
-            "minimax-cn",
-            "kilocode",
-            "xiaomi",
-            "arcee",
-            "nvidia",
-        ],
+        choices=_build_provider_choices(),
         default=None,
         help="Inference provider (default: auto)",
     )

hermes_cli/models.py

@@ -278,6 +278,14 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "trinity-large-preview",
         "trinity-mini",
     ],
+    "gmi": [
+        "zai-org/GLM-5.1-FP8",
+        "deepseek-ai/DeepSeek-V3.2",
+        "moonshotai/Kimi-K2.5",
+        "google/gemini-3.1-flash-lite-preview",
+        "anthropic/claude-sonnet-4.6",
+        "openai/gpt-5.4",
+    ],
     "opencode-zen": [
         "kimi-k2.5",
         "gpt-5.4-pro",
@@ -709,7 +717,6 @@ class ProviderEntry(NamedTuple):
     label: str
     tui_desc: str   # detailed description for `hermes model` TUI
 
-
 CANONICAL_PROVIDERS: list[ProviderEntry] = [
     ProviderEntry("nous",           "Nous Portal",              "Nous Portal (Nous Research subscription)"),
     ProviderEntry("openrouter",     "OpenRouter",               "OpenRouter (100+ models, pay-per-use)"),
@@ -735,6 +742,7 @@ CANONICAL_PROVIDERS: list[ProviderEntry] = [
     ProviderEntry("alibaba",        "Alibaba Cloud (DashScope)","Alibaba Cloud / DashScope Coding (Qwen + multi-provider)"),
     ProviderEntry("ollama-cloud",   "Ollama Cloud",             "Ollama Cloud (cloud-hosted open models — ollama.com)"),
     ProviderEntry("arcee",          "Arcee AI",                 "Arcee AI (Trinity models — direct API)"),
+    ProviderEntry("gmi",            "GMI Cloud",                "GMI Cloud (multi-model direct API)"),
     ProviderEntry("kilocode",       "Kilo Code",                "Kilo Code (Kilo Gateway API)"),
     ProviderEntry("opencode-zen",   "OpenCode Zen",             "OpenCode Zen (35+ curated models, pay-as-you-go)"),
     ProviderEntry("opencode-go",    "OpenCode Go",              "OpenCode Go (open models, $10/month subscription)"),
@@ -742,6 +750,25 @@ CANONICAL_PROVIDERS: list[ProviderEntry] = [
     ProviderEntry("azure-foundry",  "Azure Foundry",            "Azure Foundry (OpenAI-style or Anthropic-style endpoint — your Azure AI deployment)"),
 ]
 
+# Auto-extend CANONICAL_PROVIDERS with any provider registered in providers/
+# that is not already in the list above.  Adding providers/*.py is sufficient
+# to expose a new provider in the model picker, /model, and all downstream
+# consumers — no edits to this file needed.
+_canonical_slugs = {p.slug for p in CANONICAL_PROVIDERS}
+try:
+    from providers import list_providers as _list_providers_for_canonical
+    for _pp in _list_providers_for_canonical():
+        if _pp.name in _canonical_slugs:
+            continue
+        if _pp.auth_type in ("oauth_device_code", "oauth_external", "external_process", "aws_sdk", "copilot"):
+            continue  # non-api-key flows need bespoke picker UX; skip auto-inject
+        _label = _pp.display_name or _pp.name
+        _desc = _pp.description or f"{_label} (direct API)"
+        CANONICAL_PROVIDERS.append(ProviderEntry(_pp.name, _label, _desc))
+        _canonical_slugs.add(_pp.name)
+except Exception:
+    pass
+
 # Derived dicts — used throughout the codebase
 _PROVIDER_LABELS = {p.slug: p.label for p in CANONICAL_PROVIDERS}
 _PROVIDER_LABELS["custom"] = "Custom endpoint"  # special case: not a named provider
@@ -769,6 +796,8 @@ _PROVIDER_ALIASES = {
     "stepfun-coding-plan": "stepfun",
     "arcee-ai": "arcee",
     "arceeai": "arcee",
+    "gmi-cloud": "gmi",
+    "gmicloud": "gmi",
     "minimax-china": "minimax-cn",
     "minimax_cn": "minimax-cn",
     "claude": "anthropic",
@@ -1849,6 +1878,19 @@ def provider_model_ids(provider: Optional[str], *, force_refresh: bool = False)
                     return live
             except Exception:
                 pass
+    if normalized == "gmi":
+        try:
+            from hermes_cli.auth import resolve_api_key_provider_credentials
+
+            creds = resolve_api_key_provider_credentials("gmi")
+            api_key = str(creds.get("api_key") or "").strip()
+            base_url = str(creds.get("base_url") or "").strip()
+            if api_key and base_url:
+                live = fetch_api_models(api_key, base_url)
+                if live:
+                    return live
+        except Exception:
+            pass
     if normalized == "custom":
         base_url = _get_custom_base_url()
         if base_url:
@@ -1861,6 +1903,34 @@ def provider_model_ids(provider: Optional[str], *, force_refresh: bool = False)
             live = fetch_api_models(api_key, base_url)
             if live:
                 return live
+
+    # ── Profile-based generic live fetch (all simple api-key providers) ──
+    # Handles any provider registered in providers/ with auth_type="api_key".
+    # Replaces per-provider copy-paste blocks (stepfun, gmi, zai, etc.).
+    try:
+        from providers import get_provider_profile
+        from hermes_cli.auth import resolve_api_key_provider_credentials
+
+        _p = get_provider_profile(normalized)
+        if _p and _p.auth_type == "api_key" and _p.base_url:
+            try:
+                creds = resolve_api_key_provider_credentials(normalized)
+                api_key = str(creds.get("api_key") or "").strip()
+                base_url = str(creds.get("base_url") or "").strip()
+            except Exception:
+                api_key, base_url = "", _p.base_url
+            if not base_url:
+                base_url = _p.base_url
+            if api_key:
+                live = _p.fetch_models(api_key=api_key)
+                if live:
+                    return live
+            # Use profile's fallback_models if defined
+            if _p.fallback_models:
+                return list(_p.fallback_models)
+    except Exception:
+        pass
+
     curated_static = list(_PROVIDER_MODELS.get(normalized, []))
     if normalized in _MODELS_DEV_PREFERRED:
         return _merge_with_models_dev(normalized, curated_static)

hermes_cli/plugins.py

@@ -79,6 +79,20 @@ VALID_HOOKS: Set[str] = {
     #   {"action": "allow"}  /  None             -> normal dispatch
     # Kwargs: event: MessageEvent, gateway: GatewayRunner, session_store.
     "pre_gateway_dispatch",
+    # Approval lifecycle hooks. Fired by tools/approval.py when a dangerous
+    # command needs user approval -- fires BOTH for CLI-interactive prompts
+    # and for gateway/ACP approvals (Telegram, Discord, Slack, TUI, etc.).
+    # Observers only: return values are ignored. Plugins cannot veto or
+    # pre-answer an approval from these hooks (use pre_tool_call to block
+    # a tool before it reaches approval).
+    #
+    # Kwargs for pre_approval_request:
+    #   command: str, description: str, pattern_key: str, pattern_keys: list[str],
+    #   session_key: str, surface: "cli" | "gateway"
+    # Kwargs for post_approval_response: same as above plus
+    #   choice: "once" | "session" | "always" | "deny" | "timeout"
+    "pre_approval_request",
+    "post_approval_response",
 }
 
 ENTRY_POINTS_GROUP = "hermes_agent.plugins"

hermes_cli/providers.py

@@ -163,6 +163,12 @@ HERMES_OVERLAYS: Dict[str, HermesOverlay] = {
         base_url_override="https://api.arcee.ai/api/v1",
         base_url_env_var="ARCEE_BASE_URL",
     ),
+    "gmi": HermesOverlay(
+        transport="openai_chat",
+        extra_env_vars=("GMI_API_KEY",),
+        base_url_override="https://api.gmi-serving.com/v1",
+        base_url_env_var="GMI_BASE_URL",
+    ),
     "ollama-cloud": HermesOverlay(
         transport="openai_chat",
         base_url_env_var="OLLAMA_BASE_URL",
@@ -297,6 +303,10 @@ ALIASES: Dict[str, str] = {
     "arcee-ai": "arcee",
     "arceeai": "arcee",
 
+    # gmi
+    "gmi-cloud": "gmi",
+    "gmicloud": "gmi",
+
     # Local server aliases → virtual "local" concept (resolved via user config)
     "lmstudio": "lmstudio",
     "lm-studio": "lmstudio",
@@ -319,6 +329,7 @@ _LABEL_OVERRIDES: Dict[str, str] = {
     "copilot-acp": "GitHub Copilot ACP",
     "stepfun": "StepFun Step Plan",
     "xiaomi": "Xiaomi MiMo",
+    "gmi": "GMI Cloud",
     "local": "Local endpoint",
     "bedrock": "AWS Bedrock",
     "ollama-cloud": "Ollama Cloud",

hermes_cli/runtime_provider.py

@@ -214,10 +214,6 @@ def _resolve_runtime_from_pool_entry(
         base_url = cfg_base_url or base_url or "https://api.anthropic.com"
     elif provider == "openrouter":
         base_url = base_url or OPENROUTER_BASE_URL
-    elif provider == "xai":
-        api_mode = "codex_responses"
-    elif provider == "nous":
-        api_mode = "chat_completions"
     elif provider == "copilot":
         api_mode = _copilot_runtime_api_mode(model_cfg, getattr(entry, "runtime_api_key", ""))
         base_url = base_url or PROVIDER_REGISTRY["copilot"].inference_base_url
@@ -249,6 +245,14 @@ def _resolve_runtime_from_pool_entry(
             base_url = re.sub(r"/v1/?$", "", base_url)
     else:
         configured_provider = str(model_cfg.get("provider") or "").strip().lower()
+        # Use profile api_mode for all other known providers
+        try:
+            from providers import get_provider_profile
+            _p = get_provider_profile(provider)
+            if _p and _p.api_mode:
+                api_mode = _p.api_mode
+        except Exception:
+            pass
         # Honour model.base_url from config.yaml when the configured provider
         # matches this provider — same pattern as the Anthropic branch above.
         # Only override when the pool entry has no explicit base_url (i.e. it
@@ -266,12 +270,21 @@ def _resolve_runtime_from_pool_entry(
             from hermes_cli.models import opencode_model_api_mode
             api_mode = opencode_model_api_mode(provider, effective_model)
         else:
-            # Auto-detect Anthropic-compatible endpoints (/anthropic suffix,
-            # Kimi /coding, api.openai.com → codex_responses, api.x.ai →
-            # codex_responses).
-            detected = _detect_api_mode_for_url(base_url)
-            if detected:
-                api_mode = detected
+            # Try profile api_mode first, then auto-detect from URL
+            try:
+                from providers import get_provider_profile
+                _p = get_provider_profile(provider)
+                if _p and _p.api_mode:
+                    api_mode = _p.api_mode
+            except Exception:
+                pass
+            if api_mode == "chat_completions":
+                # Auto-detect Anthropic-compatible endpoints (/anthropic suffix,
+                # Kimi /coding, api.openai.com → codex_responses, api.x.ai →
+                # codex_responses).
+                detected = _detect_api_mode_for_url(base_url)
+                if detected:
+                    api_mode = detected
 
     # OpenCode base URLs end with /v1 for OpenAI-compatible models, but the
     # Anthropic SDK prepends its own /v1/messages to the base_url.  Strip the

hermes_cli/web_server.py

@@ -2212,7 +2212,7 @@ async def get_usage_analytics(days: int = 30):
         cutoff = time.time() - (days * 86400)
         cur = db._conn.execute("""
             SELECT date(started_at, 'unixepoch') as day,
-                   SUM(input_tokens + COALESCE(cache_read_tokens, 0) + COALESCE(cache_write_tokens, 0)) as input_tokens,
+                   SUM(input_tokens) as input_tokens,
                    SUM(output_tokens) as output_tokens,
                    SUM(cache_read_tokens) as cache_read_tokens,
                    SUM(reasoning_tokens) as reasoning_tokens,
@@ -2227,18 +2227,18 @@ async def get_usage_analytics(days: int = 30):
 
         cur2 = db._conn.execute("""
             SELECT model,
-                   SUM(input_tokens + COALESCE(cache_read_tokens, 0) + COALESCE(cache_write_tokens, 0)) as input_tokens,
+                   SUM(input_tokens) as input_tokens,
                    SUM(output_tokens) as output_tokens,
                    COALESCE(SUM(estimated_cost_usd), 0) as estimated_cost,
                    COUNT(*) as sessions,
                    SUM(COALESCE(api_call_count, 0)) as api_calls
             FROM sessions WHERE started_at > ? AND model IS NOT NULL
-            GROUP BY model ORDER BY SUM(input_tokens + COALESCE(cache_read_tokens, 0) + COALESCE(cache_write_tokens, 0)) + SUM(output_tokens) DESC
+            GROUP BY model ORDER BY SUM(input_tokens) + SUM(output_tokens) DESC
         """, (cutoff,))
         by_model = [dict(r) for r in cur2.fetchall()]
 
         cur3 = db._conn.execute("""
-            SELECT SUM(input_tokens + COALESCE(cache_read_tokens, 0) + COALESCE(cache_write_tokens, 0)) as total_input,
+            SELECT SUM(input_tokens) as total_input,
                    SUM(output_tokens) as total_output,
                    SUM(cache_read_tokens) as total_cache_read,
                    SUM(reasoning_tokens) as total_reasoning,

hermes_state.py

@@ -22,6 +22,8 @@ import sqlite3
 import threading
 import time
 from pathlib import Path
+
+from agent.memory_manager import sanitize_context
 from hermes_constants import get_hermes_home
 from typing import Any, Callable, Dict, List, Optional, TypeVar
 
@@ -31,7 +33,7 @@ T = TypeVar("T")
 
 DEFAULT_DB_PATH = get_hermes_home() / "state.db"
 
-SCHEMA_VERSION = 9
+SCHEMA_VERSION = 10
 
 SCHEMA_SQL = """
 CREATE TABLE IF NOT EXISTS schema_version (
@@ -119,6 +121,32 @@ CREATE TRIGGER IF NOT EXISTS messages_fts_update AFTER UPDATE ON messages BEGIN
 END;
 """
 
+# Trigram FTS5 table for CJK substring search.  The default unicode61
+# tokenizer splits CJK characters into individual tokens, breaking phrase
+# matching.  The trigram tokenizer creates overlapping 3-byte sequences so
+# substring queries work natively for any script (CJK, Thai, etc.).
+FTS_TRIGRAM_SQL = """
+CREATE VIRTUAL TABLE IF NOT EXISTS messages_fts_trigram USING fts5(
+    content,
+    content=messages,
+    content_rowid=id,
+    tokenize='trigram'
+);
+
+CREATE TRIGGER IF NOT EXISTS messages_fts_trigram_insert AFTER INSERT ON messages BEGIN
+    INSERT INTO messages_fts_trigram(rowid, content) VALUES (new.id, new.content);
+END;
+
+CREATE TRIGGER IF NOT EXISTS messages_fts_trigram_delete AFTER DELETE ON messages BEGIN
+    INSERT INTO messages_fts_trigram(messages_fts_trigram, rowid, content) VALUES('delete', old.id, old.content);
+END;
+
+CREATE TRIGGER IF NOT EXISTS messages_fts_trigram_update AFTER UPDATE ON messages BEGIN
+    INSERT INTO messages_fts_trigram(messages_fts_trigram, rowid, content) VALUES('delete', old.id, old.content);
+    INSERT INTO messages_fts_trigram(rowid, content) VALUES (new.id, new.content);
+END;
+"""
+
 
 class SessionDB:
     """
@@ -366,6 +394,18 @@ class SessionDB:
                 except sqlite3.OperationalError:
                     pass  # Column already exists
                 cursor.execute("UPDATE schema_version SET version = 9")
+            if current_version < 10:
+                # v10: trigram FTS5 table for CJK/substring search.
+                # Created via FTS_TRIGRAM_SQL below; backfill existing messages.
+                try:
+                    cursor.execute("SELECT * FROM messages_fts_trigram LIMIT 0")
+                except sqlite3.OperationalError:
+                    cursor.executescript(FTS_TRIGRAM_SQL)
+                    cursor.execute(
+                        "INSERT INTO messages_fts_trigram(rowid, content) "
+                        "SELECT id, content FROM messages WHERE content IS NOT NULL"
+                    )
+                cursor.execute("UPDATE schema_version SET version = 10")
 
         # Unique title index — always ensure it exists (safe to run after migrations
         # since the title column is guaranteed to exist at this point)
@@ -383,6 +423,12 @@ class SessionDB:
         except sqlite3.OperationalError:
             cursor.executescript(FTS_SQL)
 
+        # Trigram FTS5 for CJK/substring search
+        try:
+            cursor.execute("SELECT * FROM messages_fts_trigram LIMIT 0")
+        except sqlite3.OperationalError:
+            cursor.executescript(FTS_TRIGRAM_SQL)
+
         self._conn.commit()
 
     # =========================================================================
@@ -1155,7 +1201,10 @@ class SessionDB:
 
         messages = []
         for row in rows:
-            msg = {"role": row["role"], "content": row["content"]}
+            content = row["content"]
+            if row["role"] in {"user", "assistant"} and isinstance(content, str):
+                content = sanitize_context(content).strip()
+            msg = {"role": row["role"], "content": content}
             if row["tool_call_id"]:
                 msg["tool_call_id"] = row["tool_call_id"]
             if row["tool_name"]:
@@ -1291,6 +1340,16 @@ class SessionDB:
         return sanitized.strip()
 
 
+    @staticmethod
+    def _is_cjk_codepoint(cp: int) -> bool:
+        return (0x4E00 <= cp <= 0x9FFF or    # CJK Unified Ideographs
+                0x3400 <= cp <= 0x4DBF or    # CJK Extension A
+                0x20000 <= cp <= 0x2A6DF or  # CJK Extension B
+                0x3000 <= cp <= 0x303F or    # CJK Symbols
+                0x3040 <= cp <= 0x309F or    # Hiragana
+                0x30A0 <= cp <= 0x30FF or    # Katakana
+                0xAC00 <= cp <= 0xD7AF)      # Hangul Syllables
+
     @staticmethod
     def _contains_cjk(text: str) -> bool:
         """Check if text contains CJK (Chinese, Japanese, Korean) characters."""
@@ -1306,6 +1365,11 @@ class SessionDB:
                 return True
         return False
 
+    @classmethod
+    def _count_cjk(cls, text: str) -> int:
+        """Count CJK characters in text."""
+        return sum(1 for ch in text if cls._is_cjk_codepoint(ord(ch)))
+
     def search_messages(
         self,
         query: str,
@@ -1376,52 +1440,113 @@ class SessionDB:
             LIMIT ? OFFSET ?
         """
 
-        with self._lock:
-            try:
-                cursor = self._conn.execute(sql, params)
-            except sqlite3.OperationalError:
-                # FTS5 query syntax error despite sanitization — return empty
-                # unless query contains CJK (fall back to LIKE below)
-                if not self._contains_cjk(query):
-                    return []
-                matches = []
-            else:
-                matches = [dict(row) for row in cursor.fetchall()]
-
-        # LIKE fallback for CJK queries: FTS5 default tokenizer splits CJK
-        # characters individually, causing multi-character queries to fail.
-        if not matches and self._contains_cjk(query):
+        # CJK queries bypass the unicode61 FTS5 table.  The default tokenizer
+        # splits CJK characters into individual tokens, so "大别山项目" becomes
+        # "大 AND 别 AND 山 AND 项 AND 目" — producing false positives and
+        # missing exact phrase matches.
+        #
+        # For queries with 3+ CJK characters, we use the trigram FTS5 table
+        # (indexed substring matching with ranking and snippets).  For shorter
+        # CJK queries (1-2 chars), trigram can't match (it needs ≥9 UTF-8
+        # bytes = 3 CJK chars), so we fall back to LIKE.
+        is_cjk = self._contains_cjk(query)
+        if is_cjk:
             raw_query = query.strip('"').strip()
-            like_where = ["m.content LIKE ?"]
-            like_params: list = [f"%{raw_query}%"]
-            if source_filter is not None:
-                like_where.append(f"s.source IN ({','.join('?' for _ in source_filter)})")
-                like_params.extend(source_filter)
-            if exclude_sources is not None:
-                like_where.append(f"s.source NOT IN ({','.join('?' for _ in exclude_sources)})")
-                like_params.extend(exclude_sources)
-            if role_filter:
-                like_where.append(f"m.role IN ({','.join('?' for _ in role_filter)})")
-                like_params.extend(role_filter)
-            like_sql = f"""
-                SELECT m.id, m.session_id, m.role,
-                       substr(m.content,
-                              max(1, instr(m.content, ?) - 40),
-                              120) AS snippet,
-                       m.content, m.timestamp, m.tool_name,
-                       s.source, s.model, s.started_at AS session_started
-                FROM messages m
-                JOIN sessions s ON s.id = m.session_id
-                WHERE {' AND '.join(like_where)}
-                ORDER BY m.timestamp DESC
-                LIMIT ? OFFSET ?
-            """
-            like_params.extend([limit, offset])
-            # instr() parameter goes first in the bound list
-            like_params = [raw_query] + like_params
+            cjk_count = self._count_cjk(raw_query)
+
+            if cjk_count >= 3:
+                # Trigram FTS5 path — quote each non-operator token to handle
+                # FTS5 special chars (%, *, etc.) while preserving boolean
+                # operators (AND, OR, NOT) for multi-term queries.
+                tokens = raw_query.split()
+                parts = []
+                for tok in tokens:
+                    if tok.upper() in ("AND", "OR", "NOT"):
+                        parts.append(tok)
+                    else:
+                        parts.append('"' + tok.replace('"', '""') + '"')
+                trigram_query = " ".join(parts)
+                tri_where = ["messages_fts_trigram MATCH ?"]
+                tri_params: list = [trigram_query]
+                if source_filter is not None:
+                    tri_where.append(f"s.source IN ({','.join('?' for _ in source_filter)})")
+                    tri_params.extend(source_filter)
+                if exclude_sources is not None:
+                    tri_where.append(f"s.source NOT IN ({','.join('?' for _ in exclude_sources)})")
+                    tri_params.extend(exclude_sources)
+                if role_filter:
+                    tri_where.append(f"m.role IN ({','.join('?' for _ in role_filter)})")
+                    tri_params.extend(role_filter)
+                tri_sql = f"""
+                    SELECT
+                        m.id,
+                        m.session_id,
+                        m.role,
+                        snippet(messages_fts_trigram, 0, '>>>', '<<<', '...', 40) AS snippet,
+                        m.content,
+                        m.timestamp,
+                        m.tool_name,
+                        s.source,
+                        s.model,
+                        s.started_at AS session_started
+                    FROM messages_fts_trigram
+                    JOIN messages m ON m.id = messages_fts_trigram.rowid
+                    JOIN sessions s ON s.id = m.session_id
+                    WHERE {' AND '.join(tri_where)}
+                    ORDER BY rank
+                    LIMIT ? OFFSET ?
+                """
+                tri_params.extend([limit, offset])
+                with self._lock:
+                    try:
+                        tri_cursor = self._conn.execute(tri_sql, tri_params)
+                    except sqlite3.OperationalError:
+                        matches = []
+                    else:
+                        matches = [dict(row) for row in tri_cursor.fetchall()]
+            else:
+                # Short CJK query (1-2 chars) — trigram needs ≥3 CJK chars.
+                # Fall back to LIKE substring search.
+                escaped = raw_query.replace("\\", "\\\\").replace("%", "\\%").replace("_", "\\_")
+                like_where = ["m.content LIKE ? ESCAPE '\\'"]
+                like_params: list = [f"%{escaped}%"]
+                if source_filter is not None:
+                    like_where.append(f"s.source IN ({','.join('?' for _ in source_filter)})")
+                    like_params.extend(source_filter)
+                if exclude_sources is not None:
+                    like_where.append(f"s.source NOT IN ({','.join('?' for _ in exclude_sources)})")
+                    like_params.extend(exclude_sources)
+                if role_filter:
+                    like_where.append(f"m.role IN ({','.join('?' for _ in role_filter)})")
+                    like_params.extend(role_filter)
+                like_sql = f"""
+                    SELECT m.id, m.session_id, m.role,
+                           substr(m.content,
+                                  max(1, instr(m.content, ?) - 40),
+                                  120) AS snippet,
+                           m.content, m.timestamp, m.tool_name,
+                           s.source, s.model, s.started_at AS session_started
+                    FROM messages m
+                    JOIN sessions s ON s.id = m.session_id
+                    WHERE {' AND '.join(like_where)}
+                    ORDER BY m.timestamp DESC
+                    LIMIT ? OFFSET ?
+                """
+                like_params.extend([limit, offset])
+                # instr() parameter goes first in the bound list
+                like_params = [raw_query] + like_params
+                with self._lock:
+                    like_cursor = self._conn.execute(like_sql, like_params)
+                    matches = [dict(row) for row in like_cursor.fetchall()]
+        else:
             with self._lock:
-                like_cursor = self._conn.execute(like_sql, like_params)
-                matches = [dict(row) for row in like_cursor.fetchall()]
+                try:
+                    cursor = self._conn.execute(sql, params)
+                except sqlite3.OperationalError:
+                    # FTS5 query syntax error despite sanitization — return empty
+                    return []
+                else:
+                    matches = [dict(row) for row in cursor.fetchall()]
 
         # Add surrounding context (1 message before + after each match).
         # Done outside the lock so we don't hold it across N sequential queries.

nix/checks.nix

@@ -7,9 +7,7 @@
   perSystem = { pkgs, system, lib, ... }:
     let
       hermes-agent = inputs.self.packages.${system}.default;
-      hermesVenv = pkgs.callPackage ./python.nix {
-        inherit (inputs) uv2nix pyproject-nix pyproject-build-systems;
-      };
+      hermesVenv = hermes-agent.hermesVenv;
 
       configMergeScript = pkgs.callPackage ./configMergeScript.nix { };
 
@@ -193,6 +191,35 @@ json.dump(sorted(leaf_paths(DEFAULT_CONFIG)), sys.stdout, indent=2)
           echo "ok" > $out/result
         '';
 
+        # Verify extraPythonPackages PYTHONPATH injection
+        extra-python-packages = let
+          testPkg = pkgs.python312Packages.pyfiglet;
+          hermesWithExtra = hermes-agent.override {
+            extraPythonPackages = [ testPkg ];
+          };
+        in pkgs.runCommand "hermes-extra-python-packages" { } ''
+          set -e
+          echo "=== Checking extraPythonPackages PYTHONPATH injection ==="
+
+          grep -q "PYTHONPATH" ${hermesWithExtra}/bin/hermes || \
+            (echo "FAIL: PYTHONPATH not in wrapper"; exit 1)
+          echo "PASS: PYTHONPATH present in wrapper"
+
+          grep -q "${testPkg}" ${hermesWithExtra}/bin/hermes || \
+            (echo "FAIL: test package path not in PYTHONPATH"; exit 1)
+          echo "PASS: test package path found in wrapper"
+
+          echo "=== Checking base package has no PYTHONPATH ==="
+          if grep -q "PYTHONPATH" ${hermes-agent}/bin/hermes; then
+            echo "FAIL: base package should not have PYTHONPATH"; exit 1
+          fi
+          echo "PASS: base package clean"
+
+          echo "=== All extraPythonPackages checks passed ==="
+          mkdir -p $out
+          echo "ok" > $out/result
+        '';
+
         # ── Config merge + round-trip test ────────────────────────────────
         # Tests the merge script (Nix activation behavior) across 7
         # scenarios, then verifies Python's load_config() reads correctly.

nix/hermes-agent.nix

@@ -0,0 +1,186 @@
+# nix/hermes-agent.nix — Overridable Hermes Agent package
+#
+# callPackage auto-wires nixpkgs args; flake inputs are passed explicitly.
+# Users override via: pkgs.hermes-agent.override { extraPythonPackages = [...]; }
+{
+  lib,
+  stdenv,
+  makeWrapper,
+  callPackage,
+  python312,
+  nodejs_22,
+  ripgrep,
+  git,
+  openssh,
+  ffmpeg,
+  tirith,
+  # Flake inputs — passed explicitly by packages.nix and overlays.nix
+  uv2nix,
+  pyproject-nix,
+  pyproject-build-systems,
+  npm-lockfile-fix,
+  # Overridable parameters
+  extraPythonPackages ? [ ],
+}:
+let
+  hermesVenv = callPackage ./python.nix {
+    inherit uv2nix pyproject-nix pyproject-build-systems;
+  };
+
+  hermesNpmLib = callPackage ./lib.nix {
+    inherit npm-lockfile-fix;
+  };
+
+  hermesTui = callPackage ./tui.nix {
+    inherit hermesNpmLib;
+  };
+
+  hermesWeb = callPackage ./web.nix {
+    inherit hermesNpmLib;
+  };
+
+  bundledSkills = lib.cleanSourceWith {
+    src = ../skills;
+    filter = path: _type: !(lib.hasInfix "/index-cache/" path);
+  };
+
+  runtimeDeps = [
+    nodejs_22
+    ripgrep
+    git
+    openssh
+    ffmpeg
+    tirith
+  ];
+
+  runtimePath = lib.makeBinPath runtimeDeps;
+
+  sitePackagesPath = python312.sitePackages;
+
+  # Walk propagatedBuildInputs to include transitive Python deps in PYTHONPATH.
+  # Without this, a plugin listing e.g. requests as a dep would fail at runtime
+  # if requests isn't already in the sealed uv2nix venv.
+  allExtraPythonPackages = python312.pkgs.requiredPythonModules extraPythonPackages;
+
+  pythonPath = lib.makeSearchPath sitePackagesPath allExtraPythonPackages;
+
+  pyprojectHash = builtins.hashString "sha256" (builtins.readFile ../pyproject.toml);
+  uvLockHash =
+    if builtins.pathExists ../uv.lock then
+      builtins.hashString "sha256" (builtins.readFile ../uv.lock)
+    else
+      "none";
+in
+stdenv.mkDerivation {
+  pname = "hermes-agent";
+  version = (builtins.fromTOML (builtins.readFile ../pyproject.toml)).project.version;
+
+  dontUnpack = true;
+  dontBuild = true;
+  nativeBuildInputs = [ makeWrapper ];
+
+  installPhase = ''
+    runHook preInstall
+
+    mkdir -p $out/share/hermes-agent $out/bin
+    cp -r ${bundledSkills} $out/share/hermes-agent/skills
+    cp -r ${hermesWeb} $out/share/hermes-agent/web_dist
+
+    mkdir -p $out/ui-tui
+    cp -r ${hermesTui}/lib/hermes-tui/* $out/ui-tui/
+
+    ${lib.concatMapStringsSep "\n"
+      (name: ''
+        makeWrapper ${hermesVenv}/bin/${name} $out/bin/${name} \
+          --suffix PATH : "${runtimePath}" \
+          --set HERMES_BUNDLED_SKILLS $out/share/hermes-agent/skills \
+          --set HERMES_WEB_DIST $out/share/hermes-agent/web_dist \
+          --set HERMES_TUI_DIR $out/ui-tui \
+          --set HERMES_PYTHON ${hermesVenv}/bin/python3 \
+          --set HERMES_NODE ${nodejs_22}/bin/node \
+          ${lib.optionalString (extraPythonPackages != [ ]) ''--suffix PYTHONPATH : "${pythonPath}"''}
+      '')
+      [
+        "hermes"
+        "hermes-agent"
+        "hermes-acp"
+      ]
+    }
+
+    ${lib.optionalString (extraPythonPackages != [ ]) ''
+      echo "=== Checking for plugin/core package collisions ==="
+      ${hermesVenv}/bin/python3 -c "
+import pathlib, sys, re
+
+def canonical(name):
+    return re.sub(r'[-_.]+', '-', name).lower()
+
+# Collect core venv package names
+core = set()
+venv_sp = pathlib.Path('${hermesVenv}/${sitePackagesPath}')
+for di in venv_sp.glob('*.dist-info'):
+    meta = di / 'METADATA'
+    if meta.exists():
+        for line in meta.read_text().splitlines():
+            if line.startswith('Name:'):
+                core.add(canonical(line.split(':', 1)[1].strip()))
+                break
+
+# Check each extra package for collisions
+extras_dirs = [${lib.concatMapStringsSep ", " (p: "'${toString p}'") allExtraPythonPackages}]
+for edir in extras_dirs:
+    sp = pathlib.Path(edir) / '${sitePackagesPath}'
+    if not sp.exists():
+        continue
+    for di in sp.glob('*.dist-info'):
+        meta = di / 'METADATA'
+        if not meta.exists():
+            continue
+        for line in meta.read_text().splitlines():
+            if line.startswith('Name:'):
+                pkg = canonical(line.split(':', 1)[1].strip())
+                if pkg in core:
+                    print(f'ERROR: plugin package \"{pkg}\" collides with a package in hermes sealed venv', file=sys.stderr)
+                    print(f'  from: {di}', file=sys.stderr)
+                    print(f'  Remove this dependency from extraPythonPackages.', file=sys.stderr)
+                    sys.exit(1)
+                break
+
+print('No collisions found.')
+      "
+      echo "=== No collisions ==="
+    ''}
+
+    runHook postInstall
+  '';
+
+  passthru = {
+    inherit hermesTui hermesWeb hermesNpmLib hermesVenv;
+
+    devShellHook = ''
+      STAMP=".nix-stamps/hermes-agent"
+      STAMP_VALUE="${pyprojectHash}:${uvLockHash}"
+      if [ ! -f "$STAMP" ] || [ "$(cat "$STAMP")" != "$STAMP_VALUE" ]; then
+        echo "hermes-agent: installing Python dependencies..."
+        uv venv .venv --python ${python312}/bin/python3 2>/dev/null || true
+        source .venv/bin/activate
+        uv pip install -e ".[all]"
+        [ -d mini-swe-agent ] && uv pip install -e ./mini-swe-agent 2>/dev/null || true
+        [ -d tinker-atropos ] && uv pip install -e ./tinker-atropos 2>/dev/null || true
+        mkdir -p .nix-stamps
+        echo "$STAMP_VALUE" > "$STAMP"
+      else
+        source .venv/bin/activate
+        export HERMES_PYTHON=${hermesVenv}/bin/python3
+      fi
+    '';
+  };
+
+  meta = with lib; {
+    description = "AI agent with advanced tool-calling capabilities";
+    homepage = "https://github.com/NousResearch/hermes-agent";
+    mainProgram = "hermes";
+    license = licenses.mit;
+    platforms = platforms.unix;
+  };
+}

nix/nixosModules.nix

@@ -28,6 +28,8 @@
 
   let
     cfg = config.services.hermes-agent;
+    effectivePackage = if cfg.extraPythonPackages == [ ] then cfg.package
+      else cfg.package.override { inherit (cfg) extraPythonPackages; };
     hermes-agent = inputs.self.packages.${pkgs.stdenv.hostPlatform.system}.default;
 
     # Deep-merge config type (from 0xrsydn/nix-hermes-agent)
@@ -456,6 +458,52 @@
         description = "Extra packages available on PATH.";
       };
 
+      extraPlugins = mkOption {
+        type = types.listOf types.package;
+        default = [ ];
+        description = ''
+          Directory-based plugin packages to symlink into the hermes plugins
+          directory. Each package should contain a plugin.yaml and __init__.py
+          at its root. Hermes discovers these automatically on startup.
+        '';
+        example = literalExpression ''
+          [
+            (pkgs.fetchFromGitHub {
+              owner = "stephenschoettler";
+              repo = "hermes-lcm";
+              name = "hermes-lcm";
+              rev = "v0.7.0";
+              hash = "sha256-...";
+            })
+          ]
+        '';
+      };
+
+      extraPythonPackages = mkOption {
+        type = types.listOf types.package;
+        default = [ ];
+        description = ''
+          Python packages to add to PYTHONPATH for entry-point plugin discovery.
+          These are pip-packaged plugins that register via the
+          hermes_agent.plugins entry-point group. Each package must be built
+          with the same Python interpreter as hermes (python312).
+        '';
+        example = literalExpression ''
+          [
+            (pkgs.python312Packages.buildPythonPackage {
+              pname = "rtk-hermes";
+              version = "1.0.0";
+              src = pkgs.fetchFromGitHub {
+                owner = "ogallotti";
+                repo = "rtk-hermes";
+                rev = "main";
+                hash = "sha256-...";
+              };
+            })
+          ]
+        '';
+      };
+
       restart = mkOption {
         type = types.str;
         default = "always";
@@ -570,7 +618,7 @@
       # so interactive shells share state (sessions, skills, cron) with the
       # gateway service instead of creating a separate ~/.hermes/.
       (lib.mkIf cfg.addToSystemPackages {
-        environment.systemPackages = [ cfg.package ];
+        environment.systemPackages = [ effectivePackage ];
         environment.variables.HERMES_HOME = "${cfg.stateDir}/.hermes";
       })
 
@@ -581,6 +629,16 @@
         });
       })
 
+      # ── Assertions ─────────────────────────────────────────────────────
+      {
+        assertions = let
+          names = map lib.getName cfg.extraPlugins;
+        in [{
+          assertion = (lib.length names) == (lib.length (lib.unique names));
+          message = "services.hermes-agent.extraPlugins: duplicate plugin names detected: ${toString names}. If using fetchFromGitHub, set name = \"plugin-name\" to disambiguate.";
+        }];
+      }
+
       # ── Warnings ──────────────────────────────────────────────────────
       (lib.mkIf (cfg.container.enable && !cfg.addToSystemPackages && cfg.container.hostUsers != []) {
         warnings = [
@@ -602,6 +660,7 @@
           "d ${cfg.stateDir}/.hermes/sessions 2770 ${cfg.user} ${cfg.group} - -"
           "d ${cfg.stateDir}/.hermes/logs   2770 ${cfg.user} ${cfg.group} - -"
           "d ${cfg.stateDir}/.hermes/memories 2770 ${cfg.user} ${cfg.group} - -"
+          "d ${cfg.stateDir}/.hermes/plugins 2770 ${cfg.user} ${cfg.group} - -"
           "d ${cfg.stateDir}/home           0750 ${cfg.user} ${cfg.group} - -"
           "d ${cfg.workingDirectory}         2770 ${cfg.user} ${cfg.group} - -"
         ];
@@ -623,7 +682,7 @@
           find ${cfg.stateDir}/.hermes -maxdepth 1 \
             \( -name "*.db" -o -name "*.db-wal" -o -name "*.db-shm" -o -name "SOUL.md" \) \
             -exec chmod g+rw {} + 2>/dev/null || true
-          for _subdir in cron sessions logs memories; do
+          for _subdir in cron sessions logs memories plugins; do
             mkdir -p "${cfg.stateDir}/.hermes/$_subdir"
             chown ${cfg.user}:${cfg.group} "${cfg.stateDir}/.hermes/$_subdir"
             chmod 2770 "${cfg.stateDir}/.hermes/$_subdir"
@@ -732,6 +791,22 @@ HERMES_NIX_ENV_EOF
           ${lib.concatStringsSep "\n" (lib.mapAttrsToList (name: _value: ''
             install -o ${cfg.user} -g ${cfg.group} -m 0640 ${documentDerivation}/${name} ${cfg.workingDirectory}/${name}
           '') cfg.documents)}
+
+        # ── Declarative plugins ─────────────────────────────────────────
+        # Remove stale managed symlinks (plugins removed from config)
+        find ${cfg.stateDir}/.hermes/plugins -maxdepth 1 -type l -name 'nix-managed-*' -delete 2>/dev/null || true
+
+        ${lib.concatStringsSep "\n" (map (plugin:
+          let
+            name = lib.getName plugin;
+          in ''
+            if [ ! -f "${plugin}/plugin.yaml" ]; then
+              echo "ERROR: extraPlugins entry '${plugin}' has no plugin.yaml" >&2
+              exit 1
+            fi
+            ln -sfn ${plugin} ${cfg.stateDir}/.hermes/plugins/nix-managed-${name}
+            chown -h ${cfg.user}:${cfg.group} ${cfg.stateDir}/.hermes/plugins/nix-managed-${name}
+          '') cfg.extraPlugins)}
         '';
       }
 
@@ -762,7 +837,7 @@ HERMES_NIX_ENV_EOF
             # reads them at Python startup — no systemd EnvironmentFile needed.
 
             ExecStart = lib.concatStringsSep " " ([
-              "${cfg.package}/bin/hermes"
+              "${effectivePackage}/bin/hermes"
               "gateway"
             ] ++ cfg.extraArgs);
 
@@ -785,7 +860,7 @@ HERMES_NIX_ENV_EOF
           };
 
           path = [
-            cfg.package
+            effectivePackage
             pkgs.bash
             pkgs.coreutils
             pkgs.git
@@ -810,11 +885,11 @@ HERMES_NIX_ENV_EOF
 
           preStart = ''
             # Stable symlinks — container references these, not store paths directly
-            ln -sfn ${cfg.package} ${cfg.stateDir}/current-package
+            ln -sfn ${effectivePackage} ${cfg.stateDir}/current-package
             ln -sfn ${containerEntrypoint} ${cfg.stateDir}/current-entrypoint
 
             # GC roots so nix-collect-garbage doesn't remove store paths in use
-            ${pkgs.nix}/bin/nix-store --add-root ${cfg.stateDir}/.gc-root --indirect -r ${cfg.package} 2>/dev/null || true
+            ${pkgs.nix}/bin/nix-store --add-root ${cfg.stateDir}/.gc-root --indirect -r ${effectivePackage} 2>/dev/null || true
             ${pkgs.nix}/bin/nix-store --add-root ${cfg.stateDir}/.gc-root-entrypoint --indirect -r ${containerEntrypoint} 2>/dev/null || true
 
             # Check if container needs (re)creation

nix/overlays.nix

@@ -0,0 +1,10 @@
+# nix/overlays.nix — Expose pkgs.hermes-agent for external NixOS configs
+{ inputs, ... }:
+{
+  flake.overlays.default = final: _: {
+    hermes-agent = final.callPackage ./hermes-agent.nix {
+      inherit (inputs) uv2nix pyproject-nix pyproject-build-systems;
+      npm-lockfile-fix = inputs.npm-lockfile-fix.packages.${final.stdenv.hostPlatform.system}.default;
+    };
+  };
+}

nix/packages.nix

@@ -4,120 +4,19 @@
   perSystem =
     { pkgs, inputs', ... }:
     let
-      hermesVenv = pkgs.callPackage ./python.nix {
+      hermesAgent = pkgs.callPackage ./hermes-agent.nix {
         inherit (inputs) uv2nix pyproject-nix pyproject-build-systems;
-      };
-
-      hermesNpmLib = pkgs.callPackage ./lib.nix {
         npm-lockfile-fix = inputs'.npm-lockfile-fix.packages.default;
       };
-
-      hermesTui = pkgs.callPackage ./tui.nix {
-        inherit hermesNpmLib;
-      };
-
-      # Import bundled skills, excluding runtime caches
-      bundledSkills = pkgs.lib.cleanSourceWith {
-        src = ../skills;
-        filter = path: _type: !(pkgs.lib.hasInfix "/index-cache/" path);
-      };
-
-      hermesWeb = pkgs.callPackage ./web.nix {
-        inherit hermesNpmLib;
-      };
-
-      runtimeDeps = with pkgs; [
-        nodejs_22
-        ripgrep
-        git
-        openssh
-        ffmpeg
-        tirith
-      ];
-
-      runtimePath = pkgs.lib.makeBinPath runtimeDeps;
-
-      # Lockfile hashes for dev shell stamps
-      pyprojectHash = builtins.hashString "sha256" (builtins.readFile ../pyproject.toml);
-      uvLockHash =
-        if builtins.pathExists ../uv.lock then
-          builtins.hashString "sha256" (builtins.readFile ../uv.lock)
-        else
-          "none";
     in
     {
       packages = {
-        default = pkgs.stdenv.mkDerivation {
-          pname = "hermes-agent";
-          version = (fromTOML (builtins.readFile ../pyproject.toml)).project.version;
+        default = hermesAgent;
+        tui = hermesAgent.hermesTui;
+        web = hermesAgent.hermesWeb;
 
-          dontUnpack = true;
-          dontBuild = true;
-          nativeBuildInputs = [ pkgs.makeWrapper ];
-
-          installPhase = ''
-            runHook preInstall
-
-            mkdir -p $out/share/hermes-agent $out/bin
-            cp -r ${bundledSkills} $out/share/hermes-agent/skills
-            cp -r ${hermesWeb} $out/share/hermes-agent/web_dist
-
-            # copy pre-built TUI (same layout as dev: ui-tui/dist/ + node_modules/)
-            mkdir -p $out/ui-tui
-            cp -r ${hermesTui}/lib/hermes-tui/* $out/ui-tui/
-
-            ${pkgs.lib.concatMapStringsSep "\n"
-              (name: ''
-                makeWrapper ${hermesVenv}/bin/${name} $out/bin/${name} \
-                  --suffix PATH : "${runtimePath}" \
-                  --set HERMES_BUNDLED_SKILLS $out/share/hermes-agent/skills \
-                  --set HERMES_WEB_DIST $out/share/hermes-agent/web_dist \
-                  --set HERMES_TUI_DIR $out/ui-tui \
-                  --set HERMES_PYTHON ${hermesVenv}/bin/python3 \
-                  --set HERMES_NODE ${pkgs.nodejs_22}/bin/node
-              '')
-              [
-                "hermes"
-                "hermes-agent"
-                "hermes-acp"
-              ]
-            }
-
-            runHook postInstall
-          '';
-
-          passthru.devShellHook = ''
-            STAMP=".nix-stamps/hermes-agent"
-            STAMP_VALUE="${pyprojectHash}:${uvLockHash}"
-            if [ ! -f "$STAMP" ] || [ "$(cat "$STAMP")" != "$STAMP_VALUE" ]; then
-              echo "hermes-agent: installing Python dependencies..."
-              uv venv .venv --python ${pkgs.python312}/bin/python3 2>/dev/null || true
-              source .venv/bin/activate
-              uv pip install -e ".[all]"
-              [ -d mini-swe-agent ] && uv pip install -e ./mini-swe-agent 2>/dev/null || true
-              [ -d tinker-atropos ] && uv pip install -e ./tinker-atropos 2>/dev/null || true
-              mkdir -p .nix-stamps
-              echo "$STAMP_VALUE" > "$STAMP"
-            else
-              source .venv/bin/activate
-              export HERMES_PYTHON=${hermesVenv}/bin/python3
-            fi
-          '';
-
-          meta = with pkgs.lib; {
-            description = "AI agent with advanced tool-calling capabilities";
-            homepage = "https://github.com/NousResearch/hermes-agent";
-            mainProgram = "hermes";
-            license = licenses.mit;
-            platforms = platforms.unix;
-          };
-        };
-
-        tui = hermesTui;
-        web = hermesWeb;
-
-        fix-lockfiles = hermesNpmLib.mkFixLockfiles {
-          packages = [ hermesTui hermesWeb ];
+        fix-lockfiles = hermesAgent.hermesNpmLib.mkFixLockfiles {
+          packages = [ hermesAgent.hermesTui hermesAgent.hermesWeb ];
         };
       };
     };

nix/python.nix

@@ -7,6 +7,7 @@
   pyproject-nix,
   pyproject-build-systems,
   stdenv,
+  dependency-groups ? [ "all" ],
 }:
 let
   workspace = uv2nix.lib.workspace.loadWorkspace { workspaceRoot = ./..; };
@@ -96,5 +97,5 @@ let
       ]);
 in
 pythonSet.mkVirtualEnv "hermes-agent-env" {
-  hermes-agent = [ "all" ];
+  hermes-agent = dependency-groups;
 }

plugins/memory/honcho/__init__.py

@@ -22,6 +22,7 @@ import threading
 import time
 from typing import Any, Dict, List, Optional
 
+from agent.memory_manager import sanitize_context
 from agent.memory_provider import MemoryProvider
 from tools.registry import tool_error
 
@@ -37,7 +38,10 @@ PROFILE_SCHEMA = {
     "description": (
         "Retrieve or update a peer card from Honcho — a curated list of key facts "
         "about that peer (name, role, preferences, communication style, patterns). "
-        "Pass `card` to update; omit `card` to read."
+        "Pass `card` to update; omit `card` to read.  If the card is empty, the "
+        "result includes a `hint` field explaining why (observation disabled, "
+        "fresh peer, dialectic layer still warming up, etc.) — this is NOT an "
+        "error.  Peer cards accumulate over time from observed conversation."
     ),
     "parameters": {
         "type": "object",
@@ -1056,6 +1060,63 @@ class HonchoMemoryProvider(MemoryProvider):
 
         return chunks
 
+    def _empty_profile_hint(self, peer: str) -> Dict[str, Any]:
+        """Build a diagnostic hint when honcho_profile returns an empty card.
+
+        A literal "No profile facts available yet." tells the model nothing
+        about WHY.  The model then often surfaces it to the user as a cryptic
+        error.  This hint enumerates the likely causes so the model can
+        explain the situation (or retry with a different peer).
+
+        Ordered by likelihood for a typical deployment:
+          1. Observation is disabled for this peer
+          2. Card hasn't accumulated yet (fresh peer, not enough dialectic
+             cycles — dialectic cadence runs every N turns)
+          3. Self-hosted Honcho backend doesn't support peer cards
+             (honcho-ai server < 3.x)
+        """
+        cfg = self._config
+        reasons: List[str] = []
+
+        if cfg is not None:
+            if peer == "user":
+                observe_me = bool(getattr(cfg, "user_observe_me", True))
+                observe_others = bool(getattr(cfg, "user_observe_others", True))
+            else:
+                observe_me = bool(getattr(cfg, "ai_observe_me", True))
+                observe_others = bool(getattr(cfg, "ai_observe_others", True))
+            if not (observe_me or observe_others):
+                reasons.append(
+                    f"observation is disabled for peer '{peer}' "
+                    f"(user_observe_me/ai_observe_me in config)"
+                )
+
+        cadence = getattr(self, "_dialectic_cadence", 1)
+        turn = getattr(self, "_turn_count", 0)
+        if turn < max(2, cadence):
+            reasons.append(
+                f"this session has only {turn} turn(s); peer cards accumulate "
+                f"as the dialectic layer reasons over conversation history "
+                f"(cadence every {cadence} turn(s))"
+            )
+
+        if not reasons:
+            reasons.append(
+                "peer card has no facts yet — Honcho's dialectic layer builds "
+                "this over time from observed turns; self-hosted Honcho < 3.x "
+                "does not support peer cards at all"
+            )
+
+        return {
+            "result": "No profile facts available yet.",
+            "hint": (
+                "This is not an error.  "
+                + "; ".join(reasons)
+                + ".  Try honcho_reasoning for a synthesized answer, or "
+                "honcho_search to query raw conversation excerpts."
+            ),
+        }
+
     def sync_turn(self, user_content: str, assistant_content: str, *, session_id: str = "") -> None:
         """Record the conversation turn in Honcho (non-blocking).
 
@@ -1068,13 +1129,15 @@ class HonchoMemoryProvider(MemoryProvider):
             return
 
         msg_limit = self._config.message_max_chars if self._config else 25000
+        clean_user_content = sanitize_context(user_content or "").strip()
+        clean_assistant_content = sanitize_context(assistant_content or "").strip()
 
         def _sync():
             try:
                 session = self._manager.get_or_create(self._session_key)
-                for chunk in self._chunk_message(user_content, msg_limit):
+                for chunk in self._chunk_message(clean_user_content, msg_limit):
                     session.add_message("user", chunk)
-                for chunk in self._chunk_message(assistant_content, msg_limit):
+                for chunk in self._chunk_message(clean_assistant_content, msg_limit):
                     session.add_message("assistant", chunk)
                 self._manager._flush_session(session)
             except Exception as e:
@@ -1087,8 +1150,20 @@ class HonchoMemoryProvider(MemoryProvider):
         )
         self._sync_thread.start()
 
-    def on_memory_write(self, action: str, target: str, content: str) -> None:
-        """Mirror built-in user profile writes as Honcho conclusions."""
+    def on_memory_write(
+        self,
+        action: str,
+        target: str,
+        content: str,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """Mirror built-in user profile writes as Honcho conclusions.
+
+        ``metadata`` is accepted for compatibility with the write-origin
+        work landed in main (commit 6a957a74); it's not yet threaded into
+        the Honcho conclusion payload.  Left as a follow-up so this PR
+        stays focused on the 7-PR consolidation and its review follow-ups.
+        """
         if action != "add" or target != "user" or not content:
             return
         if self._cron_skipped:
@@ -1154,7 +1229,7 @@ class HonchoMemoryProvider(MemoryProvider):
                     return json.dumps({"result": f"Peer card updated ({len(result)} facts).", "card": result})
                 card = self._manager.get_peer_card(self._session_key, peer=peer)
                 if not card:
-                    return json.dumps({"result": "No profile facts available yet."})
+                    return json.dumps(self._empty_profile_hint(peer))
                 return json.dumps({"result": card})
 
             elif tool_name == "honcho_search":

plugins/memory/honcho/cli.py

@@ -273,9 +273,38 @@ def _write_config(cfg: dict, path: Path | None = None) -> None:
 
 
 def _resolve_api_key(cfg: dict) -> str:
-    """Resolve API key with host -> root -> env fallback."""
+    """Resolve API key with host -> root -> env fallback.
+
+    For self-hosted instances configured with ``baseUrl`` instead of an API
+    key, returns ``"local"`` so that credential guards throughout the CLI
+    don't reject a valid configuration.  The ``baseUrl`` is scheme-validated
+    (http/https only) so that a typo like ``baseUrl: true`` can't silently
+    pass the guard.  Schemeless strings that look like host:port (legacy
+    config shapes, e.g. ``localhost:8000``) still pass — the Honcho SDK
+    will reject them itself with a clearer error than ours.
+    """
     host_key = ((cfg.get("hosts") or {}).get(_host_key()) or {}).get("apiKey")
-    return host_key or cfg.get("apiKey", "") or os.environ.get("HONCHO_API_KEY", "")
+    key = host_key or cfg.get("apiKey", "") or os.environ.get("HONCHO_API_KEY", "")
+    if not key:
+        base_url = cfg.get("baseUrl") or cfg.get("base_url") or os.environ.get("HONCHO_BASE_URL", "")
+        base_url = (base_url or "").strip()
+        if base_url:
+            from urllib.parse import urlparse
+            try:
+                parsed = urlparse(base_url)
+            except (TypeError, ValueError):
+                parsed = None
+            if parsed and parsed.scheme in ("http", "https") and parsed.netloc:
+                return "local"
+            # Schemeless but looks like a host (contains '.' or ':' and isn't
+            # a boolean literal): let it through so legacy configs don't
+            # regress into "no API key configured" when they previously worked.
+            lowered = base_url.lower()
+            if lowered not in ("true", "false", "none", "null") and any(
+                c in base_url for c in ".:"
+            ) and not base_url.isdigit():
+                return "local"
+    return key
 
 
 def _prompt(label: str, default: str | None = None, secret: bool = False) -> str:

plugins/memory/honcho/client.py

@@ -16,6 +16,7 @@ from __future__ import annotations
 import json
 import os
 import logging
+import hashlib
 from dataclasses import dataclass, field
 from pathlib import Path
 
@@ -27,7 +28,6 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
-GLOBAL_CONFIG_PATH = Path.home() / ".honcho" / "config.json"
 HOST = "hermes"
 
 
@@ -53,6 +53,11 @@ def resolve_active_host() -> str:
     return HOST
 
 
+def resolve_global_config_path() -> Path:
+    """Return the shared Honcho config path for the current HOME."""
+    return Path.home() / ".honcho" / "config.json"
+
+
 def resolve_config_path() -> Path:
     """Return the active Honcho config path.
 
@@ -72,7 +77,7 @@ def resolve_config_path() -> Path:
     if default_path != local_path and default_path.exists():
         return default_path
 
-    return GLOBAL_CONFIG_PATH
+    return resolve_global_config_path()
 
 
 _RECALL_MODE_ALIASES = {"auto": "hybrid"}
@@ -138,6 +143,15 @@ def _parse_dialectic_depth_levels(host_val, root_val, depth: int) -> list[str] |
     return None
 
 
+# Default HTTP timeout (seconds) applied when no explicit timeout is
+# configured via HonchoClientConfig.timeout, honcho.timeout / requestTimeout,
+# or HONCHO_TIMEOUT. Honcho calls happen on the post-response path of
+# run_conversation; without a cap the agent can block indefinitely when
+# the Honcho backend is unreachable, preventing the gateway from
+# delivering the already-generated response.
+_DEFAULT_HTTP_TIMEOUT = 30.0
+
+
 def _resolve_optional_float(*values: Any) -> float | None:
     """Return the first non-empty value coerced to a positive float."""
     for value in values:
@@ -226,6 +240,13 @@ class HonchoClientConfig:
     # Identity
     peer_name: str | None = None
     ai_peer: str = "hermes"
+    # When True, ``peer_name`` wins over any gateway-supplied runtime
+    # identity (Telegram UID, Discord ID, …) when resolving the user peer.
+    # This keeps memory unified across platforms for single-user deployments
+    # where Honcho's one peer-name is an unambiguous identity — otherwise
+    # each platform would fork memory into its own peer (#14984).  Default
+    # ``False`` preserves existing multi-user behaviour.
+    pin_peer_name: bool = False
     # Toggles
     enabled: bool = False
     save_messages: bool = True
@@ -420,6 +441,11 @@ class HonchoClientConfig:
             timeout=timeout,
             peer_name=host_block.get("peerName") or raw.get("peerName"),
             ai_peer=ai_peer,
+            pin_peer_name=_resolve_bool(
+                host_block.get("pinPeerName"),
+                raw.get("pinPeerName"),
+                default=False,
+            ),
             enabled=enabled,
             save_messages=save_messages,
             write_frequency=write_frequency,
@@ -522,6 +548,39 @@ class HonchoClientConfig:
             pass
         return None
 
+    # Honcho enforces a 100-char limit on session IDs. Long gateway session keys
+    # (Matrix "!room:server" + thread event IDs, Telegram supergroup reply
+    # chains, Slack thread IDs with long workspace prefixes) can overflow this
+    # limit after sanitization; the Honcho API then rejects every call for that
+    # session with "session_id too long". See issue #13868.
+    _HONCHO_SESSION_ID_MAX_LEN = 100
+    _HONCHO_SESSION_ID_HASH_LEN = 8
+
+    @classmethod
+    def _enforce_session_id_limit(cls, sanitized: str, original: str) -> str:
+        """Truncate a sanitized session ID to Honcho's 100-char limit.
+
+        The common case (short keys) short-circuits with no modification.
+        For over-limit keys, keep a prefix of the sanitized ID and append a
+        deterministic ``-<sha256 prefix>`` suffix so two distinct long keys
+        that share a leading segment don't collide onto the same truncated ID.
+        The hash is taken over the *original* pre-sanitization key, so two
+        inputs that sanitize to the same string still collide intentionally
+        (same logical session), but two inputs that only share a prefix do not.
+        """
+        max_len = cls._HONCHO_SESSION_ID_MAX_LEN
+        if len(sanitized) <= max_len:
+            return sanitized
+
+        hash_len = cls._HONCHO_SESSION_ID_HASH_LEN
+        digest = hashlib.sha256(original.encode("utf-8")).hexdigest()[:hash_len]
+        # max_len - hash_len - 1 (for the '-' separator) chars of the sanitized
+        # prefix, then '-<hash>'. Strip any trailing hyphen from the prefix so
+        # the result doesn't double up on separators.
+        prefix_len = max_len - hash_len - 1
+        prefix = sanitized[:prefix_len].rstrip("-")
+        return f"{prefix}-{digest}"
+
     def resolve_session_name(
         self,
         cwd: str | None = None,
@@ -566,7 +625,7 @@ class HonchoClientConfig:
         if gateway_session_key:
             sanitized = re.sub(r'[^a-zA-Z0-9_-]+', '-', gateway_session_key).strip('-')
             if sanitized:
-                return sanitized
+                return self._enforce_session_id_limit(sanitized, gateway_session_key)
 
         # per-session: inherit Hermes session_id (new Honcho session each run)
         if self.session_strategy == "per-session" and session_id:
@@ -646,6 +705,11 @@ def get_honcho_client(config: HonchoClientConfig | None = None) -> Honcho:
         except Exception:
             pass
 
+    # Fall back to the default so an unconfigured install cannot hang
+    # indefinitely on a stalled Honcho request.
+    if resolved_timeout is None:
+        resolved_timeout = _DEFAULT_HTTP_TIMEOUT
+
     if resolved_base_url:
         logger.info("Initializing Honcho client (base_url: %s, workspace: %s)", resolved_base_url, config.workspace_id)
     else:

plugins/memory/honcho/session.py

@@ -95,6 +95,7 @@ class HonchoSessionManager:
         self._config = config
         self._runtime_user_peer_name = runtime_user_peer_name
         self._cache: dict[str, HonchoSession] = {}
+        self._cache_lock = threading.RLock()
         self._peers_cache: dict[str, Any] = {}
         self._sessions_cache: dict[str, Any] = {}
 
@@ -273,17 +274,35 @@ class HonchoSessionManager:
         Returns:
             The session.
         """
-        if key in self._cache:
-            logger.debug("Local session cache hit: %s", key)
-            return self._cache[key]
+        with self._cache_lock:
+            if key in self._cache:
+                logger.debug("Local session cache hit: %s", key)
+                return self._cache[key]
 
-        # Gateway sessions should use the runtime user identity when available.
-        if self._runtime_user_peer_name:
+        # Determine peer IDs — no lock needed (read-only, no shared state mutation).
+        # Gateway sessions normally use the runtime user identity (the
+        # platform-native ID: Telegram UID, Discord snowflake, Slack user,
+        # etc.) so multi-user bots scope memory per user.  For a single-user
+        # deployment the config-supplied ``peer_name`` is an unambiguous
+        # identity and we should keep it unified across platforms — see
+        # #14984.  Opt into that with ``hosts.<host>.pinPeerName: true`` in
+        # ``honcho.json`` (or root-level ``pinPeerName: true``).
+        # `is True` (not `bool(...)`) is deliberate: several multi-user tests
+        # pass a ``MagicMock`` for ``config`` where ``mock.pin_peer_name``
+        # silently returns another MagicMock — truthy by default.  Requiring
+        # strict ``True`` keeps pinning as opt-in even for callers that
+        # haven't updated their mocks yet; real configs built via
+        # ``from_global_config`` always produce a proper boolean.
+        pin_peer_name = (
+            self._config is not None
+            and bool(getattr(self._config, "peer_name", None))
+            and getattr(self._config, "pin_peer_name", False) is True
+        )
+        if self._runtime_user_peer_name and not pin_peer_name:
             user_peer_id = self._sanitize_id(self._runtime_user_peer_name)
         elif self._config and self._config.peer_name:
             user_peer_id = self._sanitize_id(self._config.peer_name)
         else:
-            # Fallback: derive from session key
             parts = key.split(":", 1)
             channel = parts[0] if len(parts) > 1 else "default"
             chat_id = parts[1] if len(parts) > 1 else key
@@ -293,19 +312,14 @@ class HonchoSessionManager:
             self._config.ai_peer if self._config else "hermes-assistant"
         )
 
-        # Sanitize session ID for Honcho
+        # All expensive I/O outside the lock — Honcho's persistence is source of truth
         honcho_session_id = self._sanitize_id(key)
-
-        # Get or create peers
         user_peer = self._get_or_create_peer(user_peer_id)
         assistant_peer = self._get_or_create_peer(assistant_peer_id)
-
-        # Get or create Honcho session
         honcho_session, existing_messages = self._get_or_create_honcho_session(
             honcho_session_id, user_peer, assistant_peer
         )
 
-        # Convert Honcho messages to local format
         local_messages = []
         for msg in existing_messages:
             role = "assistant" if msg.peer_id == assistant_peer_id else "user"
@@ -313,10 +327,9 @@ class HonchoSessionManager:
                 "role": role,
                 "content": msg.content,
                 "timestamp": msg.created_at.isoformat() if msg.created_at else "",
-                "_synced": True,  # Already in Honcho
+                "_synced": True,
             })
 
-        # Create local session wrapper with existing messages
         session = HonchoSession(
             key=key,
             user_peer_id=user_peer_id,
@@ -325,7 +338,9 @@ class HonchoSessionManager:
             messages=local_messages,
         )
 
-        self._cache[key] = session
+        # Write to cache under lock — only one writer wins
+        with self._cache_lock:
+            self._cache[key] = session
         return session
 
     def _flush_session(self, session: HonchoSession) -> bool:
@@ -356,13 +371,15 @@ class HonchoSessionManager:
             for msg in new_messages:
                 msg["_synced"] = True
             logger.debug("Synced %d messages to Honcho for %s", len(honcho_messages), session.key)
-            self._cache[session.key] = session
+            with self._cache_lock:
+                self._cache[session.key] = session
             return True
         except Exception as e:
             for msg in new_messages:
                 msg["_synced"] = False
             logger.error("Failed to sync messages to Honcho: %s", e)
-            self._cache[session.key] = session
+            with self._cache_lock:
+                self._cache[session.key] = session
             return False
 
     def _async_writer_loop(self) -> None:
@@ -434,7 +451,9 @@ class HonchoSessionManager:
         Called at session end for "session" write_frequency, or to force
         a sync before process exit regardless of mode.
         """
-        for session in list(self._cache.values()):
+        with self._cache_lock:
+            sessions = list(self._cache.values())
+        for session in sessions:
             try:
                 self._flush_session(session)
             except Exception as e:
@@ -459,9 +478,10 @@ class HonchoSessionManager:
 
     def delete(self, key: str) -> bool:
         """Delete a session from local cache."""
-        if key in self._cache:
-            del self._cache[key]
-            return True
+        with self._cache_lock:
+            if key in self._cache:
+                del self._cache[key]
+                return True
         return False
 
     def new_session(self, key: str) -> HonchoSession:
@@ -473,20 +493,25 @@ class HonchoSessionManager:
         """
         import time
 
-        # Remove old session from caches (but don't delete from Honcho)
-        old_session = self._cache.pop(key, None)
-        if old_session:
-            self._sessions_cache.pop(old_session.honcho_session_id, None)
+        # Hold the reentrant lock across get_or_create so a concurrent caller
+        # can't observe the (old-popped, new-not-yet-inserted) gap and create
+        # its own session under the raw key.  `_cache_lock` is an RLock so
+        # nested reacquisition inside get_or_create is safe.
+        with self._cache_lock:
+            # Remove old session from caches (but don't delete from Honcho)
+            old_session = self._cache.pop(key, None)
+            if old_session:
+                self._sessions_cache.pop(old_session.honcho_session_id, None)
 
-        # Create new session with timestamp suffix
-        timestamp = int(time.time())
-        new_key = f"{key}:{timestamp}"
+            # Create new session with timestamp suffix
+            timestamp = int(time.time())
+            new_key = f"{key}:{timestamp}"
 
-        # get_or_create will create a fresh session
-        session = self.get_or_create(new_key)
+            # get_or_create will create a fresh session
+            session = self.get_or_create(new_key)
 
-        # Cache under the original key so callers find it by the expected name
-        self._cache[key] = session
+            # Cache under the original key so callers find it by the expected name
+            self._cache[key] = session
 
         logger.info("Created new session for %s (honcho: %s)", key, session.honcho_session_id)
         return session

providers/README.md

@@ -0,0 +1,307 @@
+# providers/
+
+Single source of truth for every inference provider Hermes knows about.
+
+Each provider is declared once here as a `ProviderProfile`. Every other layer —
+auth resolution, transport kwargs, model listing, runtime routing — reads from
+these profiles instead of maintaining its own parallel data.
+
+---
+
+## Directory layout
+
+```
+providers/
+├── base.py           ProviderProfile dataclass + OMIT_TEMPERATURE sentinel
+├── __init__.py       Registry: register_provider(), get_provider_profile()
+├── README.md         This file
+│
+├── # Simple providers — just identity + auth + endpoint
+├── alibaba.py        Alibaba Cloud DashScope
+├── arcee.py          Arcee AI
+├── bedrock.py        AWS Bedrock  (api_mode=bedrock_converse)
+├── deepseek.py       DeepSeek
+├── huggingface.py    Hugging Face Inference API
+├── kilocode.py       Kilo Code
+├── minimax.py        MiniMax (international + CN)
+├── nvidia.py         NVIDIA NIM  (default_max_tokens=16384)
+├── ollama_cloud.py   Ollama Cloud
+├── stepfun.py        StepFun
+├── xiaomi.py         Xiaomi MiMo
+├── xai.py            xAI Grok  (api_mode=codex_responses)
+├── zai.py            Z.AI / GLM
+│
+├── # Medium — one or two quirks
+├── anthropic.py      Native Anthropic  (x-api-key header, api_mode=anthropic_messages)
+├── copilot.py        GitHub Copilot  (auth_type=copilot, reasoning per model)
+├── copilot_acp.py    Copilot ACP subprocess  (api_mode=copilot_acp)
+├── custom.py         Custom/Ollama local  (think=false, num_ctx)
+├── gemini.py         Google Gemini AI Studio + Cloud Code OAuth
+├── kimi.py           Kimi Coding  (OMIT_TEMPERATURE, thinking, dual endpoint)
+├── openai_codex.py   OpenAI Codex OAuth  (api_mode=codex_responses)
+├── opencode.py       OpenCode Zen + Go  (per-model api_mode routing)
+│
+├── # Complex — subclasses with multiple overrides
+├── nous.py           Nous Portal  (tags, attribution, reasoning omit-when-disabled)
+├── openrouter.py     OpenRouter  (provider preferences, public model fetch)
+├── qwen.py           Qwen OAuth  (message normalization, cache_control, vl_hires)
+└── vercel.py         Vercel AI Gateway  (attribution headers, reasoning passthrough)
+```
+
+---
+
+## ProviderProfile fields
+
+```python
+@dataclass
+class ProviderProfile:
+    # Identity
+    name: str                    # canonical ID — auto-registered as PROVIDER_REGISTRY key for new api-key providers
+    api_mode: str                # "chat_completions" | "anthropic_messages" |
+                                 # "codex_responses" | "bedrock_converse" | "copilot_acp"
+    aliases: tuple               # alternate names resolved by get_provider_profile()
+
+    # Auth & endpoints
+    env_vars: tuple              # env var names holding the API key, in priority order
+    base_url: str                # default inference endpoint
+    models_url: str              # explicit models endpoint; falls back to {base_url}/models
+                                 # set when the models catalog lives at a different URL
+                                 # (e.g. OpenRouter: public /api/v1/models vs /api/v1 inference)
+    auth_type: str               # "api_key" | "oauth_device_code" | "oauth_external" |
+                                 # "copilot" | "aws" | "external_process"
+
+    # Client-level quirks
+    default_headers: dict        # extra HTTP headers sent on every request
+
+    # Request-level quirks
+    fixed_temperature: Any       # None = use caller's default; OMIT_TEMPERATURE = don't send
+    default_max_tokens: int|None # inject max_tokens when caller omits it
+    default_aux_model: str       # cheap model for auxiliary tasks (compression, vision, etc.)
+                                 # empty string = use main model (default)
+```
+
+---
+
+## Hooks (override in a subclass)
+
+| Method | When to override |
+|--------|-----------------|
+| `prepare_messages(messages)` | Provider needs message pre-processing (Qwen: string → list-of-parts, cache_control) |
+| `build_extra_body(*, session_id, **ctx)` | Provider-specific `extra_body` fields (Nous: tags, OpenRouter: provider preferences) |
+| `build_api_kwargs_extras(*, reasoning_config, **ctx)` | Returns `(extra_body_additions, top_level_kwargs)` — use when some fields go to `extra_body` and some go top-level (Kimi: `reasoning_effort` top-level; OpenRouter: `reasoning` in extra_body) |
+| `fetch_models(*, api_key, timeout)` | Custom model listing (Anthropic: x-api-key header; OpenRouter: public endpoint, no auth; Bedrock/copilot-acp: return None) |
+
+All hooks have safe defaults — only override what differs from the base.
+
+---
+
+## How to add a new provider
+
+### 1. Simple (standard OpenAI-compatible endpoint)
+
+```python
+# providers/myprovider.py
+from providers import register_provider
+from providers.base import ProviderProfile
+
+myprovider = ProviderProfile(
+    name="myprovider",           # must match id in hermes_cli/auth.py PROVIDER_REGISTRY
+    aliases=("my-provider", "myp"),
+    api_mode="chat_completions",
+    env_vars=("MYPROVIDER_API_KEY",),
+    base_url="https://api.myprovider.com/v1",
+    auth_type="api_key",
+)
+
+register_provider(myprovider)
+```
+
+The default `fetch_models()` will call `GET https://api.myprovider.com/v1/models`
+with Bearer auth automatically. No override needed for standard `/v1/models`.
+
+### 2. With quirks (subclass)
+
+```python
+# providers/myprovider.py
+from typing import Any
+from providers import register_provider
+from providers.base import ProviderProfile
+
+
+class MyProviderProfile(ProviderProfile):
+    """My provider — custom reasoning header."""
+
+    def build_api_kwargs_extras(
+        self,
+        *,
+        reasoning_config: dict | None = None,
+        **ctx: Any,
+    ) -> tuple[dict[str, Any], dict[str, Any]]:
+        extra_body: dict[str, Any] = {}
+        if reasoning_config:
+            extra_body["my_reasoning"] = reasoning_config.get("effort", "medium")
+        return extra_body, {}
+
+    def fetch_models(
+        self,
+        *,
+        api_key: str | None = None,
+        timeout: float = 8.0,
+    ) -> list[str] | None:
+        # Override only if your endpoint differs from standard /v1/models
+        return super().fetch_models(api_key=api_key, timeout=timeout)
+
+
+myprovider = MyProviderProfile(
+    name="myprovider",
+    aliases=("myp",),
+    env_vars=("MYPROVIDER_API_KEY",),
+    base_url="https://api.myprovider.com/v1",
+)
+
+register_provider(myprovider)
+```
+
+### 3. Wire it up
+
+After creating the file, add `name` to the `_PROFILE_ACTIVE_PROVIDERS` set in
+`run_agent.py` once you've verified parity against the legacy flag path. Start
+with a simple provider (no message prep, no reasoning quirks) and work up.
+
+---
+
+## fetch_models contract
+
+```python
+def fetch_models(
+    self,
+    *,
+    api_key: str | None = None,
+    timeout: float = 8.0,
+) -> list[str] | None:
+    ...
+```
+
+- Returns `list[str]`: model IDs from the provider's live endpoint.
+- Returns `None`: provider doesn't support REST model listing (Bedrock, copilot-acp),
+  or the request failed. Callers **must** fall back to `_PROVIDER_MODELS` on `None`.
+- Never raises — swallow exceptions and return `None`.
+- Default implementation: `GET {base_url}/models` with Bearer auth. Works for any
+  standard OpenAI-compatible provider.
+
+**Override when:**
+- Auth header is not `Bearer` (Anthropic: `x-api-key`)
+- Endpoint path differs from `/models` AND you can't just set `models_url` (OpenRouter: public endpoint, pass `api_key=None` explicitly)
+- Response format differs (extra wrapping, non-standard `id` field)
+- Provider has no REST endpoint (Bedrock, copilot-acp → return `None`)
+- Filtering needed post-fetch (only tool-capable models, etc.)
+
+Use `models_url` instead of overriding when the only difference is the URL:
+
+```python
+# No subclass needed — just set models_url
+myprovider = ProviderProfile(
+    name="myprovider",
+    base_url="https://api.myprovider.com/v1",
+    models_url="https://catalog.myprovider.com/models",  # different host
+)
+```
+
+---
+
+## Debugging
+
+### Check if a provider resolves
+
+```python
+from providers import get_provider_profile
+
+p = get_provider_profile("myprovider")
+print(p)           # ProviderProfile(name='myprovider', ...)
+print(p.base_url)
+print(p.api_mode)
+```
+
+### Check all registered providers
+
+```python
+from providers import _REGISTRY
+print(list(_REGISTRY.keys()))
+```
+
+### Test live model fetch
+
+```python
+import os
+from providers import get_provider_profile
+
+p = get_provider_profile("myprovider")
+key = os.getenv("MYPROVIDER_API_KEY")
+models = p.fetch_models(api_key=key, timeout=5.0)
+print(models)      # list of model IDs, or None on failure
+```
+
+### Test alias resolution
+
+```python
+from providers import get_provider_profile
+
+# All of these should return the same profile
+assert get_provider_profile("openrouter").name == "openrouter"
+assert get_provider_profile("or").name == "openrouter"
+```
+
+### Run the provider test suite
+
+```bash
+# From the repo root
+source venv/bin/activate
+python -m pytest tests/providers/ -v
+```
+
+### Check ruff + ty compliance
+
+```bash
+source venv/bin/activate
+ruff format providers/*.py
+ruff check providers/*.py --select UP,E,F,I,W
+ty check providers/*.py
+```
+
+---
+
+## Common mistakes
+
+**Wrong `name`** — must be the same string that appears as the key in
+`hermes_cli/auth.py` `PROVIDER_REGISTRY`. New api-key providers auto-register
+into `PROVIDER_REGISTRY` from the profile, so the name IS the key. For providers
+with a pre-existing `PROVIDER_REGISTRY` entry, use the exact `id` field value.
+
+**Wrong `env_vars`** — separate API-key vars from base-URL override vars in the
+tuple. Env vars that end with `_BASE_URL` or `_URL` are treated as URL overrides;
+everything else is treated as an API key. Getting this wrong causes the doctor
+health check to send a URL string as a Bearer token.
+
+**Wrong `base_url`** — several providers have non-obvious paths:
+`stepfun: /step_plan/v1`, `opencode-go: /zen/go/v1`. The profile's `base_url`
+is also used as the `inference_base_url` when auto-registering into `PROVIDER_REGISTRY`
+for new providers, so it must be correct for auth resolution to work.
+
+**Skipping `api_mode`** — defaults to `chat_completions`. Providers that use
+`anthropic_messages`, `codex_responses`, `bedrock_converse`, or `copilot_acp`
+must set it explicitly.
+
+**Forgetting `register_provider()`** — auto-discovery runs `pkgutil.iter_modules`
+over the package and imports each module, but only if `register_provider()` is
+called at module level. Without it the profile is never in `_REGISTRY`.
+
+**`fetch_models` returning the wrong shape** — must return `list[str]` (plain
+model IDs), not `list[tuple]` or `list[dict]`. Callers expect plain strings.
+
+**Wrong `build_api_kwargs_extras` return shape** — must return a 2-tuple
+`(extra_body_dict, top_level_dict)`. Returning a single dict causes a
+`ValueError: not enough values to unpack` in the transport.
+
+**`build_api_kwargs_extras` wrong tuple** — must return `(extra_body_dict,
+top_level_dict)`. Returning a flat dict or swapping the order silently sends
+fields to the wrong place.

providers/__init__.py

@@ -0,0 +1,76 @@
+"""Provider module registry.
+
+Auto-discovers ProviderProfile instances from providers/*.py modules.
+Each module should define a module-level PROVIDER or PROVIDERS list.
+
+Usage:
+    from providers import get_provider_profile
+    profile = get_provider_profile("nvidia")  # returns ProviderProfile or None
+    profile = get_provider_profile("kimi")    # checks name + aliases
+"""
+
+from __future__ import annotations
+
+from providers.base import OMIT_TEMPERATURE, ProviderProfile  # noqa: F401
+
+_REGISTRY: dict[str, ProviderProfile] = {}
+_ALIASES: dict[str, str] = {}
+_discovered = False
+
+
+def register_provider(profile: ProviderProfile) -> None:
+    """Register a provider profile by name and aliases."""
+    _REGISTRY[profile.name] = profile
+    for alias in profile.aliases:
+        _ALIASES[alias] = profile.name
+
+
+def get_provider_profile(name: str) -> ProviderProfile | None:
+    """Look up a provider profile by name or alias.
+
+    Returns None if the provider has no profile (falls back to generic).
+    """
+    if not _discovered:
+        _discover_providers()
+    canonical = _ALIASES.get(name, name)
+    return _REGISTRY.get(canonical)
+
+
+def list_providers() -> list[ProviderProfile]:
+    """Return all registered provider profiles (one per canonical name)."""
+    if not _discovered:
+        _discover_providers()
+    # Deduplicate: _REGISTRY has canonical names; _ALIASES points to same objects
+    seen: set[int] = set()
+    result: list[ProviderProfile] = []
+    for profile in _REGISTRY.values():
+        pid = id(profile)
+        if pid not in seen:
+            seen.add(pid)
+            result.append(profile)
+    return result
+
+
+def _discover_providers() -> None:
+    """Import all provider modules to trigger registration."""
+    global _discovered
+    if _discovered:
+        return
+    _discovered = True
+
+    import importlib
+    import pkgutil
+
+    import providers as _pkg
+
+    for _importer, modname, _ispkg in pkgutil.iter_modules(_pkg.__path__):
+        if modname.startswith("_") or modname == "base":
+            continue
+        try:
+            importlib.import_module(f"providers.{modname}")
+        except ImportError as e:
+            import logging
+
+            logging.getLogger(__name__).warning(
+                "Failed to import provider module %s: %s", modname, e
+            )

providers/alibaba.py

@@ -0,0 +1,13 @@
+"""Alibaba Cloud DashScope provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+alibaba = ProviderProfile(
+    name="alibaba",
+    aliases=("dashscope", "alibaba-cloud", "qwen-dashscope"),
+    env_vars=("DASHSCOPE_API_KEY",),
+    base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
+)
+
+register_provider(alibaba)

providers/anthropic.py

@@ -0,0 +1,52 @@
+"""Native Anthropic provider profile."""
+
+import json
+import logging
+import urllib.request
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+logger = logging.getLogger(__name__)
+
+
+class AnthropicProfile(ProviderProfile):
+    """Native Anthropic — uses x-api-key header, not Bearer."""
+
+    def fetch_models(
+        self,
+        *,
+        api_key: str | None = None,
+        timeout: float = 8.0,
+    ) -> list[str] | None:
+        """Anthropic uses x-api-key header and anthropic-version."""
+        if not api_key:
+            return None
+        try:
+            req = urllib.request.Request("https://api.anthropic.com/v1/models")
+            req.add_header("x-api-key", api_key)
+            req.add_header("anthropic-version", "2023-06-01")
+            req.add_header("Accept", "application/json")
+            with urllib.request.urlopen(req, timeout=timeout) as resp:
+                data = json.loads(resp.read().decode())
+            return [
+                m["id"]
+                for m in data.get("data", [])
+                if isinstance(m, dict) and "id" in m
+            ]
+        except Exception as exc:
+            logger.debug("fetch_models(anthropic): %s", exc)
+            return None
+
+
+anthropic = AnthropicProfile(
+    name="anthropic",
+    aliases=("claude", "claude-oauth", "claude-code"),
+    api_mode="anthropic_messages",
+    env_vars=("ANTHROPIC_API_KEY", "ANTHROPIC_TOKEN", "CLAUDE_CODE_OAUTH_TOKEN"),
+    base_url="https://api.anthropic.com",
+    auth_type="api_key",
+    default_aux_model="claude-haiku-4-5-20251001",
+)
+
+register_provider(anthropic)

providers/arcee.py

@@ -0,0 +1,13 @@
+"""Arcee AI provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+arcee = ProviderProfile(
+    name="arcee",
+    aliases=("arcee-ai", "arceeai"),
+    env_vars=("ARCEEAI_API_KEY",),
+    base_url="https://api.arcee.ai/api/v1",
+)
+
+register_provider(arcee)

providers/base.py

@@ -0,0 +1,165 @@
+"""Provider profile base class.
+
+A ProviderProfile declares everything about an inference provider in one place:
+auth, endpoints, client quirks, request-time quirks. The transport reads this
+instead of receiving 20+ boolean flags.
+
+Provider profiles are DECLARATIVE — they describe the provider's behavior.
+They do NOT own client construction, credential rotation, or streaming.
+Those stay on AIAgent.
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, field
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+# Sentinel for "omit temperature entirely" (Kimi: server manages it)
+OMIT_TEMPERATURE = object()
+
+
+@dataclass
+class ProviderProfile:
+    """Base provider profile — subclass or instantiate with overrides."""
+
+    # ── Identity ─────────────────────────────────────────────
+    name: str
+    api_mode: str = "chat_completions"
+    aliases: tuple = ()
+
+    # ── Human-readable metadata ───────────────────────────────
+    display_name: str = ""       # e.g. "GMI Cloud" — shown in picker/labels
+    description: str = ""        # e.g. "GMI Cloud (multi-model direct API)" — picker subtitle
+    signup_url: str = ""         # e.g. "https://www.gmicloud.ai/" — shown during setup
+
+    # ── Auth & endpoints ─────────────────────────────────────
+    env_vars: tuple = ()
+    base_url: str = ""
+    models_url: str = ""  # explicit models endpoint; falls back to {base_url}/models
+    auth_type: str = "api_key"   # api_key|oauth_device_code|oauth_external|copilot|aws_sdk
+
+    # ── Model catalog ─────────────────────────────────────────
+    # fallback_models: curated list shown in /model picker when live fetch fails.
+    # Only agentic models that support tool calling should appear here.
+    fallback_models: tuple = ()
+
+    # hostname: base hostname for URL→provider reverse-mapping in model_metadata.py
+    # e.g. "api.gmi-serving.com". Derived from base_url when empty.
+    hostname: str = ""
+
+    # ── Client-level quirks (set once at client construction) ─
+    default_headers: dict[str, str] = field(default_factory=dict)
+
+    # ── Request-level quirks ─────────────────────────────────
+    # Temperature: None = use caller's default, OMIT_TEMPERATURE = don't send
+    fixed_temperature: Any = None
+    default_max_tokens: int | None = None
+    default_aux_model: str = (
+        ""  # cheap model for auxiliary tasks (compression, vision, etc.)
+    )
+    # empty = use main model
+
+    # ── Hooks (override in subclass for complex providers) ───
+
+    def get_hostname(self) -> str:
+        """Return the provider's base hostname for URL-based detection.
+
+        Uses self.hostname if set explicitly, otherwise derives it from base_url.
+        e.g. 'https://api.gmi-serving.com/v1' → 'api.gmi-serving.com'
+        """
+        if self.hostname:
+            return self.hostname
+        if self.base_url:
+            from urllib.parse import urlparse
+            return urlparse(self.base_url).hostname or ""
+        return ""
+
+    def prepare_messages(self, messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        """Provider-specific message preprocessing.
+
+        Called AFTER codex field sanitization, BEFORE developer role swap.
+        Default: pass-through.
+        """
+        return messages
+
+    def build_extra_body(
+        self, *, session_id: str | None = None, **context: Any
+    ) -> dict[str, Any]:
+        """Provider-specific extra_body fields.
+
+        Merged into the API kwargs extra_body. Default: empty dict.
+        """
+        return {}
+
+    def build_api_kwargs_extras(
+        self,
+        *,
+        reasoning_config: dict | None = None,
+        **context: Any,
+    ) -> tuple[dict[str, Any], dict[str, Any]]:
+        """Provider-specific kwargs split between extra_body and top-level api_kwargs.
+
+        Returns (extra_body_additions, top_level_kwargs).
+        The transport merges extra_body_additions into extra_body, and
+        top_level_kwargs directly into api_kwargs.
+
+        This split exists because some providers put reasoning config in
+        extra_body (OpenRouter: extra_body.reasoning) while others put it
+        as top-level api_kwargs (Kimi: api_kwargs.reasoning_effort).
+
+        Default: ({}, {}).
+        """
+        return {}, {}
+
+    def fetch_models(
+        self,
+        *,
+        api_key: str | None = None,
+        timeout: float = 8.0,
+    ) -> list[str] | None:
+        """Fetch the live model list from the provider's models endpoint.
+
+        Returns a list of model ID strings, or None if the fetch failed or
+        the provider does not support live model listing.
+
+        Resolution order for the endpoint URL:
+          1. self.models_url  (explicit override — use when the models
+             endpoint differs from the inference base URL, e.g. OpenRouter
+             exposes a public catalog at /api/v1/models while inference is
+             at /api/v1)
+          2. self.base_url + "/models"  (standard OpenAI-compat fallback)
+
+        The default implementation sends Bearer auth when api_key is given
+        and forwards self.default_headers. Override to customise auth, path,
+        response shape, or to return None for providers with no REST catalog.
+
+        Callers must always fall back to the static _PROVIDER_MODELS list
+        when this returns None.
+        """
+        url = (self.models_url or "").strip()
+        if not url:
+            if not self.base_url:
+                return None
+            url = self.base_url.rstrip("/") + "/models"
+
+        import json
+        import urllib.request
+
+        req = urllib.request.Request(url)
+        if api_key:
+            req.add_header("Authorization", f"Bearer {api_key}")
+        req.add_header("Accept", "application/json")
+        for k, v in self.default_headers.items():
+            req.add_header(k, v)
+
+        try:
+            with urllib.request.urlopen(req, timeout=timeout) as resp:
+                data = json.loads(resp.read().decode())
+            items = data if isinstance(data, list) else data.get("data", [])
+            return [m["id"] for m in items if isinstance(m, dict) and "id" in m]
+        except Exception as exc:
+            logger.debug("fetch_models(%s): %s", self.name, exc)
+            return None

providers/bedrock.py

@@ -0,0 +1,29 @@
+"""AWS Bedrock provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+
+class BedrockProfile(ProviderProfile):
+    """AWS Bedrock — no REST /v1/models endpoint; uses AWS SDK."""
+
+    def fetch_models(
+        self,
+        *,
+        api_key: str | None = None,
+        timeout: float = 8.0,
+    ) -> list[str] | None:
+        """Bedrock model listing requires AWS SDK, not a REST call."""
+        return None
+
+
+bedrock = BedrockProfile(
+    name="bedrock",
+    aliases=("aws", "aws-bedrock", "amazon-bedrock", "amazon"),
+    api_mode="bedrock_converse",
+    env_vars=(),  # AWS SDK credentials — not env vars
+    base_url="https://bedrock-runtime.us-east-1.amazonaws.com",
+    auth_type="aws_sdk",
+)
+
+register_provider(bedrock)

providers/copilot.py

@@ -0,0 +1,58 @@
+"""Copilot / GitHub Models provider profile.
+
+Copilot uses per-model api_mode routing:
+  - GPT-5+ / Codex models → codex_responses
+  - Claude models → anthropic_messages
+  - Everything else → chat_completions (this profile covers that subset)
+
+Key quirks for the chat_completions subset:
+  - Editor attribution headers (via copilot_default_headers())
+  - GitHub Models reasoning extra_body (model-catalog gated)
+"""
+
+from typing import Any
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+
+class CopilotProfile(ProviderProfile):
+    """GitHub Copilot / GitHub Models — editor headers + reasoning."""
+
+    def build_api_kwargs_extras(
+        self,
+        *,
+        model: str | None = None,
+        reasoning_config: dict | None = None,
+        supports_reasoning: bool = False,
+        **ctx,
+    ) -> tuple[dict[str, Any], dict[str, Any]]:
+        extra_body: dict[str, Any] = {}
+        if supports_reasoning and model:
+            try:
+                from hermes_cli.models import github_model_reasoning_efforts
+
+                supported_efforts = github_model_reasoning_efforts(model)
+                if supported_efforts and reasoning_config:
+                    effort = reasoning_config.get("effort", "medium")
+                    # Normalize non-standard effort levels to the nearest supported
+                    if effort == "xhigh":
+                        effort = "high"
+                    if effort in supported_efforts:
+                        extra_body["reasoning"] = {"effort": effort}
+                elif supported_efforts:
+                    extra_body["reasoning"] = {"effort": "medium"}
+            except Exception:
+                pass
+        return extra_body, {}
+
+
+copilot = CopilotProfile(
+    name="copilot",
+    aliases=("github-copilot", "github-models", "github-model", "github"),
+    env_vars=("COPILOT_GITHUB_TOKEN", "GH_TOKEN", "GITHUB_TOKEN"),
+    base_url="https://api.githubcopilot.com",
+    auth_type="copilot",
+)
+
+register_provider(copilot)

providers/copilot_acp.py

@@ -0,0 +1,34 @@
+"""GitHub Copilot ACP provider profile.
+
+copilot-acp uses an external ACP subprocess — NOT the standard
+transport. api_mode="copilot_acp" is handled separately in run_agent.py.
+The profile captures auth + endpoint metadata for registry migration.
+"""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+
+class CopilotACPProfile(ProviderProfile):
+    """GitHub Copilot ACP — external process, no REST models endpoint."""
+
+    def fetch_models(
+        self,
+        *,
+        api_key: str | None = None,
+        timeout: float = 8.0,
+    ) -> list[str] | None:
+        """Model listing is handled by the ACP subprocess."""
+        return None
+
+
+copilot_acp = CopilotACPProfile(
+    name="copilot-acp",
+    aliases=("github-copilot-acp", "copilot-acp-agent"),
+    api_mode="chat_completions",  # ACP subprocess uses chat_completions routing
+    env_vars=(),  # Managed by ACP subprocess
+    base_url="acp://copilot",  # ACP internal scheme
+    auth_type="external_process",
+)
+
+register_provider(copilot_acp)

providers/custom.py

@@ -0,0 +1,71 @@
+"""Custom / Ollama (local) provider profile.
+
+Covers any endpoint registered as provider="custom", including local
+Ollama instances. Key quirks:
+  - ollama_num_ctx → extra_body.options.num_ctx (local context window)
+  - reasoning_config disabled → extra_body.think = False
+"""
+
+from typing import Any
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+
+class CustomProfile(ProviderProfile):
+    """Custom/Ollama local provider — think=false and num_ctx support."""
+
+    def build_api_kwargs_extras(
+        self,
+        *,
+        reasoning_config: dict | None = None,
+        ollama_num_ctx: int | None = None,
+        **ctx: Any,
+    ) -> tuple[dict[str, Any], dict[str, Any]]:
+        extra_body: dict[str, Any] = {}
+
+        # Ollama context window
+        if ollama_num_ctx:
+            options = extra_body.get("options", {})
+            options["num_ctx"] = ollama_num_ctx
+            extra_body["options"] = options
+
+        # Disable thinking when reasoning is turned off
+        if reasoning_config and isinstance(reasoning_config, dict):
+            _effort = (reasoning_config.get("effort") or "").strip().lower()
+            _enabled = reasoning_config.get("enabled", True)
+            if _effort == "none" or _enabled is False:
+                extra_body["think"] = False
+
+        return extra_body, {}
+
+    def fetch_models(
+        self,
+        *,
+        api_key: str | None = None,
+        timeout: float = 8.0,
+    ) -> list[str] | None:
+        """Custom/Ollama: base_url is user-configured; fetch if set."""
+        if not self.base_url:
+            return None
+        return super().fetch_models(api_key=api_key, timeout=timeout)
+
+
+custom = CustomProfile(
+    name="custom",
+    aliases=(
+        "ollama",
+        "local",
+        "lmstudio",
+        "lm-studio",
+        "lm_studio",
+        "vllm",
+        "llamacpp",
+        "llama.cpp",
+        "llama-cpp",
+    ),
+    env_vars=(),  # No fixed key — custom endpoint
+    base_url="",  # User-configured
+)
+
+register_provider(custom)

providers/deepseek.py

@@ -0,0 +1,20 @@
+"""DeepSeek provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+deepseek = ProviderProfile(
+    name="deepseek",
+    aliases=("deepseek-chat",),
+    env_vars=("DEEPSEEK_API_KEY",),
+    display_name="DeepSeek",
+    description="DeepSeek — native DeepSeek API",
+    signup_url="https://platform.deepseek.com/",
+    fallback_models=(
+        "deepseek-chat",
+        "deepseek-reasoner",
+    ),
+    base_url="https://api.deepseek.com/v1",
+)
+
+register_provider(deepseek)

providers/gemini.py

@@ -0,0 +1,34 @@
+"""Google Gemini provider profiles.
+
+gemini:            Google AI Studio (API key) — uses GeminiNativeClient
+google-gemini-cli: Google Cloud Code Assist (OAuth) — uses GeminiCloudCodeClient
+
+Both report api_mode="chat_completions" but use custom native clients
+that bypass the standard OpenAI transport. The profile captures auth
+and endpoint metadata for auth.py / runtime_provider.py migration.
+"""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+gemini = ProviderProfile(
+    name="gemini",
+    aliases=("google", "google-gemini", "google-ai-studio"),
+    api_mode="chat_completions",
+    env_vars=("GOOGLE_API_KEY", "GEMINI_API_KEY"),
+    base_url="https://generativelanguage.googleapis.com/v1beta",
+    auth_type="api_key",
+    default_aux_model="gemini-3-flash-preview",
+)
+
+google_gemini_cli = ProviderProfile(
+    name="google-gemini-cli",
+    aliases=("gemini-cli", "gemini-oauth"),
+    api_mode="chat_completions",
+    env_vars=(),  # OAuth — no API key
+    base_url="cloudcode-pa://google",  # Cloud Code Assist internal scheme
+    auth_type="oauth_external",
+)
+
+register_provider(gemini)
+register_provider(google_gemini_cli)

providers/gmi.py

@@ -0,0 +1,26 @@
+"""GMI Cloud provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+gmi = ProviderProfile(
+    name="gmi",
+    aliases=("gmi-cloud", "gmicloud"),
+    display_name="GMI Cloud",
+    description="GMI Cloud — multi-model direct API (slash-form model IDs)",
+    signup_url="https://www.gmicloud.ai/",
+    env_vars=("GMI_API_KEY", "GMI_BASE_URL"),
+    base_url="https://api.gmi-serving.com/v1",
+    auth_type="api_key",
+    default_aux_model="google/gemini-3.1-flash-lite-preview",
+    fallback_models=(
+        "zai-org/GLM-5.1-FP8",
+        "deepseek-ai/DeepSeek-V3.2",
+        "moonshotai/Kimi-K2.5",
+        "google/gemini-3.1-flash-lite-preview",
+        "anthropic/claude-sonnet-4.6",
+        "openai/gpt-5.4",
+    ),
+)
+
+register_provider(gmi)

providers/huggingface.py

@@ -0,0 +1,20 @@
+"""Hugging Face provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+huggingface = ProviderProfile(
+    name="huggingface",
+    aliases=("hf", "hugging-face", "huggingface-hub"),
+    env_vars=("HF_TOKEN",),
+    display_name="HuggingFace",
+    description="HuggingFace Inference API",
+    signup_url="https://huggingface.co/settings/tokens",
+    fallback_models=(
+        "Qwen/Qwen3.5-72B-Instruct",
+        "deepseek-ai/DeepSeek-V3.2",
+    ),
+    base_url="https://router.huggingface.co/v1",
+)
+
+register_provider(huggingface)

providers/kilocode.py

@@ -0,0 +1,14 @@
+"""Kilo Code provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+kilocode = ProviderProfile(
+    name="kilocode",
+    aliases=("kilo-code", "kilo", "kilo-gateway"),
+    env_vars=("KILOCODE_API_KEY",),
+    base_url="https://api.kilo.ai/api/gateway",
+    default_aux_model="google/gemini-3-flash-preview",
+)
+
+register_provider(kilocode)

providers/kimi.py

@@ -0,0 +1,71 @@
+"""Kimi / Moonshot provider profiles.
+
+Kimi has dual endpoints:
+  - sk-kimi-* keys → api.kimi.com/coding (Anthropic Messages API)
+  - legacy keys → api.moonshot.ai/v1 (OpenAI chat completions)
+
+This module covers the chat_completions path (/v1 endpoint).
+"""
+
+from typing import Any
+
+from providers import register_provider
+from providers.base import OMIT_TEMPERATURE, ProviderProfile
+
+
+class KimiProfile(ProviderProfile):
+    """Kimi/Moonshot — temperature omitted, thinking + reasoning_effort."""
+
+    def build_api_kwargs_extras(
+        self, *, reasoning_config: dict | None = None, **context
+    ) -> tuple[dict[str, Any], dict[str, Any]]:
+        """Kimi uses extra_body.thinking + top-level reasoning_effort."""
+        extra_body = {}
+        top_level = {}
+
+        if not reasoning_config or not isinstance(reasoning_config, dict):
+            # No config → thinking enabled, default effort
+            extra_body["thinking"] = {"type": "enabled"}
+            top_level["reasoning_effort"] = "medium"
+            return extra_body, top_level
+
+        enabled = reasoning_config.get("enabled", True)
+        if enabled is False:
+            extra_body["thinking"] = {"type": "disabled"}
+            return extra_body, top_level
+
+        # Enabled
+        extra_body["thinking"] = {"type": "enabled"}
+        effort = (reasoning_config.get("effort") or "").strip().lower()
+        if effort in ("low", "medium", "high"):
+            top_level["reasoning_effort"] = effort
+        else:
+            top_level["reasoning_effort"] = "medium"
+
+        return extra_body, top_level
+
+
+kimi = KimiProfile(
+    name="kimi-coding",
+    aliases=("kimi", "moonshot", "kimi-for-coding"),
+    env_vars=("KIMI_API_KEY", "KIMI_CODING_API_KEY"),
+    base_url="https://api.moonshot.ai/v1",
+    fixed_temperature=OMIT_TEMPERATURE,
+    default_max_tokens=32000,
+    default_headers={"User-Agent": "hermes-agent/1.0"},
+    default_aux_model="kimi-k2-turbo-preview",
+)
+
+kimi_cn = KimiProfile(
+    name="kimi-coding-cn",
+    aliases=("kimi-cn", "moonshot-cn"),
+    env_vars=("KIMI_CN_API_KEY",),
+    base_url="https://api.moonshot.cn/v1",
+    fixed_temperature=OMIT_TEMPERATURE,
+    default_max_tokens=32000,
+    default_headers={"User-Agent": "hermes-agent/1.0"},
+    default_aux_model="kimi-k2-turbo-preview",
+)
+
+register_provider(kimi)
+register_provider(kimi_cn)

providers/minimax.py

@@ -0,0 +1,31 @@
+"""MiniMax provider profiles (international + China).
+
+Both use anthropic_messages api_mode — their inference_base_url
+ends with /anthropic which triggers auto-detection to anthropic_messages.
+"""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+minimax = ProviderProfile(
+    name="minimax",
+    aliases=("mini-max",),
+    api_mode="anthropic_messages",
+    env_vars=("MINIMAX_API_KEY",),
+    base_url="https://api.minimax.io/anthropic",
+    auth_type="api_key",
+    default_aux_model="MiniMax-M2.7",
+)
+
+minimax_cn = ProviderProfile(
+    name="minimax-cn",
+    aliases=("minimax-china", "minimax_cn"),
+    api_mode="anthropic_messages",
+    env_vars=("MINIMAX_CN_API_KEY",),
+    base_url="https://api.minimaxi.com/anthropic",
+    auth_type="api_key",
+    default_aux_model="MiniMax-M2.7",
+)
+
+register_provider(minimax)
+register_provider(minimax_cn)

providers/nous.py

@@ -0,0 +1,53 @@
+"""Nous Portal provider profile."""
+
+from typing import Any
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+
+class NousProfile(ProviderProfile):
+    """Nous Portal — product tags, reasoning with Nous-specific omission."""
+
+    def build_extra_body(
+        self, *, session_id: str | None = None, **context
+    ) -> dict[str, Any]:
+        return {"tags": ["product=hermes-agent"]}
+
+    def build_api_kwargs_extras(
+        self,
+        *,
+        reasoning_config: dict | None = None,
+        supports_reasoning: bool = False,
+        **context,
+    ) -> tuple[dict[str, Any], dict[str, Any]]:
+        """Nous: passes full reasoning_config, but OMITS when disabled."""
+        extra_body = {}
+        if supports_reasoning:
+            if reasoning_config is not None:
+                rc = dict(reasoning_config)
+                if rc.get("enabled") is False:
+                    pass  # Nous omits reasoning when disabled
+                else:
+                    extra_body["reasoning"] = rc
+            else:
+                extra_body["reasoning"] = {"enabled": True, "effort": "medium"}
+        return extra_body, {}
+
+
+nous = NousProfile(
+    name="nous",
+    aliases=("nous-portal", "nousresearch"),
+    env_vars=("NOUS_API_KEY",),
+    display_name="Nous Research",
+    description="Nous Research — Hermes model family",
+    signup_url="https://nousresearch.com/",
+    fallback_models=(
+        "hermes-3-405b",
+        "hermes-3-70b",
+    ),
+    base_url="https://inference.nousresearch.com/v1",
+    auth_type="oauth_device_code",
+)
+
+register_provider(nous)

providers/nvidia.py

@@ -0,0 +1,21 @@
+"""NVIDIA NIM provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+nvidia = ProviderProfile(
+    name="nvidia",
+    aliases=("nvidia-nim",),
+    env_vars=("NVIDIA_API_KEY",),
+    display_name="NVIDIA NIM",
+    description="NVIDIA NIM — accelerated inference",
+    signup_url="https://build.nvidia.com/",
+    fallback_models=(
+        "nvidia/llama-3.1-nemotron-70b-instruct",
+        "nvidia/llama-3.3-70b-instruct",
+    ),
+    base_url="https://integrate.api.nvidia.com/v1",
+    default_max_tokens=16384,
+)
+
+register_provider(nvidia)

providers/ollama_cloud.py

@@ -0,0 +1,14 @@
+"""Ollama Cloud provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+ollama_cloud = ProviderProfile(
+    name="ollama-cloud",
+    aliases=("ollama_cloud",),
+    default_aux_model="nemotron-3-nano:30b",
+    env_vars=("OLLAMA_API_KEY",),
+    base_url="https://ollama.com/v1",
+)
+
+register_provider(ollama_cloud)

providers/openai_codex.py

@@ -0,0 +1,15 @@
+"""OpenAI Codex (Responses API) provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+openai_codex = ProviderProfile(
+    name="openai-codex",
+    aliases=("codex", "openai_codex"),
+    api_mode="codex_responses",
+    env_vars=(),  # OAuth external — no API key
+    base_url="https://chatgpt.com/backend-api/codex",
+    auth_type="oauth_external",
+)
+
+register_provider(openai_codex)

providers/opencode.py

@@ -0,0 +1,30 @@
+"""OpenCode provider profiles (Zen + Go).
+
+Both use per-model api_mode routing:
+  - OpenCode Zen: Claude → anthropic_messages, GPT-5/Codex → codex_responses,
+    everything else → chat_completions (this profile)
+  - OpenCode Go: MiniMax → anthropic_messages, GLM/Kimi → chat_completions
+    (this profile)
+"""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+opencode_zen = ProviderProfile(
+    name="opencode-zen",
+    aliases=("opencode", "opencode_zen", "zen"),
+    env_vars=("OPENCODE_ZEN_API_KEY",),
+    base_url="https://opencode.ai/zen/v1",
+    default_aux_model="gemini-3-flash",
+)
+
+opencode_go = ProviderProfile(
+    name="opencode-go",
+    aliases=("opencode_go", "go", "opencode-go-sub"),
+    env_vars=("OPENCODE_GO_API_KEY",),
+    base_url="https://opencode.ai/zen/go/v1",
+    default_aux_model="glm-5",
+)
+
+register_provider(opencode_zen)
+register_provider(opencode_go)

providers/openrouter.py

@@ -0,0 +1,86 @@
+"""OpenRouter provider profile."""
+
+import logging
+from typing import Any
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+logger = logging.getLogger(__name__)
+
+_CACHE: list[str] | None = None
+
+
+class OpenRouterProfile(ProviderProfile):
+    """OpenRouter aggregator — provider preferences, reasoning config passthrough."""
+
+    def fetch_models(
+        self,
+        *,
+        api_key: str | None = None,
+        timeout: float = 8.0,
+    ) -> list[str] | None:
+        """Fetch from public OpenRouter catalog — no auth required.
+
+        Note: Tool-call capability filtering is applied by hermes_cli/models.py
+        via fetch_openrouter_models() → _openrouter_model_supports_tools(), not
+        here. The picker early-returns via the dedicated openrouter path before
+        reaching this method, so filtering here would be unreachable.
+        """
+        global _CACHE  # noqa: PLW0603
+        if _CACHE is not None:
+            return _CACHE
+        try:
+            result = super().fetch_models(api_key=None, timeout=timeout)
+            if result is not None:
+                _CACHE = result
+            return result
+        except Exception as exc:
+            logger.debug("fetch_models(openrouter): %s", exc)
+            return None
+
+    def build_extra_body(
+        self, *, session_id: str | None = None, **context: Any
+    ) -> dict[str, Any]:
+        body: dict[str, Any] = {}
+        prefs = context.get("provider_preferences")
+        if prefs:
+            body["provider"] = prefs
+        return body
+
+    def build_api_kwargs_extras(
+        self,
+        *,
+        reasoning_config: dict | None = None,
+        supports_reasoning: bool = False,
+        **context: Any,
+    ) -> tuple[dict[str, Any], dict[str, Any]]:
+        """OpenRouter passes the full reasoning_config dict as extra_body.reasoning."""
+        extra_body: dict[str, Any] = {}
+        if supports_reasoning:
+            if reasoning_config is not None:
+                extra_body["reasoning"] = dict(reasoning_config)
+            else:
+                extra_body["reasoning"] = {"enabled": True, "effort": "medium"}
+        return extra_body, {}
+
+
+openrouter = OpenRouterProfile(
+    name="openrouter",
+    aliases=("or",),
+    env_vars=("OPENROUTER_API_KEY",),
+    display_name="OpenRouter",
+    description="OpenRouter — unified API for 200+ models",
+    signup_url="https://openrouter.ai/keys",
+    base_url="https://openrouter.ai/api/v1",
+    models_url="https://openrouter.ai/api/v1/models",
+    fallback_models=(
+        "anthropic/claude-sonnet-4.6",
+        "openai/gpt-5.4",
+        "deepseek/deepseek-chat",
+        "google/gemini-3-flash-preview",
+        "qwen/qwen3-plus",
+    ),
+)
+
+register_provider(openrouter)

providers/qwen.py

@@ -0,0 +1,82 @@
+"""Qwen Portal provider profile."""
+
+import copy
+from typing import Any
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+
+class QwenProfile(ProviderProfile):
+    """Qwen Portal — message normalization, vl_high_resolution, metadata top-level."""
+
+    def prepare_messages(self, messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        """Normalize content to list-of-dicts format.
+
+        Inject cache_control on system message.
+
+        Matches the behavior of run_agent.py:_qwen_prepare_chat_messages().
+        """
+        prepared = copy.deepcopy(messages)
+        if not prepared:
+            return prepared
+
+        for msg in prepared:
+            if not isinstance(msg, dict):
+                continue
+            content = msg.get("content")
+            if isinstance(content, str):
+                msg["content"] = [{"type": "text", "text": content}]
+            elif isinstance(content, list):
+                normalized_parts = []
+                for part in content:
+                    if isinstance(part, str):
+                        normalized_parts.append({"type": "text", "text": part})
+                    elif isinstance(part, dict):
+                        normalized_parts.append(part)
+                if normalized_parts:
+                    msg["content"] = normalized_parts
+
+        # Inject cache_control on the last part of the system message.
+        for msg in prepared:
+            if isinstance(msg, dict) and msg.get("role") == "system":
+                content = msg.get("content")
+                if (
+                    isinstance(content, list)
+                    and content
+                    and isinstance(content[-1], dict)
+                ):
+                    content[-1]["cache_control"] = {"type": "ephemeral"}
+                break
+
+        return prepared
+
+    def build_extra_body(
+        self, *, session_id: str | None = None, **context
+    ) -> dict[str, Any]:
+        return {"vl_high_resolution_images": True}
+
+    def build_api_kwargs_extras(
+        self,
+        *,
+        reasoning_config: dict | None = None,
+        qwen_session_metadata: dict | None = None,
+        **context,
+    ) -> tuple[dict[str, Any], dict[str, Any]]:
+        """Qwen metadata goes to top-level api_kwargs, not extra_body."""
+        top_level = {}
+        if qwen_session_metadata:
+            top_level["metadata"] = qwen_session_metadata
+        return {}, top_level
+
+
+qwen = QwenProfile(
+    name="qwen-oauth",
+    aliases=("qwen", "qwen-portal", "qwen-cli"),
+    env_vars=("QWEN_API_KEY",),
+    base_url="https://portal.qwen.ai/v1",
+    auth_type="oauth_external",
+    default_max_tokens=65536,
+)
+
+register_provider(qwen)

providers/stepfun.py

@@ -0,0 +1,14 @@
+"""StepFun provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+stepfun = ProviderProfile(
+    name="stepfun",
+    aliases=("step", "stepfun-coding-plan"),
+    default_aux_model="step-3.5-flash",
+    env_vars=("STEPFUN_API_KEY",),
+    base_url="https://api.stepfun.ai/step_plan/v1",
+)
+
+register_provider(stepfun)

providers/vercel.py

@@ -0,0 +1,43 @@
+"""Vercel AI Gateway provider profile.
+
+AI Gateway routes to multiple backends. Hermes sends attribution
+headers and full reasoning config passthrough.
+"""
+
+from typing import Any
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+
+class VercelAIGatewayProfile(ProviderProfile):
+    """Vercel AI Gateway — attribution headers + reasoning passthrough."""
+
+    def build_api_kwargs_extras(
+        self,
+        *,
+        reasoning_config: dict | None = None,
+        supports_reasoning: bool = True,
+        **ctx: Any,
+    ) -> tuple[dict[str, Any], dict[str, Any]]:
+        extra_body: dict[str, Any] = {}
+        if supports_reasoning and reasoning_config is not None:
+            extra_body["reasoning"] = dict(reasoning_config)
+        elif supports_reasoning:
+            extra_body["reasoning"] = {"enabled": True, "effort": "medium"}
+        return extra_body, {}
+
+
+vercel = VercelAIGatewayProfile(
+    name="ai-gateway",
+    aliases=("vercel", "vercel-ai-gateway", "ai_gateway", "aigateway"),
+    env_vars=("AI_GATEWAY_API_KEY",),
+    base_url="https://ai-gateway.vercel.sh/v1",
+    default_headers={
+        "HTTP-Referer": "https://hermes-agent.nousresearch.com",
+        "X-Title": "Hermes Agent",
+    },
+    default_aux_model="google/gemini-3-flash",
+)
+
+register_provider(vercel)

providers/xai.py

@@ -0,0 +1,15 @@
+"""xAI (Grok) provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+xai = ProviderProfile(
+    name="xai",
+    aliases=("grok", "x-ai", "x.ai"),
+    api_mode="codex_responses",
+    env_vars=("XAI_API_KEY",),
+    base_url="https://api.x.ai/v1",
+    auth_type="api_key",
+)
+
+register_provider(xai)

providers/xiaomi.py

@@ -0,0 +1,13 @@
+"""Xiaomi MiMo provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+xiaomi = ProviderProfile(
+    name="xiaomi",
+    aliases=("mimo", "xiaomi-mimo"),
+    env_vars=("XIAOMI_API_KEY",),
+    base_url="https://api.xiaomimimo.com/v1",
+)
+
+register_provider(xiaomi)

providers/zai.py

@@ -0,0 +1,21 @@
+"""ZAI / GLM provider profile."""
+
+from providers import register_provider
+from providers.base import ProviderProfile
+
+zai = ProviderProfile(
+    name="zai",
+    aliases=("glm", "z-ai", "z.ai", "zhipu"),
+    env_vars=("GLM_API_KEY", "ZAI_API_KEY", "Z_AI_API_KEY"),
+    display_name="Z.AI (GLM)",
+    description="Z.AI / GLM — Zhipu AI models",
+    signup_url="https://z.ai/",
+    fallback_models=(
+        "glm-5",
+        "glm-4-9b",
+    ),
+    base_url="https://api.z.ai/api/paas/v4",
+    default_aux_model="glm-4.5-flash",
+)
+
+register_provider(zai)

pyproject.toml

@@ -137,7 +137,7 @@ py-modules = ["run_agent", "model_tools", "toolsets", "batch_runner", "trajector
 hermes_cli = ["web_dist/**/*"]
 
 [tool.setuptools.packages.find]
-include = ["agent", "agent.*", "tools", "tools.*", "hermes_cli", "gateway", "gateway.*", "tui_gateway", "tui_gateway.*", "cron", "acp_adapter", "plugins", "plugins.*"]
+include = ["agent", "agent.*", "tools", "tools.*", "hermes_cli", "gateway", "gateway.*", "tui_gateway", "tui_gateway.*", "cron", "acp_adapter", "plugins", "plugins.*", "providers", "providers.*"]
 
 [tool.pytest.ini_options]
 testpaths = ["tests"]

run_agent.py

@@ -86,7 +86,7 @@ from tools.browser_tool import cleanup_browser
 
 
 # Agent internals extracted to agent/ package for modularity
-from agent.memory_manager import build_memory_context_block, sanitize_context
+from agent.memory_manager import StreamingContextScrubber, build_memory_context_block, sanitize_context
 from agent.retry_utils import jittered_backoff
 from agent.error_classifier import classify_api_error, FailoverReason
 from agent.prompt_builder import (
@@ -1218,6 +1218,10 @@ class AIAgent:
         # Deferred paragraph break flag — set after tool iterations so a
         # single "\n\n" is prepended to the next real text delta.
         self._stream_needs_break = False
+        # Stateful scrubber for <memory-context> spans split across stream
+        # deltas (#5719).  sanitize_context() alone can't survive chunk
+        # boundaries because the block regex needs both tags in one string.
+        self._stream_context_scrubber = StreamingContextScrubber()
         # Visible assistant text already delivered through live token callbacks
         # during the current model response. Used to avoid re-sending the same
         # commentary when the provider later returns it as a completed interim
@@ -1367,6 +1371,17 @@ class AIAgent:
                 elif base_url_host_matches(effective_base, "chatgpt.com"):
                     from agent.auxiliary_client import _codex_cloudflare_headers
                     client_kwargs["default_headers"] = _codex_cloudflare_headers(api_key)
+                elif "default_headers" not in client_kwargs:
+                    # Fall back to profile.default_headers for providers that
+                    # declare custom headers (e.g. Vercel AI Gateway attribution,
+                    # Kimi User-Agent on non-kimi.com endpoints).
+                    try:
+                        from providers import get_provider_profile as _gpf
+                        _ph = _gpf(self.provider)
+                        if _ph and _ph.default_headers:
+                            client_kwargs["default_headers"] = dict(_ph.default_headers)
+                    except Exception:
+                        pass
             else:
                 # No explicit creds — use the centralized provider router
                 from agent.auxiliary_client import resolve_provider_client
@@ -5033,7 +5048,7 @@ class AIAgent:
         _validate_proxy_env_urls()
         _validate_base_url(client_kwargs.get("base_url"))
         if self.provider == "copilot-acp" or str(client_kwargs.get("base_url", "")).startswith("acp://copilot"):
-            from agent.copilot_acp_client import CopilotACPClient
+            from acp_adapter.copilot_client import CopilotACPClient
 
             client = CopilotACPClient(**client_kwargs)
             logger.info(
@@ -5722,7 +5737,19 @@ class AIAgent:
                 self._client_kwargs.get("api_key", "")
             )
         else:
-            self._client_kwargs.pop("default_headers", None)
+            # No URL-specific headers — check profile.default_headers before clearing.
+            _ph_headers = None
+            try:
+                from providers import get_provider_profile as _gpf2
+                _ph2 = _gpf2(self.provider)
+                if _ph2 and _ph2.default_headers:
+                    _ph_headers = dict(_ph2.default_headers)
+            except Exception:
+                pass
+            if _ph_headers:
+                self._client_kwargs["default_headers"] = _ph_headers
+            else:
+                self._client_kwargs.pop("default_headers", None)
 
     def _swap_credential(self, entry) -> None:
         runtime_key = getattr(entry, "runtime_api_key", None) or getattr(entry, "access_token", "")
@@ -6019,6 +6046,20 @@ class AIAgent:
 
     def _reset_stream_delivery_tracking(self) -> None:
         """Reset tracking for text delivered during the current model response."""
+        # Flush any benign partial-tag tail held by the context scrubber so it
+        # reaches the UI before we clear state for the next model call.  If
+        # the scrubber is mid-span, flush() drops the orphaned content.
+        scrubber = getattr(self, "_stream_context_scrubber", None)
+        if scrubber is not None:
+            tail = scrubber.flush()
+            if tail:
+                callbacks = [cb for cb in (self.stream_delta_callback, self._stream_callback) if cb is not None]
+                for cb in callbacks:
+                    try:
+                        cb(tail)
+                    except Exception:
+                        pass
+                self._record_streamed_assistant_text(tail)
         self._current_streamed_assistant_text = ""
 
     def _record_streamed_assistant_text(self, text: str) -> None:
@@ -6069,6 +6110,28 @@ class AIAgent:
         if getattr(self, "_stream_needs_break", False) and text and text.strip():
             self._stream_needs_break = False
             text = "\n\n" + text
+            prepended_break = True
+        else:
+            prepended_break = False
+        if isinstance(text, str):
+            # Strip <think> blocks first (per-delta is safe for closed pairs; the
+            # unterminated-tag path is handled downstream by stream_consumer).
+            # Then feed through the stateful context scrubber so memory-context
+            # spans split across chunks cannot leak to the UI (#5719).
+            text = self._strip_think_blocks(text or "")
+            scrubber = getattr(self, "_stream_context_scrubber", None)
+            if scrubber is not None:
+                text = scrubber.feed(text)
+            else:
+                # Defensive: legacy callers without the scrubber attribute.
+                text = sanitize_context(text)
+            # Only strip leading newlines on the first delta — mid-stream "\n" is legitimate markdown.
+            if not prepended_break and not getattr(
+                self, "_current_streamed_assistant_text", ""
+            ):
+                text = text.lstrip("\n")
+        if not text:
+            return
         callbacks = [cb for cb in (self.stream_delta_callback, self._stream_callback) if cb is not None]
         delivered = False
         for cb in callbacks:
@@ -7817,66 +7880,79 @@ class AIAgent:
         # ── chat_completions (default) ─────────────────────────────────────
         _ct = self._get_transport()
 
-        # Provider detection flags
-        _is_qwen = self._is_qwen_portal()
-        _is_or = self._is_openrouter_url()
-        _is_gh = (
-            base_url_host_matches(self._base_url_lower, "models.github.ai")
-            or base_url_host_matches(self._base_url_lower, "api.githubcopilot.com")
-        )
-        _is_nous = "nousresearch" in self._base_url_lower
-        _is_nvidia = "integrate.api.nvidia.com" in self._base_url_lower
-        _is_kimi = (
-            base_url_host_matches(self.base_url, "api.kimi.com")
-            or base_url_host_matches(self.base_url, "moonshot.ai")
-            or base_url_host_matches(self.base_url, "moonshot.cn")
-        )
-
-        # Temperature: _fixed_temperature_for_model may return OMIT_TEMPERATURE
-        # sentinel (temperature omitted entirely), a numeric override, or None.
+        # ── Provider profile path (all chat_completions providers) ─────────
+        # Profiles handle per-provider quirks via hooks. We compute the shared
+        # per-call context here and pass it through so hooks can use it.
         try:
-            from agent.auxiliary_client import _fixed_temperature_for_model, OMIT_TEMPERATURE
-            _ft = _fixed_temperature_for_model(self.model, self.base_url)
-            _omit_temp = _ft is OMIT_TEMPERATURE
-            _fixed_temp = _ft if not _omit_temp else None
+            from providers import get_provider_profile
+            _profile = get_provider_profile(self.provider)
         except Exception:
-            _omit_temp = False
-            _fixed_temp = None
+            _profile = None
 
-        # Provider preferences (OpenRouter-specific)
-        _prefs: Dict[str, Any] = {}
-        if self.providers_allowed:
-            _prefs["only"] = self.providers_allowed
-        if self.providers_ignored:
-            _prefs["ignore"] = self.providers_ignored
-        if self.providers_order:
-            _prefs["order"] = self.providers_order
-        if self.provider_sort:
-            _prefs["sort"] = self.provider_sort
-        if self.provider_require_parameters:
-            _prefs["require_parameters"] = True
-        if self.provider_data_collection:
-            _prefs["data_collection"] = self.provider_data_collection
+        if _profile:
+            _ephemeral_out = getattr(self, "_ephemeral_max_output_tokens", None)
+            if _ephemeral_out is not None:
+                self._ephemeral_max_output_tokens = None
 
-        # Anthropic max output for Claude on OpenRouter/Nous
-        _ant_max = None
-        if (_is_or or _is_nous) and "claude" in (self.model or "").lower():
-            try:
-                from agent.anthropic_adapter import _get_anthropic_max_output
-                _ant_max = _get_anthropic_max_output(self.model)
-            except Exception:
-                pass  # fail open — let the proxy pick its default
+            # Per-call context for profile hooks — mirrors the legacy flag block.
+            # Computed here so profiles receive live per-call values (not stale).
+            _prefs: Dict[str, Any] = {}
+            if self.providers_allowed:
+                _prefs["only"] = self.providers_allowed
+            if self.providers_ignored:
+                _prefs["ignore"] = self.providers_ignored
+            if self.providers_order:
+                _prefs["order"] = self.providers_order
+            if self.provider_sort:
+                _prefs["sort"] = self.provider_sort
+            if self.provider_require_parameters:
+                _prefs["require_parameters"] = True
+            if self.provider_data_collection:
+                _prefs["data_collection"] = self.provider_data_collection
 
-        # Qwen session metadata precomputed here (promptId is per-call random)
-        _qwen_meta = None
-        if _is_qwen:
-            _qwen_meta = {
-                "sessionId": self.session_id or "hermes",
-                "promptId": str(uuid.uuid4()),
-            }
+            _is_or = self._is_openrouter_url()
+            _is_nous = "nousresearch" in self._base_url_lower
+            _ant_max = None
+            if (_is_or or _is_nous) and "claude" in (self.model or "").lower():
+                try:
+                    from agent.anthropic_adapter import _get_anthropic_max_output
+                    _ant_max = _get_anthropic_max_output(self.model)
+                except Exception:
+                    pass
 
-        # Ephemeral max output override — consume immediately so the next
-        # turn doesn't inherit it.
+            _is_qwen = self._is_qwen_portal()
+            _qwen_meta = None
+            if _is_qwen:
+                _qwen_meta = {
+                    "sessionId": self.session_id or "hermes",
+                    "promptId": str(uuid.uuid4()),
+                }
+
+            return _ct.build_kwargs(
+                model=self.model,
+                messages=api_messages,
+                tools=self.tools,
+                timeout=self._resolved_api_call_timeout(),
+                max_tokens=self.max_tokens,
+                ephemeral_max_output_tokens=_ephemeral_out,
+                max_tokens_param_fn=self._max_tokens_param,
+                reasoning_config=self.reasoning_config,
+                request_overrides=self.request_overrides,
+                session_id=getattr(self, "session_id", None),
+                provider_profile=_profile,
+                ollama_num_ctx=self._ollama_num_ctx,
+                # Context forwarded to profile hooks:
+                provider_preferences=_prefs or None,
+                anthropic_max_output=_ant_max,
+                supports_reasoning=self._supports_reasoning_extra_body(),
+                qwen_session_metadata=_qwen_meta,
+            )
+
+        # ── Legacy flag path ────────────────────────────────────────────
+        # Reached only when get_provider_profile() returns None — i.e. a
+        # completely unknown provider not in providers/ registry.
+        # Best-effort: send a clean chat_completions request with no
+        # provider-specific quirks.
         _ephemeral_out = getattr(self, "_ephemeral_max_output_tokens", None)
         if _ephemeral_out is not None:
             self._ephemeral_max_output_tokens = None
@@ -7895,24 +7971,7 @@ class AIAgent:
             reasoning_config=self.reasoning_config,
             request_overrides=self.request_overrides,
             session_id=getattr(self, "session_id", None),
-            model_lower=(self.model or "").lower(),
-            is_openrouter=_is_or,
-            is_nous=_is_nous,
-            is_qwen_portal=_is_qwen,
-            is_github_models=_is_gh,
-            is_nvidia_nim=_is_nvidia,
-            is_kimi=_is_kimi,
-            is_custom_provider=self.provider == "custom",
             ollama_num_ctx=self._ollama_num_ctx,
-            provider_preferences=_prefs or None,
-            qwen_prepare_fn=self._qwen_prepare_chat_messages if _is_qwen else None,
-            qwen_prepare_inplace_fn=self._qwen_prepare_chat_messages_inplace if _is_qwen else None,
-            qwen_session_metadata=_qwen_meta,
-            fixed_temperature=_fixed_temp,
-            omit_temperature=_omit_temp,
-            supports_reasoning=self._supports_reasoning_extra_body(),
-            github_reasoning_extra=self._github_models_reasoning_extra_body() if _is_gh else None,
-            anthropic_max_output=_ant_max,
         )
 
     def _supports_reasoning_extra_body(self) -> bool:
@@ -8420,6 +8479,23 @@ class AIAgent:
                     f"⚠ Compression summary failed: {summary_error}. "
                     "Inserted a fallback context marker."
                 )
+        else:
+            # No hard failure — but did the configured aux model error out
+            # and get recovered by retrying on main?  Surface that so users
+            # know their auxiliary.compression.model setting is broken even
+            # though compression succeeded.
+            _aux_fail_model = getattr(self.context_compressor, "_last_aux_model_failure_model", None)
+            _aux_fail_err = getattr(self.context_compressor, "_last_aux_model_failure_error", None)
+            if _aux_fail_model:
+                # Dedup on (model, error) so we don't spam on every compaction
+                _aux_key = (_aux_fail_model, _aux_fail_err)
+                if getattr(self, "_last_aux_fallback_warning_key", None) != _aux_key:
+                    self._last_aux_fallback_warning_key = _aux_key
+                    self._emit_warning(
+                        f"ℹ Configured compression model '{_aux_fail_model}' failed "
+                        f"({_aux_fail_err or 'unknown error'}). Recovered using main model — "
+                        "check auxiliary.compression.model in config.yaml."
+                    )
 
         todo_snapshot = self._todo_store.format_for_injection()
         if todo_snapshot:
@@ -9592,16 +9668,6 @@ class AIAgent:
         if isinstance(persist_user_message, str):
             persist_user_message = _sanitize_surrogates(persist_user_message)
 
-        # Strip leaked <memory-context> blocks from user input.  When Honcho's
-        # saveMessages persists a turn that included injected context, the block
-        # can reappear in the next turn's user message via message history.
-        # Stripping here prevents stale memory tags from leaking into the
-        # conversation and being visible to the user or the model as user text.
-        if isinstance(user_message, str):
-            user_message = sanitize_context(user_message)
-        if isinstance(persist_user_message, str):
-            persist_user_message = sanitize_context(persist_user_message)
-
         # Store stream callback for _interruptible_api_call to pick up
         self._stream_callback = stream_callback
         self._persist_user_message_idx = None
@@ -9680,6 +9746,13 @@ class AIAgent:
         # Track user turns for memory flush and periodic nudge logic
         self._user_turn_count += 1
 
+        # Reset the streaming context scrubber at the top of each turn so a
+        # hung span from a prior interrupted stream can't taint this turn's
+        # output.
+        scrubber = getattr(self, "_stream_context_scrubber", None)
+        if scrubber is not None:
+            scrubber.reset()
+
         # Preserve the original user message (no nudge injection).
         original_user_message = persist_user_message if persist_user_message is not None else user_message
 
@@ -12711,7 +12784,6 @@ class AIAgent:
                         truncated_response_prefix = ""
                         length_continue_retries = 0
                     
-                    # Strip <think> blocks from user-facing response (keep raw in messages for trajectory)
                     final_response = self._strip_think_blocks(final_response).strip()
                     
                     final_msg = self._build_assistant_message(assistant_message, finish_reason)

scripts/release.py

@@ -557,6 +557,12 @@ AUTHOR_MAP = {
     "mor.aleksandr@yahoo.com": "MorAlekss",
     "ash@users.noreply.github.com": "ash",
     "andrewho.sf@gmail.com": "andrewhosf",
+    # April 2026 Honcho bug-fix consolidation (#15381)
+    "HiddenPuppy@users.noreply.github.com": "HiddenPuppy",
+    "code@sasha.id": "sasha-id",
+    "dontcallmejames@users.noreply.github.com": "dontcallmejames",
+    "hekaru.agent@gmail.com": "hekaru-agent",
+    "jas9000@gmail.com": "twozle",
 }
 
 

skills/creative/touchdesigner-mcp/SKILL.md

@@ -204,8 +204,9 @@ win.par.winopen.pulse()
 | `td_input_clear` | Stop input automation |
 | `td_op_screen_rect` | Get screen coords of a node |
 | `td_click_screen_point` | Click a point in a screenshot |
+| `td_screen_point_to_global` | Convert screenshot pixel to absolute screen coords |
 
-See `references/mcp-tools.md` for full parameter schemas.
+The table above covers the 32 tools used in typical creative workflows. The remaining 4 tools (`td_project_quit`, `td_test_session`, `td_dev_log`, `td_clear_dev_log`) are admin/dev-mode utilities — see `references/mcp-tools.md` for the full 36-tool reference with complete parameter schemas.
 
 ## Key Implementation Rules
 
@@ -338,6 +339,15 @@ See `references/network-patterns.md` for complete build scripts + shader code.
 | `references/operator-tips.md` | Wireframe rendering, feedback TOP setup |
 | `references/geometry-comp.md` | Geometry COMP: instancing, POP vs SOP, morphing |
 | `references/audio-reactive.md` | Audio band extraction, beat detection, envelope following |
+| `references/animation.md` | LFOs, timers, keyframes, easing, expression-driven motion |
+| `references/midi-osc.md` | MIDI/OSC controllers, TouchOSC, multi-machine sync |
+| `references/particles.md` | POPs and legacy particleSOP — emission, forces, collisions |
+| `references/projection-mapping.md` | Multi-window output, corner pin, mesh warp, edge blending |
+| `references/external-data.md` | HTTP, WebSocket, MQTT, Serial, TCP, webserverDAT |
+| `references/panel-ui.md` | Custom params, panel COMPs, button/slider/field, panelExecuteDAT |
+| `references/replicator.md` | replicatorCOMP — data-driven cloning, layouts, callbacks |
+| `references/dat-scripting.md` | Execute DAT family — chop/dat/parameter/panel/op/executeDAT |
+| `references/3d-scene.md` | Lighting rigs, shadows, IBL/cubemaps, multi-camera, PBR |
 | `scripts/setup.sh` | Automated setup script |
 
 ---

skills/creative/touchdesigner-mcp/references/3d-scene.md

@@ -0,0 +1,275 @@
+# 3D Scene Reference
+
+Lighting rigs, shadows, IBL/cubemaps, multi-camera, and PBR materials. For wireframe rendering and feedback TOPs see `operator-tips.md`. For instancing geometry see `geometry-comp.md`. For shader code see `glsl.md`.
+
+---
+
+## Anatomy of a 3D Scene
+
+```
+[Geometry COMP]    ← contains SOPs (the shapes)
+[Material]         ← Phong/PBR/GLSL/Constant MAT
+[Light COMPs]      ← point/directional/spot/area/environment
+[Camera COMP]      ← view position, FOV
+        │
+        ▼
+   [Render TOP]    ← combines geo + lights + camera into a 2D image
+        │
+        ▼
+   [post-FX chain] ← bloomTOP, glsl shaders, etc.
+        │
+        ▼
+   [windowCOMP]    ← actual display
+```
+
+Render TOP is the heart. It takes an explicit `geometry` path, an explicit `camera` path, and lights via the lights table or an envlight reference.
+
+---
+
+## Minimal Scene
+
+```python
+# Geometry
+geo = root.create(geometryCOMP, 'scene_geo')
+sphere = geo.create(sphereSOP, 'shape')
+sphere.par.rad = 1.0; sphere.par.rows = 64; sphere.par.cols = 64
+
+# Material — start with PBR
+mat = root.create(pbrMAT, 'mat')
+mat.par.basecolorr = 0.7; mat.par.basecolorg = 0.7; mat.par.basecolorb = 0.7
+mat.par.metallic = 0.0
+mat.par.roughness = 0.4
+
+geo.par.material = mat.path
+
+# Camera
+cam = root.create(cameraCOMP, 'cam1')
+cam.par.tx = 0; cam.par.ty = 0; cam.par.tz = 4
+cam.par.fov = 45
+cam.par.near = 0.1; cam.par.far = 100
+
+# Key light
+key = root.create(lightCOMP, 'key_light')
+key.par.lighttype = 'point'
+key.par.tx = 3; key.par.ty = 3; key.par.tz = 3
+key.par.dimmer = 1.5
+
+# Render
+render = root.create(renderTOP, 'render1')
+render.par.outputresolution = 'custom'
+render.par.resolutionw = 1920; render.par.resolutionh = 1080
+render.par.camera = cam.path
+render.par.geometry = geo.path
+render.par.lights = key.path                 # single light path; for multi, see below
+render.par.bgcolorr = 0; render.par.bgcolorg = 0; render.par.bgcolorb = 0
+```
+
+For multiple lights, leave `par.lights` blank — Render TOP scans the network for all `lightCOMP` and `envlightCOMP` ops by default. To restrict to specific lights, set `par.lights = '/project1/key_light /project1/fill_light'` (space-separated paths).
+
+---
+
+## Light Types
+
+| Type | What | Common params |
+|---|---|---|
+| `point` | Omnidirectional, falls off with distance | `dimmer`, `coneangle` (n/a), `attenuation` |
+| `directional` | Parallel rays, infinite distance (sun) | `dimmer`, light's rotation only matters |
+| `spot` | Cone, falls off with distance + angle | `coneangle`, `conedelta`, `dimmer` |
+| `cone` | Like spot but harder edge | same |
+| `area` | Rectangular soft light source | `sizex`, `sizey` |
+
+For all: `colorr`, `colorg`, `colorb`, `tx/ty/tz`, `rx/ry/rz`, `dimmer`.
+
+### Three-Point Lighting (Studio Setup)
+
+```python
+# Key — main light, ~45° front
+key = root.create(lightCOMP, 'key')
+key.par.lighttype = 'point'
+key.par.tx = 4; key.par.ty = 3; key.par.tz = 4
+key.par.dimmer = 1.5
+key.par.colorr = 1.0; key.par.colorg = 0.95; key.par.colorb = 0.85
+
+# Fill — softer, opposite side
+fill = root.create(lightCOMP, 'fill')
+fill.par.lighttype = 'area'
+fill.par.tx = -4; fill.par.ty = 2; fill.par.tz = 3
+fill.par.dimmer = 0.5
+fill.par.colorr = 0.7; fill.par.colorg = 0.8; fill.par.colorb = 1.0
+fill.par.sizex = 4; fill.par.sizey = 4
+
+# Rim/back — outline from behind
+rim = root.create(lightCOMP, 'rim')
+rim.par.lighttype = 'spot'
+rim.par.tx = 0; rim.par.ty = 4; rim.par.tz = -4
+rim.par.coneangle = 30
+rim.par.dimmer = 1.0
+
+# Optional: ambient lift to prevent pure-black shadows
+amb = root.create(ambientlightCOMP, 'ambient')
+amb.par.dimmer = 0.15
+```
+
+---
+
+## Shadows
+
+Spot and directional lights cast shadows when `par.shadowtype != 'none'`.
+
+```python
+key.par.shadowtype = 'softshadow'        # 'none' | 'hardshadow' | 'softshadow'
+key.par.shadowsize = 1024                # shadow map resolution
+key.par.shadowsoftness = 0.02            # softshadow only
+```
+
+**Tips:**
+- Soft shadows are GPU-expensive. Start with `shadowsize = 1024` and only go higher (2048/4096) if shadow edges look pixelated at your resolution.
+- Set the spot light's `near`/`far` to JUST contain the scene. Wider range = wasted shadow map precision.
+- Multiple shadow-casting lights compound cost. Limit to 1-2 in real-time work; pre-bake the rest into the materials.
+
+---
+
+## Image-Based Lighting (IBL) / Environment Light
+
+For realistic PBR materials you need a cubemap for reflections.
+
+```python
+# Environment light from an HDR
+env = root.create(envlightCOMP, 'env')
+env.par.envmap = '/project1/cube_in'         # path to a TOP that produces a cubemap
+env.par.envlightmap = ...                    # diffuse irradiance map (often same as envmap)
+env.par.dimmer = 1.0
+
+# Cubemap source — option A: built-in cubeTOP from 6 faces
+cube = root.create(cubeTOP, 'cube_in')
+# (assign 6 face TOPs)
+
+# Option B: HDR equirectangular → cubemap conversion
+# Use a moviefileinTOP loading .hdr or .exr, then projectTOP type='cubemapfromequirect'
+hdr = root.create(moviefileinTOP, 'hdr_src')
+hdr.par.file = '/path/to/environment.hdr'
+
+proj = root.create(projectTOP, 'cube_proj')
+proj.par.projecttype = 'cubemapfromequirect'
+proj.inputConnectors[0].connect(hdr)
+```
+
+PBR materials sample the environment automatically when `envlightCOMP` is in the scene. Verify param names with `td_get_par_info(op_type='envlightCOMP')` — TD versions vary.
+
+---
+
+## PBR Material Setup
+
+```python
+mat = root.create(pbrMAT, 'pbr_metal')
+mat.par.basecolorr = 0.95; mat.par.basecolorg = 0.65; mat.par.basecolorb = 0.4
+mat.par.metallic = 1.0
+mat.par.roughness = 0.25
+mat.par.specularlevel = 0.5
+mat.par.emitcolorr = 0; mat.par.emitcolorg = 0; mat.par.emitcolorb = 0
+
+# Texture maps
+mat.par.basecolormap = '/project1/textures/albedo'         # TOP path
+mat.par.metallicroughnessmap = '/project1/textures/mr'      # G=roughness, B=metallic (glTF convention)
+mat.par.normalmap = '/project1/textures/normal'
+mat.par.emitmap = '/project1/textures/emit'
+mat.par.occlusionmap = '/project1/textures/ao'
+```
+
+**Material idioms:**
+
+| Look | metallic | roughness | basecolor |
+|---|---|---|---|
+| Brushed steel | 1.0 | 0.4 | (0.7, 0.7, 0.7) |
+| Polished gold | 1.0 | 0.1 | (1.0, 0.85, 0.4) |
+| Plastic | 0.0 | 0.5 | mid-saturated |
+| Rubber | 0.0 | 0.9 | dark |
+| Glass | 0.0 | 0.05 | (1, 1, 1), low alpha + transmission |
+| Glowing emitter | 0.0 | 1.0 | dark, high `emitcolor` |
+
+For glass/transmission, recent TD versions support `transmission` in PBR; older versions need glslMAT.
+
+---
+
+## Multi-Camera Setups
+
+For comparison views, instant replay, multi-screen mapping, etc.
+
+```python
+# Camera A — main scene
+cam_a = root.create(cameraCOMP, 'cam_main')
+cam_a.par.tz = 5
+
+# Camera B — orbiting top-down
+cam_b = root.create(cameraCOMP, 'cam_top')
+cam_b.par.ty = 6; cam_b.par.rx = -90
+
+# Render each via separate Render TOPs
+render_a = root.create(renderTOP, 'render_main')
+render_a.par.camera = cam_a.path
+render_a.par.geometry = geo.path
+
+render_b = root.create(renderTOP, 'render_top')
+render_b.par.camera = cam_b.path
+render_b.par.geometry = geo.path
+```
+
+Composite both with a `multiplyTOP`/`compositeTOP` for picture-in-picture, or route to separate `windowCOMP`s for multi-display.
+
+### Camera animation
+
+Drive camera params via expressions (orbit), animationCOMP (waypoint), or LFO (oscillation):
+
+```python
+# Orbiting camera
+cam_a.par.tx.mode = ParMode.EXPRESSION
+cam_a.par.tx.expr = "cos(absTime.seconds * 0.3) * 6"
+cam_a.par.tz.mode = ParMode.EXPRESSION
+cam_a.par.tz.expr = "sin(absTime.seconds * 0.3) * 6"
+cam_a.par.lookat = '/project1/scene_geo'        # auto-aim at target
+```
+
+`par.lookat` is the simplest "always look at target" mechanism.
+
+### Depth of field
+
+PBR + Render TOP supports DOF when `par.dof = 'on'`.
+
+```python
+render.par.dof = 'on'
+render.par.focusdistance = 5.0
+render.par.aperture = 0.05         # blur strength
+render.par.bokehshape = 'hexagon'
+```
+
+DOF is GPU-heavy. Render at lower res then upscale for performance.
+
+---
+
+## Common Pitfalls
+
+1. **Render TOP shows black** — most common cause: no light. Even with PBR you need at least one `lightCOMP` or `envlightCOMP`. Add an `ambientlightCOMP` at low dimmer as a safety net.
+2. **Material doesn't appear** — `geo.par.material` must be a string PATH, not the material op itself. Use `mat.path`, not `mat`.
+3. **Lights ignored** — by default Render TOP picks up ALL `lightCOMP`s in the network. If you have leftover lights from another scene, they leak in. Set `par.lights` explicitly.
+4. **PBR looks flat** — without an `envlightCOMP` providing reflections, PBR materials look like Phong. Add one even if you don't have an HDR (use a `constantTOP` cubemap as fallback).
+5. **Shadow acne / striping** — increase `par.shadowbias` slightly. Tune per-light.
+6. **Camera inside geometry** — if `cam.par.tz` is INSIDE a sphere, you see the inside (or nothing if backface culled). Move the camera further out.
+7. **Light range too small** — point lights have implicit attenuation. Far-away geometry receives little light. Increase `par.dimmer` or move lights closer.
+8. **Multiple cameras conflict** — one render TOP = one camera. Don't try to share. Use multiple render TOPs.
+9. **Wrong handedness** — TD is right-handed Y-up. Imported assets from Z-up apps (Blender, Maya in Z-up) need a 90° X rotation on the geo COMP.
+10. **Cooking budget** — PBR + IBL + shadows + DOF at 1080p60 is fine on modern GPUs but 4K + 4 lights + soft shadows + DOF will tank. Profile via `td_get_perf` and downgrade settings before adding more.
+
+---
+
+## Quick Recipes
+
+| Goal | Recipe |
+|---|---|
+| Studio portrait | 3-point rig (key + fill + rim) + ambient + PBR mat + DOF |
+| Outdoor daylight | One directional `lightCOMP` (sun) + envlight (sky HDR) + soft shadows |
+| Dramatic / film noir | Single spot light from upper side, hard shadows, deep ambient = 0.05 |
+| Abstract / dreamy | Multiple area lights at low dimmer, no shadows, `bloomTOP` post |
+| Product render | Three-point + IBL + neutral PBR + `bgcolorr=g=b=1` (white seamless) |
+| Game-style | Phong MAT + 1-2 lights + no IBL + flat ambient (cheap, stylized) |
+| Wireframe + solid | Two render TOPs (one with wireframeMAT, one with PBR), composite via `addTOP` |
+| Orbiting camera | `par.lookat` + expressions on tx/tz using sin/cos |

skills/creative/touchdesigner-mcp/references/animation.md

@@ -0,0 +1,221 @@
+# Animation Reference
+
+Patterns for time-based motion — keyframes, LFOs, timers, easing, expression-driven animation.
+
+Always call `td_get_par_info` for the op type before setting params. Param names below reflect TD 2025.32 but verify if errors fire.
+
+---
+
+## Time Sources
+
+TD has three time references — pick the right one.
+
+| Expression | Behavior | Use for |
+|---|---|---|
+| `absTime.seconds` | Wall-clock seconds since TD started. Never resets. | Continuous motion, GLSL `uTime`, infinite loops |
+| `absTime.frame` | Wall-clock frame count. | Frame-accurate triggers |
+| `me.time.frame` | Local component frame count (resets on play/stop). | Per-COMP animation timeline |
+| `me.time.seconds` | Local component seconds. | Same, in seconds |
+
+**Rule:** for shaders and continuous motion use `absTime.seconds`. For triggered/looping animations inside a COMP use `me.time.*`.
+
+---
+
+## LFO CHOP — Cyclic Motion
+
+The simplest periodic driver. Fast, GPU-cheap, expression-friendly.
+
+```python
+lfo = root.create(lfoCHOP, 'rot_driver')
+lfo.par.type = 'sin'        # 'sin' | 'cos' | 'ramp' | 'square' | 'triangle' | 'pulse'
+lfo.par.frequency = 0.25    # cycles per second
+lfo.par.amplitude = 1.0
+lfo.par.offset = 0.0
+lfo.par.phase = 0.0         # 0-1, useful for offsetting parallel LFOs
+```
+
+**Drive a parameter via export:**
+
+```python
+op('/project1/geo1').par.rx.mode = ParMode.EXPRESSION
+op('/project1/geo1').par.rx.expr = "op('rot_driver')['chan1'] * 360"
+```
+
+**Multiple synced LFOs (X/Y/Z rotation with phase offsets):**
+Create one LFO with three channels and phase-offset each, or use three LFOs and offset their `phase` params (0.0, 0.33, 0.66).
+
+---
+
+## Timer CHOP — Triggered Sequences
+
+For run-once animations, beat-locked sequences, or stage-based logic.
+
+```python
+timer = root.create(timerCHOP, 'fade_timer')
+timer.par.length = 4.0       # cycle length in seconds
+timer.par.cycle = False      # run once vs. loop
+timer.par.outputseconds = True
+```
+
+Output channels: `timer_fraction` (0→1 across the cycle), `running`, `done`, `cycles`.
+
+**Start the timer:**
+```python
+timer.par.start.pulse()
+```
+
+**Drive a fade:**
+```python
+op('/project1/level1').par.opacity.mode = ParMode.EXPRESSION
+op('/project1/level1').par.opacity.expr = "op('fade_timer')['timer_fraction']"
+```
+
+**Easing on the timer fraction** — apply in the expression itself:
+
+```python
+# Smoothstep: ease in/out
+expr = "smoothstep(0, 1, op('fade_timer')['timer_fraction'])"
+# Cubic ease-out: 1 - (1-t)^3
+expr = "1 - pow(1 - op('fade_timer')['timer_fraction'], 3)"
+```
+
+---
+
+## Pattern CHOP — Custom Curves
+
+For arbitrary waveforms (saw ramps, easing curves, custom envelopes).
+
+```python
+pat = root.create(patternCHOP, 'envelope')
+pat.par.type = 'gaussian'    # 'gaussian' | 'ramp' | 'square' | 'sin' | etc.
+pat.par.length = 60          # samples
+pat.par.cyclelength = 1.0    # seconds at TD framerate
+```
+
+Combine with `lookupCHOP` to remap a 0-1 driver through a custom curve.
+
+---
+
+## Animation COMP — Keyframe-Based
+
+For multi-keyframe motion graphics. Each animationCOMP holds channels with keyframes editable in the Animation Editor.
+
+```python
+anim = root.create(animationCOMP, 'intro_anim')
+# By default has channels chan1..chanN; access via:
+# op('intro_anim').par.length, .par.play, .par.cue, etc.
+
+# Drive a parameter from a channel
+op('/project1/text1').par.tx.mode = ParMode.EXPRESSION
+op('/project1/text1').par.tx.expr = "op('intro_anim/out1')['chan1']"
+```
+
+**Keyframes are typically edited in the UI** (Animation Editor), but can be set via `keyframes` table internally. For programmatic keyframe creation, use `td_execute_python`:
+
+```python
+# Get the channel CHOP inside an animationCOMP
+ch = op('/project1/intro_anim/chans')
+# Insert a key (advanced API — verify with td_get_par_info(op_type='animationCOMP'))
+ch.appendKey('chan1', frame=0, value=0.0, expression=None)
+ch.appendKey('chan1', frame=120, value=1.0)
+```
+
+For most use cases, drive params with LFO/Timer/Pattern CHOPs instead — simpler and scriptable.
+
+---
+
+## Easing in Expressions
+
+TD's expression evaluator supports Python math. Common easing forms:
+
+```python
+# Linear
+"t"
+
+# Smoothstep (classic ease-in-out)
+"smoothstep(0, 1, t)"
+
+# Ease-out cubic
+"1 - pow(1 - t, 3)"
+
+# Ease-in cubic
+"pow(t, 3)"
+
+# Ease-in-out cubic
+"3*t*t - 2*t*t*t"
+
+# Bounce (manual, simplified)
+"abs(sin(t * 6.28 * 3) * (1 - t))"
+```
+
+Where `t` is `op('fade_timer')['timer_fraction']` or any 0-1 driver.
+
+---
+
+## Filter CHOP — Smoothing Existing Channels
+
+Smooth out jittery values (e.g., audio analysis, sensor data) before driving visuals.
+
+```python
+filt = root.create(filterCHOP, 'smooth')
+filt.par.filter = 'gaussian'   # or 'lowpass'
+filt.par.width = 0.5            # smoothing window in seconds
+filt.inputConnectors[0].connect(op('raw_signal'))
+```
+
+**WARNING:** Do NOT use Filter CHOP on AudioSpectrum output in timeslice mode — it expands the sample count and averages bins to near-zero. See `audio-reactive.md`.
+
+---
+
+## Lag CHOP — Asymmetric Attack/Release
+
+Different speeds for rising vs. falling values. Standard for visualizing audio envelopes.
+
+```python
+lag = root.create(lagCHOP, 'env_smooth')
+lag.par.lag1 = 0.02   # attack (rise time, seconds)
+lag.par.lag2 = 0.30   # release (fall time, seconds)
+lag.inputConnectors[0].connect(op('raw_envelope'))
+```
+
+Fast attack, slow release = classic VU-meter feel.
+
+---
+
+## Per-Frame Driving via Script DAT
+
+For complex per-frame logic that doesn't fit expressions, use a `executeDAT` (`onFrameStart` callback) or a `chopExecuteDAT`.
+
+```python
+# In an executeDAT (frameStart):
+def onFrameStart(frame):
+    t = absTime.seconds
+    op('/project1/circle').par.tx = math.sin(t * 2.0) * 3.0
+    op('/project1/circle').par.ty = math.cos(t * 2.0) * 3.0
+    return
+```
+
+Heavy logic should still be in CHOPs (CPU-cheap, deterministic). Reserve scripts for one-shots or non-realtime branching.
+
+---
+
+## Pitfalls
+
+1. **Frame rate dependency** — `me.time.frame` is in TD project frames (default 60). If your project rate changes, motion speed changes. Use `seconds` for rate-independent timing.
+2. **Cooking budget** — every CHOP that drives a parameter cooks every frame. Consolidate drivers (one big mathCHOP > many small ones).
+3. **Expression mode** — params default to `CONSTANT`. `par.X.expr = ...` is ignored unless `par.X.mode = ParMode.EXPRESSION`.
+4. **Animation editor edits** — keyframes set via UI live in the animationCOMP's internal keyframe table. They survive save/reopen. Programmatic keys via `appendKey()` work but verify the API with `td_get_docs(topic='animation')` first.
+5. **Looping animations** — for seamless loops, `length` must equal `cyclelength` and the start/end values must match. Otherwise expect a visible jump.
+
+---
+
+## Quick Recipes
+
+| Goal | Simplest path |
+|---|---|
+| Continuous rotation | LFO CHOP `type='ramp'`, expr → `geo.par.rx` |
+| Fade in over 2s | Timer CHOP `length=2`, smoothstep expr → `level.par.opacity` |
+| Pulse on every beat | `triggerCHOP` from audio → drive scale via expression |
+| 3D Lissajous orbit | Two LFOs with different freq, drive `tx`/`ty`/`tz` |
+| Random jitter | `noiseCHOP` (low-freq) added to position |
+| Timed scene switch | Timer CHOP → switchTOP/CHOP `index` |

skills/creative/touchdesigner-mcp/references/dat-scripting.md

@@ -0,0 +1,352 @@
+# DAT-Based Scripting Reference
+
+TD's event/callback model — Python that runs in response to network events. The full set of "Execute DATs" plus their idiomatic patterns.
+
+For arbitrary Python execution (not callback-based), see `python-api.md`. For the MCP's `td_execute_python` tool, see `mcp-tools.md`.
+
+---
+
+## The Execute DAT Family
+
+Every type watches one kind of event source and fires Python on changes.
+
+| DAT | Watches | Use for |
+|---|---|---|
+| `chopExecuteDAT` | A CHOP's channel values | Audio triggers, threshold callbacks, state machines on numeric input |
+| `datExecuteDAT` | A DAT's content (table cells, text) | Reacting to data updates from APIs, parsing webDAT responses |
+| `parameterExecuteDAT` | A parameter's value or pulse | Reacting to user-changed params, custom pulse buttons |
+| `panelExecuteDAT` | A panel COMP's interaction | Button clicks, slider drags, field commits |
+| `opExecuteDAT` | Operator lifecycle | New operator created, deleted, name changed |
+| `executeDAT` | Project lifecycle, frame events | Run-once setup, per-frame logic, save/load hooks |
+
+All have a docked DAT with predefined callback functions. You only fill in the bodies of the ones you care about.
+
+---
+
+## chopExecuteDAT — Numeric Triggers
+
+```python
+ce = root.create(chopExecuteDAT, 'kick_handler')
+ce.par.chop = '/project1/audio/out_kick'      # source CHOP
+ce.par.offtoon = True                          # fire when channel rises above 0
+ce.par.ontooff = False
+ce.par.whileon = False
+ce.par.valuechange = False
+```
+
+In the docked callback DAT:
+
+```python
+def offToOn(channel, sampleIndex, val, prev):
+    """Channel went from 0 to non-zero. Classic beat trigger."""
+    op('/project1/strobe').par.flash.pulse()
+    op('/project1/scene').par.index = (op('/project1/scene').par.index + 1) % 8
+    return
+
+def onToOff(channel, sampleIndex, val, prev):
+    """Channel went from non-zero to 0."""
+    return
+
+def whileOn(channel, sampleIndex, val, prev):
+    """Fires every frame while channel is non-zero. Use sparingly."""
+    return
+
+def valueChange(channel, sampleIndex, val, prev):
+    """Fires every frame the value changes (continuous). Heavy."""
+    return
+```
+
+`channel` is a `Channel` object — `.name`, `.owner`, `.vals[]`. Use `channel.name == 'chan1'` to filter.
+
+**Threshold-based custom triggers:** wire the source CHOP through a `triggerCHOP` first to get clean 0/1 pulses, then watch with `offtoon`.
+
+---
+
+## datExecuteDAT — Table/Text Changes
+
+```python
+de = root.create(datExecuteDAT, 'api_response')
+de.par.dat = '/project1/api/web1'              # source DAT
+de.par.tablechange = True                      # any cell change
+de.par.cellchange = False
+de.par.rowchange = False
+de.par.colchange = False
+```
+
+```python
+def onTableChange(dat):
+    """Whole table changed (including text DAT content updates)."""
+    if dat.numRows == 0:
+        return
+    # If it's a webDAT response, parse JSON
+    import json
+    try:
+        data = json.loads(dat.text)
+    except json.JSONDecodeError:
+        debug(f'Bad JSON: {dat.text[:100]}')
+        return
+    # Write to a CHOP
+    op('/project1/api_value').par.value0 = float(data.get('count', 0))
+    return
+
+def onCellChange(dat, cells, prev):
+    """Specific cells changed."""
+    for cell in cells:
+        # cell.row, cell.col, cell.val
+        pass
+    return
+```
+
+`debug()` prints to the textport — readable via `td_read_textport`.
+
+---
+
+## parameterExecuteDAT — Param Changes & Pulse
+
+```python
+pe = root.create(parameterExecuteDAT, 'comp_params')
+pe.par.op = '/project1/my_component'           # COMP whose params to watch
+pe.par.parameters = '*'                         # or specific names like 'Intensity Reset'
+pe.par.valuechange = True
+pe.par.pulse = True
+```
+
+```python
+def onValueChange(par, prev):
+    """par is a Par object. par.name, par.eval(), par.owner."""
+    if par.name == 'Intensity':
+        op('/project1/bloom').par.threshold = par.eval()
+    return
+
+def onPulse(par):
+    """Pulse param was triggered."""
+    if par.name == 'Reset':
+        op('/project1/scene').par.index = 0
+        op('/project1/audio_player').par.cuepoint = 0
+        op('/project1/audio_player').par.cuepulse.pulse()
+    return
+
+def onExpressionChange(par, val, prev):
+    """User changed the expression on a param."""
+    return
+
+def onExportChange(par, val, prev):
+    """Export source changed."""
+    return
+
+def onModeChange(par, val, prev):
+    """Param mode changed (CONSTANT / EXPRESSION / EXPORT / etc)."""
+    return
+```
+
+---
+
+## panelExecuteDAT — UI Events
+
+For interactive control surfaces. See `panel-ui.md` for the full panel COMP context.
+
+```python
+pe = root.create(panelExecuteDAT, 'btn_handler')
+pe.par.panel = '/project1/play_btn'
+pe.par.click = True              # mouse click events
+pe.par.value = True              # state changes (toggle)
+pe.par.lockedchange = False
+```
+
+```python
+def onOffToOn(panelValue):
+    """Panel value rose to 1 (button pressed, slider crossed threshold)."""
+    op('/project1/scene_timer').par.start.pulse()
+    return
+
+def onOnToOff(panelValue):
+    """Panel value dropped to 0."""
+    return
+
+def onValueChange(panelValue):
+    """Continuous: every frame the value changes."""
+    val = panelValue.eval()
+    op('/project1/master').par.opacity = val
+    return
+
+def onClick(panelValue):
+    """Discrete click event, fires once per click."""
+    return
+```
+
+`panelValue` is a `Par` object on the panel COMP.
+
+---
+
+## opExecuteDAT — Operator Lifecycle
+
+Watches creation/deletion/renaming of operators in a parent COMP.
+
+```python
+oe = root.create(opExecuteDAT, 'lifecycle')
+oe.par.op = '/project1'
+oe.par.create = True
+oe.par.destroy = True
+oe.par.namechange = True
+oe.par.flagchange = False
+```
+
+```python
+def onCreate(opCreated):
+    """A new operator was created. Useful for auto-applying conventions."""
+    if opCreated.OPType == 'glslTOP':
+        # Always wrap with a null
+        n = opCreated.parent().create(nullTOP, opCreated.name + '_out')
+        n.inputConnectors[0].connect(opCreated)
+    return
+
+def onDestroy(opDestroyed):
+    """Operator was deleted. opDestroyed.path is still valid for one frame."""
+    return
+
+def onNameChange(opChanged):
+    """Operator was renamed."""
+    return
+```
+
+Useful for dev-time scaffolding (auto-create downstream nullTOPs, auto-name conventions). Disable in production projects to avoid surprise side effects.
+
+---
+
+## executeDAT — Project Lifecycle & Per-Frame
+
+The catch-all. Gets you hooks into project start, save, load, frame-start, frame-end.
+
+```python
+exec_dat = root.create(executeDAT, 'lifecycle')
+exec_dat.par.start = True
+exec_dat.par.create = True
+exec_dat.par.framestart = True
+exec_dat.par.frameend = False
+```
+
+```python
+def onStart():
+    """Project just started cooking. Run once."""
+    op('/project1/scene').par.index = 0
+    debug('Project started')
+    return
+
+def onCreate():
+    """Component was just created (only fires for component executeDATs, not project root)."""
+    return
+
+def onFrameStart(frame):
+    """Per-frame, BEFORE network cooks. Heavy logic here = bottleneck."""
+    return
+
+def onFrameEnd(frame):
+    """Per-frame, AFTER network cooks. Use for capture, recording, post-network logic."""
+    return
+
+def onPlayStateChange(playing):
+    """Project play/pause toggled."""
+    return
+
+def onProjectPreSave():
+    """Right before saving the .toe file."""
+    return
+
+def onProjectPostSave():
+    return
+```
+
+Heavy per-frame logic in `onFrameStart` is one of the top performance regressions in TD projects. Use CHOPs for per-frame computation, scripts for events.
+
+---
+
+## Pattern: Triggering an Animation Sequence on Beat
+
+```python
+# Source: a kick trigger CHOP
+# Goal: on each kick, run a 1.5s scale pulse + color flash
+
+# Setup (create once)
+animator = root.create(timerCHOP, 'pulse_anim')
+animator.par.length = 1.5
+animator.par.cycle = False
+
+# Param expressions on visual targets:
+op('logo').par.sx.expr = "1.0 + (1 - op('pulse_anim')['timer_fraction']) * 0.3"
+op('logo').par.sx.mode = ParMode.EXPRESSION
+op('logo').par.sy.expr = "1.0 + (1 - op('pulse_anim')['timer_fraction']) * 0.3"
+op('logo').par.sy.mode = ParMode.EXPRESSION
+
+# In a chopExecuteDAT watching the kick CHOP:
+def offToOn(channel, sampleIndex, val, prev):
+    op('pulse_anim').par.start.pulse()
+    return
+```
+
+---
+
+## Pattern: Live Editing a CHOP from API Data
+
+```python
+# webDAT polls an API every 5 seconds
+# datExecuteDAT parses the response and writes to a constantCHOP
+
+def onTableChange(dat):
+    import json
+    try:
+        data = json.loads(dat.text)
+    except:
+        return
+    target = op('/project1/external_state')
+    target.par.name0 = 'temperature'
+    target.par.value0 = float(data['temp_c'])
+    target.par.name1 = 'humidity'
+    target.par.value1 = float(data['humidity'])
+    return
+```
+
+Visuals just reference `op('external_state')['temperature']` — they update live.
+
+---
+
+## Pattern: Self-Cleaning Network
+
+```python
+# An opExecuteDAT watching for orphaned helper ops, deleting them after their parent disappears
+
+def onDestroy(opDestroyed):
+    parent_name = opDestroyed.name
+    helper = op(f'/project1/{parent_name}_helper')
+    if helper:
+        helper.destroy()
+    return
+```
+
+---
+
+## Pitfalls
+
+1. **Callbacks crash silently** — exceptions print to the textport but don't show up in the UI. Always `td_clear_textport` before debugging, then `td_read_textport` after.
+2. **`debug()` vs `print()`** — both write to textport, but `debug()` includes the file/line of the calling DAT. Prefer `debug()` for scripts.
+3. **`val` is the new value, `prev` is old** — easy to swap. Always: `def offToOn(channel, sampleIndex, val, prev)`. Check parameter order in TD docs if confused.
+4. **`whileOn` and `valueChange` are per-frame** — heavy. Avoid unless absolutely needed. Drive via expressions instead.
+5. **Callbacks don't run during cooking-paused state** — if the parent COMP has `allowCooking=False`, callbacks freeze. Useful for "disable me" toggles.
+6. **`par` vs `panelValue`** — parameterExecuteDAT gives `par` (a Par object), panelExecuteDAT gives `panelValue` (also a Par-like object). Both have `.name` and `.eval()` but their context differs.
+7. **`opExecuteDAT` fires for itself** — when you create an opExecuteDAT, it can fire `onCreate` for itself if `par.create=True` and parent matches. Filter by `if opCreated == me: return`.
+8. **Reload behavior** — when reloading an extension (`td_reinit_extension`), all callback DATs reset their internal state. Module-level vars are lost. Persist state in tableDATs or the docked DAT itself, not in module globals.
+9. **Cooking dependencies** — if a callback writes to an op that's upstream of the callback's source, you get a cooking loop. TD warns about it but doesn't always block. Keep dataflow one-directional.
+10. **Active flag** — every Execute DAT has `par.active`. False = silent. Easy to toggle for testing without deleting wiring.
+
+---
+
+## Quick Recipes
+
+| Goal | Setup |
+|---|---|
+| Beat trigger | `chopExecuteDAT.par.offtoon=True` watching a `triggerCHOP` |
+| API response handler | `datExecuteDAT.par.tablechange=True` watching a `webDAT` |
+| Custom button → action | `parameterExecuteDAT.par.pulse=True` watching a custom pulse param |
+| Slider → continuous param | `panelExecuteDAT.par.value=True` watching a `sliderCOMP` |
+| Run-once setup | `executeDAT.par.start=True` with logic in `onStart()` |
+| Per-frame metrics | `executeDAT.par.frameend=True` recording values to a CHOP |
+| Auto-name new ops | `opExecuteDAT.par.create=True` enforcing naming conventions |

skills/creative/touchdesigner-mcp/references/external-data.md

@@ -0,0 +1,322 @@
+# External Data Reference
+
+Network and device I/O — HTTP requests, WebSockets, MQTT, Serial, TCP, UDP. For MIDI/OSC specifically see `midi-osc.md`.
+
+Common production needs:
+- API polling / webhook ingestion
+- Real-time data streams (sensors, market data, chat)
+- IoT device control (Arduino, ESP32, smart lights)
+- Inter-application messaging
+- Hosting a tiny TD-side HTTP server for remote control
+
+---
+
+## Web DAT — HTTP Requests
+
+```python
+web = root.create(webDAT, 'api_call')
+web.par.url = 'https://api.example.com/v1/status'
+web.par.fetchmethod = 'get'           # 'get' | 'post' | 'put' | 'delete'
+web.par.format = 'auto'                # 'auto' | 'text' | 'json'
+web.par.timeout = 5.0
+```
+
+**Triggering a request:**
+
+`webDAT` does NOT auto-fetch on cook. Trigger explicitly:
+
+```python
+web.par.fetch.pulse()
+```
+
+Or via expression on a CHOP value-change (chopExecuteDAT — see `dat-scripting.md`).
+
+**Authentication headers:**
+
+Use `webclientDAT` (more flexible) or set `webDAT` headers via the headers DAT:
+
+```python
+web_headers = root.create(tableDAT, 'headers')
+web_headers.appendRow(['Authorization', 'Bearer YOUR_TOKEN'])
+web_headers.appendRow(['Accept', 'application/json'])
+web.par.headers = web_headers.path
+```
+
+**Parsing JSON response:**
+
+```python
+import json
+
+def onTableChange(dat):
+    response = dat.text          # raw response body
+    data = json.loads(response)
+    # Update a tableDAT or store in a constantCHOP for downstream use
+    op('/project1/api_status').par.value0 = data['count']
+    return
+```
+
+Wire this in a `datExecuteDAT` watching the webDAT.
+
+**Polling pattern:**
+
+```python
+# timerCHOP fires every N seconds
+timer = root.create(timerCHOP, 'poll_timer')
+timer.par.length = 5.0
+timer.par.cycle = True
+
+# chopExecuteDAT on the timer's 'cycles' channel pulses the webDAT
+def offToOn(channel, sampleIndex, val, prev):
+    op('/project1/api_call').par.fetch.pulse()
+    return
+```
+
+---
+
+## Web Client DAT — More Robust HTTP
+
+`webclientDAT` is the modern replacement for `webDAT` — supports streaming responses, chunked transfer, custom auth.
+
+```python
+client = root.create(webclientDAT, 'api')
+client.par.method = 'POST'
+client.par.url = 'https://api.example.com/events'
+client.par.uploadtype = 'json'
+client.par.uploaddata = '{"event": "scene_change", "scene": 3}'
+client.par.request.pulse()
+```
+
+Output goes to its child `webclient1_response` DAT. Use a `datExecuteDAT` to react.
+
+---
+
+## Web Server DAT — TD as HTTP Server
+
+Hosts a tiny HTTP server inside TD. Useful for:
+- Status/health endpoints
+- Remote control from a phone or another machine
+- Webhook receivers from external services
+
+```python
+server = root.create(webserverDAT, 'control_server')
+server.par.port = 8080
+server.par.active = True
+
+# Define handler in the docked callback DAT
+```
+
+In the auto-created `webserver1_callbacks` DAT:
+
+```python
+def onHTTPRequest(webServerDAT, request, response):
+    path = request['uri']
+    if path == '/status':
+        response['statusCode'] = 200
+        response['data'] = '{"fps": 60, "scene": "active"}'
+    elif path == '/scene':
+        idx = int(request['args'].get('index', 0))
+        op('/project1/scene_switch').par.index = idx
+        response['statusCode'] = 200
+        response['data'] = 'OK'
+    else:
+        response['statusCode'] = 404
+        response['data'] = 'Not Found'
+    return response
+```
+
+Test from terminal: `curl http://localhost:8080/status`.
+
+**Security:** No auth by default. Bind to localhost only or add a token check in the callback. Never expose to the public internet without auth.
+
+---
+
+## WebSocket DAT — Bidirectional Real-Time
+
+For low-latency bidirectional streams (chat, live data feeds, controllers).
+
+### Client
+
+```python
+ws = root.create(websocketDAT, 'ws_client')
+ws.par.netaddress = 'wss://api.example.com/socket'
+ws.par.active = True
+```
+
+In the docked callbacks DAT:
+
+```python
+def onConnect(dat):
+    dat.sendText('{"action": "subscribe", "channel": "ticks"}')
+    return
+
+def onReceiveText(dat, rowIndex, message):
+    # message is a string; parse JSON, dispatch to ops
+    import json
+    data = json.loads(message)
+    op('/project1/price_chop').par.value0 = data['price']
+    return
+
+def onDisconnect(dat):
+    # Optionally schedule a reconnect
+    return
+```
+
+### Server
+
+```python
+ws = root.create(websocketDAT, 'ws_server')
+ws.par.mode = 'server'
+ws.par.port = 9001
+ws.par.active = True
+```
+
+Same callback structure with an additional `clientID` arg.
+
+---
+
+## MQTT — Pub/Sub for IoT
+
+```python
+mqtt = root.create(mqttClientDAT, 'iot')
+mqtt.par.brokeraddress = 'broker.hivemq.com'
+mqtt.par.brokerport = 1883
+mqtt.par.clientid = 'td_install_01'
+mqtt.par.connect.pulse()
+
+# Subscribe in callbacks DAT:
+def onConnect(dat):
+    dat.subscribe('home/lights/+', qos=1)
+    return
+
+def onReceive(dat, topic, payload, qos, retained, dup):
+    # payload is bytes — decode if JSON
+    msg = payload.decode('utf-8')
+    # Dispatch by topic
+    return
+
+# Publish from anywhere:
+op('iot').publish('show/scene', 'sunset', qos=0, retain=False)
+```
+
+For Mosquitto / HiveMQ self-hosted brokers use the same setup with `tcp://192.168.x.x` and your local port.
+
+---
+
+## Serial DAT — Arduino, USB Devices
+
+```python
+serial = root.create(serialDAT, 'arduino')
+serial.par.port = '/dev/cu.usbmodem14101'   # macOS — check Arduino IDE
+# Windows: 'COM3', 'COM4', etc.
+serial.par.baudrate = 115200
+serial.par.active = True
+```
+
+In callbacks:
+
+```python
+def onReceive(dat, rowIndex, line):
+    # Each newline-terminated line from Arduino arrives here
+    parts = line.split(',')
+    op('/project1/sensors').par.value0 = float(parts[0])
+    op('/project1/sensors').par.value1 = float(parts[1])
+    return
+```
+
+Send to Arduino:
+```python
+op('arduino').send('LED_ON\n')
+```
+
+---
+
+## TCP/IP DAT — Custom Protocols
+
+For talking to non-HTTP servers (game servers, custom protocols, legacy systems).
+
+```python
+tcp = root.create(tcpipDAT, 'show_control')
+tcp.par.netaddress = '192.168.1.50'
+tcp.par.port = 7000
+tcp.par.protocol = 'tcp'        # 'tcp' | 'udp'
+tcp.par.active = True
+```
+
+Send / receive via callbacks similar to websocketDAT.
+
+For UDP-only (fire-and-forget, no connection), use `udpoutDAT` + `udpinDAT` — simpler but unreliable across networks.
+
+---
+
+## Common Patterns
+
+### REST API → Visual
+
+```
+timerCHOP (5s loop)
+   → chopExecuteDAT (pulse webDAT.par.fetch on cycle)
+   → webDAT (returns JSON)
+   → datExecuteDAT (parse, write to constantCHOP)
+   → CHOP drives glsl uniform → visuals
+```
+
+### Webhook receiver
+
+```
+webserverDAT (port 8080, /webhook endpoint)
+   → callback writes to a tableDAT log + triggers a scene change
+```
+
+### Real-time stock/crypto ticker
+
+```
+websocketDAT (subscribe to feed)
+   → onReceiveText callback parses JSON
+   → writes to constantCHOP
+   → drives bar chart / typography animation
+```
+
+### IoT-controlled installation
+
+```
+MQTT → callback dispatches by topic
+   → /lights/main → constantCHOP drives lighting render
+   → /audio/volume → mathCHOP for master fader
+```
+
+### Two-way phone control
+
+```
+WebSocket server in TD
+   → simple HTML page on phone connects, sends slider values
+   → callback writes to ops
+   → TD pushes status back via dat.sendText() to phone UI
+```
+
+---
+
+## Pitfalls
+
+1. **`webDAT` doesn't auto-fetch** — must explicitly pulse `par.fetch`. Easy to forget.
+2. **Blocking on slow APIs** — `webDAT` runs on the cook thread. A 30s API call freezes TD for 30s. Use `webclientDAT` (async) for anything potentially slow.
+3. **WebSocket reconnection** — TD does NOT auto-reconnect on disconnect. Implement backoff in `onDisconnect`.
+4. **Serial port permissions on macOS** — TD needs Full Disk Access OR the port needs to be unlocked via `sudo chmod 666 /dev/cu.usbmodem...` per session.
+5. **MQTT broker connection state** — `mqttClientDAT` may show `connected=true` but messages don't flow if QoS is wrong or topic ACL blocks. Check broker logs.
+6. **JSON parse errors crash callbacks silently** — wrap parses in try/except and log to textport. Otherwise the callback just stops firing.
+7. **Firewall on Windows** — first time `webserverDAT` binds, Windows pops a firewall dialog. Approve it or the server is unreachable.
+8. **CORS** — `webserverDAT` doesn't add CORS headers by default. If serving a webapp from a different origin, add `Access-Control-Allow-Origin: *` in the response.
+9. **Polling vs push** — polling burns API quota. Always prefer WebSocket / webhook / MQTT for high-frequency data.
+10. **Floating-point parsing** — sensor data over Serial often comes as strings. `float()` will crash on `'\n'` or `'NaN'`. Validate before converting.
+
+---
+
+## Quick Recipes
+
+| Goal | Op chain |
+|---|---|
+| Periodic API fetch | `timerCHOP` → `chopExecuteDAT` pulses → `webDAT` → `datExecuteDAT` parses |
+| Webhook receiver | `webserverDAT` (port + path), callback writes to ops |
+| Real-time stream | `websocketDAT` client → onReceiveText → CHOP/DAT |
+| Arduino sensor → visual | `serialDAT` → callback → `constantCHOP` → expression on visual op |
+| TD ↔ phone control | `websocketDAT` server + simple HTML page on phone |
+| MQTT IoT integration | `mqttClientDAT` subscribe → callback dispatches by topic |

skills/creative/touchdesigner-mcp/references/midi-osc.md

@@ -0,0 +1,211 @@
+# MIDI / OSC Reference
+
+External controller input and output — MIDI hardware, TouchOSC mobile UIs, OSC routing across the network.
+
+For audio-driven MIDI patterns (track triggers from spectrum analysis), see also `audio-reactive.md`.
+
+---
+
+## MIDI Input — Hardware Controllers
+
+### Discovery
+
+List connected MIDI devices first. Use a `midiinDAT` to enumerate:
+
+```python
+mdat = root.create(midiinDAT, 'mid_devices')
+# Read available device names from the DAT after one cook
+```
+
+Or via Python directly:
+
+```python
+# In td_execute_python
+import td
+devices = [d for d in op.MIDI.devices]   # verify with td_get_docs('midi')
+```
+
+Verify the API with `td_get_docs(topic='midi')` since this varies between TD versions.
+
+### MIDI In CHOP
+
+Standard pattern:
+
+```python
+midi_in = root.create(midiinCHOP, 'midi_in')
+midi_in.par.device = 0               # device index from discovery
+midi_in.par.activechan = True
+```
+
+Output channels follow the convention `chCcN` and `chCnN`:
+- `ch1c74` — channel 1, CC 74
+- `ch1n60` — channel 1, note 60 (middle C) — value is velocity 0-127
+
+**Map a CC to a parameter:**
+
+```python
+op('/project1/bloom1').par.threshold.mode = ParMode.EXPRESSION
+op('/project1/bloom1').par.threshold.expr = "op('midi_in')['ch1c74'][0] / 127.0"
+```
+
+**Map a note as a trigger:**
+
+Notes in `midiinCHOP` output velocity while held, 0 when released. Use a `triggerCHOP` to convert a held note into pulses:
+
+```python
+trig = root.create(triggerCHOP, 'note_trig')
+trig.par.threshold = 1
+trig.par.triggeron = 'increase'
+trig.inputConnectors[0].connect(op('midi_in'))
+# Filter to a single channel via a selectCHOP if desired
+```
+
+### MIDI Learn Pattern
+
+Build a reusable learn pattern when you don't know the controller's CC layout in advance:
+
+1. Drop a `midiinCHOP` and `selectCHOP` after it.
+2. User wiggles the controller knob.
+3. Use `td_read_chop` on the midiinCHOP to identify which channel is non-zero — that's the active CC.
+4. Set the `selectCHOP.par.channames` to that channel name.
+5. Save the mapping to a `tableDAT` so it persists across sessions.
+
+---
+
+## MIDI Output
+
+```python
+midi_out = root.create(midioutCHOP, 'midi_out')
+midi_out.par.device = 0
+midi_out.par.outputformat = 'continuous'    # 'continuous' | 'event'
+
+# Drive an output: send out a CC mapped from any 0-1 source
+src = root.create(constantCHOP, 'cc_src')
+src.par.name0 = 'ch1c20'
+src.par.value0 = 0.5
+midi_out.inputConnectors[0].connect(src)
+```
+
+For note events specifically, use `event` mode and pulse the value with a `pulseCHOP` or `triggerCHOP`.
+
+---
+
+## OSC Input — Network Control
+
+OSC is the more flexible cousin of MIDI. Used heavily for:
+- TouchOSC / Lemur mobile control surfaces
+- Show control systems (QLab, Watchout)
+- Inter-application sync (Ableton via Max for Live, Resolume, etc.)
+
+### OSC In CHOP
+
+```python
+osc_in = root.create(oscinCHOP, 'osc_in')
+osc_in.par.port = 7000             # listen on UDP 7000
+osc_in.par.localaddress = ''       # empty = all interfaces
+osc_in.par.queued = False          # immediate vs. queued processing
+```
+
+Each incoming OSC address becomes a channel. `/scene/1/intensity` becomes a channel named `scene_1_intensity` (TD sanitizes slashes to underscores).
+
+**Common gotcha:** TD only creates the channel after the FIRST message arrives at that address. Send a "hello" message from the controller during setup, or pre-declare channel names manually.
+
+### OSC In DAT (for raw events)
+
+Use a `oscinDAT` when you need full message access (multiple typed args, addresses with brackets/regex).
+
+```python
+osc_dat = root.create(oscinDAT, 'osc_events')
+osc_dat.par.port = 7001
+# Each row: timestamp, address, type tags, args...
+```
+
+Drive logic via a `datExecuteDAT` watching the `oscinDAT`:
+
+```python
+def onTableChange(dat):
+    last = dat[dat.numRows - 1, 'message']
+    parsed = last.val.split()
+    addr = parsed[0]
+    args = parsed[1:]
+    if addr == '/scene/trigger':
+        op('/project1/scene_switcher').par.index = int(args[0])
+    return
+```
+
+---
+
+## OSC Output — Sending to External Apps
+
+```python
+osc_out = root.create(oscoutCHOP, 'osc_out')
+osc_out.par.netaddress = '127.0.0.1'    # destination IP
+osc_out.par.port = 9000
+
+# Channel names become OSC addresses
+src = root.create(constantCHOP, 'send')
+src.par.name0 = 'scene/intensity'        # → /scene/intensity
+src.par.value0 = 0.7
+osc_out.inputConnectors[0].connect(src)
+```
+
+**Channel-to-address mapping:** TD prepends `/` automatically. Use `/` in channel names to nest.
+
+For one-shot string/typed messages, use `oscoutDAT` and call `.sendOSC(address, args)`:
+
+```python
+op('osc_out_dat').sendOSC('/scene/trigger', [1, 'fade'])
+```
+
+---
+
+## TouchOSC / Mobile UI Pattern
+
+Common setup for live VJ control from a phone/tablet:
+
+1. **Configure TouchOSC layout** — assign each control an OSC address like `/vj/master`, `/vj/scene/1`, etc.
+2. **Find your machine's LAN IP** — TouchOSC needs to point at it.
+3. **TD listens** on `oscinCHOP.par.port = 8000` (or whichever).
+4. **Map channels to params** via expressions:
+
+```python
+op('/project1/master_level').par.opacity.mode = ParMode.EXPRESSION
+op('/project1/master_level').par.opacity.expr = "op('osc_in')['vj_master']"
+```
+
+5. **Send feedback** to the controller via `oscoutCHOP` — useful for syncing state across multiple devices.
+
+---
+
+## Network / Multi-Machine
+
+OSC over LAN works out-of-the-box. For multi-TD-instance sync (e.g., projection cluster):
+
+- One TD acts as **master**, broadcasts `/sync/...` over OSC
+- Worker TDs run `oscinCHOP` listening on the same port
+- Use UDP **broadcast address** (e.g., `192.168.1.255`) on the master's `oscoutCHOP.par.netaddress` to hit all peers
+
+For reliability over WAN, use `webserverDAT` or `websocketDAT` with an external relay instead — UDP loss is invisible.
+
+---
+
+## Pitfalls
+
+1. **MIDI device indexing** — device `0` is whichever device TD enumerated first. Reorder may shift it. Pin by name when possible.
+2. **OSC channel names** — TD doesn't create a channel until the first message lands. New channels invalidate cooked dependents on first arrival, causing a one-frame stutter.
+3. **OSC queued mode** — `par.queued = True` defers processing to a single per-frame batch. Lower latency but messages arriving same frame collapse to the last value. Off for triggers, on for continuous knobs.
+4. **MIDI clock vs. transport** — `midiinCHOP` reports clock if available. Use `midisyncCHOP` (if your TD version exposes it) or compute BPM from clock pulses (24 per quarter note).
+5. **Latency** — wired MIDI is ~1-3ms. WiFi OSC is 10-30ms with jitter. Use wired for tight beat-locked work.
+6. **Port conflicts** — only one process can bind a UDP port on most OS. If `oscinCHOP` shows no traffic, check that another app (Max, Ableton, etc.) isn't already listening on that port.
+
+---
+
+## Quick Recipes
+
+| Goal | Op chain |
+|---|---|
+| Knob → bloom intensity | `midiinCHOP` → expression on `bloom.par.threshold` |
+| Note → scene change | `midiinCHOP` → `triggerCHOP` → `selectCHOP` → drive `switchTOP.par.index` |
+| Phone slider → master fader | TouchOSC `/master` → `oscinCHOP` → expression on output `level.par.opacity` |
+| TD → Resolume scene trigger | `oscoutCHOP` channel `composition/layers/1/clips/1/connect` → Resolume listening on 7000 |
+| Multi-projector sync | Master TD `oscoutCHOP` broadcast → workers `oscinCHOP` |

skills/creative/touchdesigner-mcp/references/panel-ui.md

@@ -0,0 +1,281 @@
+# Panel & UI Reference
+
+Interactive control surfaces inside TouchDesigner — buttons, sliders, fields, custom parameter pages, panel callbacks. For HUD overlays (rendered text on visuals) see `layout-compositor.md`.
+
+Use cases:
+- VJ control rack (master fader, scene buttons, FX toggles)
+- Installation operator console
+- Self-contained TOX components with their own parameter UIs
+- Phone-style touch interfaces displayed on a tablet
+
+---
+
+## Two Layers of UI
+
+| Layer | What it is | Use for |
+|---|---|---|
+| **Custom Parameters** | Params on any COMP, edited like built-in TD params | Configurable components, presets, "settings" panels |
+| **Panel COMPs** | Visible widgets (button, slider, field) inside a containerCOMP | Interactive control surfaces, real-time UIs |
+
+Combine both: build a containerCOMP with panel widgets that read/write custom parameters on a parent component.
+
+---
+
+## Custom Parameters
+
+Add user-editable params to any COMP. Params persist with the COMP, drive expressions, and survive save/reload.
+
+```python
+# Add a custom page to a baseCOMP
+comp = op('/project1/my_component')
+page = comp.appendCustomPage('Controls')
+
+# Add typed params
+page.appendFloat('Intensity', label='Intensity')[0]   # returns a Par
+page.appendInt('Count', label='Count')[0]
+page.appendToggle('Enabled', label='Enabled')[0]
+page.appendMenu('Mode', menuNames=['off', 'soft', 'hard'], menuLabels=['Off', 'Soft', 'Hard'])[0]
+page.appendStr('Title', label='Title')[0]
+page.appendRGB('Color', label='Color')                # returns 3 pars
+page.appendXY('Offset', label='Offset')               # returns 2 pars
+page.appendPulse('Reset', label='Reset')[0]
+page.appendFile('TextureFile', label='Texture')[0]
+```
+
+**Read/write from anywhere:**
+
+```python
+val = op('/project1/my_component').par.Intensity.eval()
+op('/project1/my_component').par.Intensity = 0.7
+```
+
+**Drive other params via expression:**
+
+```python
+op('bloom1').par.threshold.mode = ParMode.EXPRESSION
+op('bloom1').par.threshold.expr = "op('/project1/my_component').par.Intensity"
+```
+
+**Pulse handler (Reset button):**
+
+Use a `parameterExecuteDAT` watching the COMP's pulse params. See `dat-scripting.md`.
+
+---
+
+## Panel COMPs — The Widgets
+
+Each is a COMP that renders as a clickable/draggable widget inside a `containerCOMP`.
+
+| Type | Type Name | Use |
+|---|---|---|
+| Button | `buttonCOMP` | Click action — momentary or toggle |
+| Slider | `sliderCOMP` | Drag to set 0-1 value (1D or 2D) |
+| Field | `fieldCOMP` | Text input |
+| Container | `containerCOMP` | Layout + visual styling, holds children |
+| Select | `selectCOMP` | Reference and display content from another COMP |
+| List | `listCOMP` | Scrollable list with row callbacks |
+
+### Button
+
+```python
+btn = root.create(buttonCOMP, 'play_btn')
+btn.par.w = 120; btn.par.h = 40
+btn.par.buttontype = 'momentary'    # 'momentary' | 'toggleup' | 'togglepress' | 'radio'
+btn.par.bgcolorr = 0.1; btn.par.bgcolorg = 0.1; btn.par.bgcolorb = 0.1
+btn.par.text = 'Play'
+
+# Read state
+state = btn.panel.state          # 1 when active
+```
+
+### Slider
+
+```python
+sld = root.create(sliderCOMP, 'master_fader')
+sld.par.w = 60; sld.par.h = 300
+sld.par.style = 'vertical'        # 'vertical' | 'horizontal' | 'xy'
+sld.par.value0min = 0.0
+sld.par.value0max = 1.0
+
+# Drive a parameter via expression (always-on, no callback needed)
+op('/project1/master_level').par.opacity.mode = ParMode.EXPRESSION
+op('/project1/master_level').par.opacity.expr = "op('master_fader').panel.u"
+```
+
+`panel.u` and `panel.v` give the 0-1 normalized values. For 2D sliders both are populated.
+
+### Field (Text Input)
+
+```python
+fld = root.create(fieldCOMP, 'scene_name')
+fld.par.w = 200; fld.par.h = 30
+fld.par.fieldtype = 'string'      # 'string' | 'integer' | 'float'
+
+# Read current text
+text = fld.panel.field            # the text content
+```
+
+### List
+
+For scrollable lists with selectable rows, use the docked `list1_callbacks` DAT to handle row interactions. Set up cells via the `list_definition` table DAT.
+
+---
+
+## Container COMP — Layout & Styling
+
+`containerCOMP` is the primary parent for grouping widgets and arranging layouts.
+
+```python
+panel = root.create(containerCOMP, 'control_panel')
+panel.par.w = 400; panel.par.h = 600
+panel.par.bgcolorr = 0.05
+panel.par.bgcolorg = 0.05
+panel.par.bgcolorb = 0.05
+panel.par.bgalpha = 1.0
+
+# Layout child panels in vertical stack
+panel.par.align = 'lefttoright'   # 'lefttoright' | 'toptobottom' | etc.
+```
+
+Children are positioned automatically based on `par.align`. For absolute positioning use `par.align = 'fillresize'` and set each child's `par.x` / `par.y`.
+
+### Layout Strategies
+
+| `par.align` | Behavior |
+|---|---|
+| `lefttoright` | Children stacked horizontally |
+| `toptobottom` | Children stacked vertically |
+| `righttoleft` / `bottomtotop` | Reversed stacks |
+| `fillresize` | Children sized to fill, manual positioning |
+| `top` / `bottom` / `left` / `right` | Fixed positioning |
+
+For complex grids: nest containers — vertical container holding horizontal containers.
+
+---
+
+## Panel Callbacks — Reacting to Events
+
+`panelExecuteDAT` watches a panel and fires Python callbacks on user interaction.
+
+```python
+pe = root.create(panelExecuteDAT, 'btn_handler')
+pe.par.panel = '/project1/play_btn'
+pe.par.click = True              # respond to clicks
+pe.par.value = True              # respond to value changes
+```
+
+In its docked DAT:
+
+```python
+def onOffToOn(panelValue):
+    # Click pressed
+    op('/project1/scene_timer').par.start.pulse()
+    return
+
+def onOnToOff(panelValue):
+    # Click released
+    return
+
+def onValueChange(panelValue):
+    # Slider drag, field change, etc.
+    new_val = panelValue.eval()
+    op('/project1/master').par.opacity = new_val
+    return
+```
+
+For pulse params on custom-parameter pages, use a `parameterExecuteDAT` instead.
+
+---
+
+## Building a Complete VJ Control Panel
+
+End-to-end pattern:
+
+```python
+# 1. Top-level container
+panel = root.create(containerCOMP, 'vj_control')
+panel.par.w = 800; panel.par.h = 200
+panel.par.align = 'lefttoright'
+
+# 2. Master fader column
+master_col = panel.create(containerCOMP, 'master')
+master_col.par.w = 120; master_col.par.h = 200
+master_col.par.align = 'toptobottom'
+
+master_label = master_col.create(textTOP, 'lbl')
+master_label.par.text = 'MASTER'
+
+master_sld = master_col.create(sliderCOMP, 'fader')
+master_sld.par.w = 60; master_sld.par.h = 150
+master_sld.par.style = 'vertical'
+
+# 3. Scene buttons row
+scene_col = panel.create(containerCOMP, 'scenes')
+scene_col.par.w = 400; scene_col.par.h = 200
+scene_col.par.align = 'lefttoright'
+for i in range(8):
+    b = scene_col.create(buttonCOMP, f'scene_{i+1}')
+    b.par.w = 50; b.par.h = 50
+    b.par.text = str(i+1)
+    b.par.buttontype = 'radio'      # only one active at a time
+
+# 4. FX toggle column
+fx_col = panel.create(containerCOMP, 'fx')
+fx_col.par.w = 280; fx_col.par.h = 200
+fx_col.par.align = 'toptobottom'
+for fx in ['Bloom', 'CRT', 'Glitch', 'Strobe']:
+    t = fx_col.create(buttonCOMP, fx.lower())
+    t.par.w = 220; t.par.h = 35
+    t.par.text = fx
+    t.par.buttontype = 'toggleup'
+
+# 5. Display in a window
+win = root.create(windowCOMP, 'control_win')
+win.par.winop = panel.path
+win.par.winw = 800; win.par.winh = 200
+win.par.borders = True
+win.par.winopen.pulse()
+```
+
+Then wire panel values to ops via expressions or panelExecuteDATs.
+
+---
+
+## Showing the Panel — Window or Embedded
+
+| Approach | When |
+|---|---|
+| `windowCOMP` pointing at panel | Standalone control surface, separate display |
+| Render the containerCOMP via `renderTOP` | Composite UI over visuals (HUD-style) |
+| Use a `panelCOMP` directly inside a network editor pane | Designer/dev preview only — panel is fully interactive |
+
+For a touch-screen tablet, use a `windowCOMP` on a second display routed to the tablet's HDMI input.
+
+---
+
+## Pitfalls
+
+1. **Panel won't respond to clicks** — likely `par.disabled = True` or the parent container has `par.disableinputs = True`. Check the panel hierarchy.
+2. **Slider value not updating** — `panel.u/v` reads the visual position. If you set `par.value0` directly, the visual lags. Use `par.value0` AS the source of truth and let the slider follow.
+3. **Custom param won't appear** — must call `appendCustomPage` first, then append params. Pages with no params don't show.
+4. **Custom param disappears on reload** — params added via Python at runtime persist only if the COMP is saved AFTER. Use a `tox` save (`comp.save('mycomp.tox')`) or commit via `td_execute_python` then save the project.
+5. **Event callback fires twice** — both `onOffToOn` and `onValueChange` may fire on a single button press. Pick one to handle the action; don't double-trigger.
+6. **Pulse params need `.pulse()`** — setting `par.X = True` on a pulse param does nothing. Always use `.pulse()`.
+7. **Field text doesn't commit until Tab/Enter** — fields don't fire callbacks while typing. Use `par.committemode = 'all'` to fire on every keystroke (heavy).
+8. **`par.text` vs panel content** — `buttonCOMP.par.text` is the LABEL on the button. The button's STATE is `panel.state` (0/1). Don't confuse them.
+9. **Touch input on macOS** — multi-touch via direct touch panels works but TD's gesture handling is rudimentary. For complex multi-touch (pinch/rotate), use TouchOSC on a tablet instead.
+10. **Layout doesn't update** — changing `par.align` requires the container to re-cook. Touch a child or pulse the container to trigger.
+
+---
+
+## Quick Recipes
+
+| Goal | Setup |
+|---|---|
+| Master fader | `sliderCOMP` (vertical) → expression on `level.par.opacity` |
+| Scene picker | 8 `buttonCOMP` (radio) → `selectCHOP` on their state → drive `switchTOP.par.index` |
+| FX toggle | `buttonCOMP` (toggleup) → expression on `bypass` of an FX op |
+| Numeric input | `fieldCOMP` (float) → expression on target par |
+| Component settings | Custom params on the component COMP, panel widgets inside drive them |
+| Touch tablet UI | `containerCOMP` with widgets → `windowCOMP` to second display |
+| Status display | `textTOP` rendered into the panel via `selectCOMP` |

skills/creative/touchdesigner-mcp/references/particles.md

@@ -0,0 +1,245 @@
+# Particles Reference
+
+Particle systems in TouchDesigner — modern POPs (Particle Operators) and the legacy particleSOP path.
+
+For instancing static geometry (without per-instance lifetime/velocity), see `geometry-comp.md`. For GLSL-driven feedback simulations (no particle abstraction), see `operator-tips.md` (Feedback TOP section).
+
+Always call `td_get_par_info` for the op type before setting params. Param names below reflect TD 2025.32 — verify before relying on them.
+
+---
+
+## Two Paths: POPs vs. SOPs
+
+| | **POP family** (modern) | **particleSOP** (legacy) |
+|---|---|---|
+| GPU? | Yes (compute) | No (CPU) |
+| Particle count | 100k+ comfortably | ~5k before slowdown |
+| API style | Source / Force / Solver / Render chain | Single op with many params |
+| Use for | New projects, anything intensive | Quick demos, low counts, TD < 2023 |
+
+**Default to POPs.** Only fall back to particleSOP if a POP variant of an op you need doesn't exist.
+
+---
+
+## POP Pipeline Overview
+
+A POP system is a chain of operators inside a `geometryCOMP`:
+
+```
+popSourceTOP / popSourceSOP   ← spawn new particles
+        ↓
+popForceTOP (gravity, wind, etc.)
+        ↓
+popForceTOP (attractor, vortex, ...)
+        ↓
+popDeleteTOP (lifetime, bounds)
+        ↓
+popSolverTOP                  ← integrates velocity, updates positions
+        ↓
+[render via geometryCOMP / glslMAT instancing]
+```
+
+POP buffers carry standard channels: `P` (position), `v` (velocity), `life`, `id`, `Cd` (color), plus any custom channels you add.
+
+---
+
+## Minimal POP Setup
+
+```python
+# Create a geometry COMP to hold the POP network
+geo = root.create(geometryCOMP, 'particles_geo')
+
+# 1. Source — emit particles from a point
+src = geo.create(popSourceTOP, 'src')
+src.par.birthrate = 500          # per second
+src.par.life = 4.0                # seconds
+
+# 2. Gravity force
+grav = geo.create(popForceTOP, 'gravity')
+grav.par.forcetype = 'gravity'
+grav.par.fy = -9.8
+
+# 3. Lifetime cleanup
+delp = geo.create(popDeleteTOP, 'cull')
+delp.par.condition = 'lifeleq'    # delete when life <= 0
+delp.par.value = 0
+
+# 4. Solver
+solv = geo.create(popSolverTOP, 'solver')
+solv.par.timestep = 'frame'
+
+# Wire: source → force → delete → solver
+src.outputConnectors[0].connect(grav.inputConnectors[0])
+grav.outputConnectors[0].connect(delp.inputConnectors[0])
+delp.outputConnectors[0].connect(solv.inputConnectors[0])
+```
+
+The `popSolverTOP` output IS the live particle buffer. Render it via `glslMAT` instancing on a small SOP (sphere, point) as the "shape" of each particle.
+
+---
+
+## Common Forces
+
+| Force type | Effect | Common params |
+|---|---|---|
+| `gravity` | Constant directional pull | `fx`, `fy`, `fz` |
+| `wind` | Constant velocity addition | `wx`, `wy`, `wz` |
+| `drag` | Velocity damping over time | `dragstrength` |
+| `noise` | Curl-noise turbulence | `noiseamp`, `noisefreq`, `noiseseed` |
+| `attractor` | Pull toward a point | `position`, `strength`, `falloff` |
+| `vortex` | Swirl around an axis | `axis`, `strength` |
+| `point` (custom) | GLSL-evaluated arbitrary force | via `popforceadvancedTOP` |
+
+Stack multiple `popForceTOP`s in series — each modifies velocity additively.
+
+---
+
+## Lifecycle Patterns
+
+### Continuous emission (e.g. smoke plume)
+
+```python
+src.par.birthrate = 800
+src.par.life = 6.0       # variance via 'lifevariance'
+src.par.lifevariance = 1.5
+```
+
+### Burst emission (e.g. explosion)
+
+```python
+src.par.birthrate = 0    # no continuous emission
+src.par.burst.pulse()    # one burst on demand (verify param name)
+src.par.burstcount = 5000
+src.par.life = 1.5
+```
+
+### Beat-triggered burst
+
+Wire a `triggerCHOP` (from audio or MIDI) to pulse the burst:
+
+```python
+op('/project1/audio_kick_trigger').outputConnectors[0].connect(...)
+# Then via a chopExecuteDAT, on each kick:
+def offToOn(channel, sampleIndex, val, prev):
+    op('/project1/particles_geo/src').par.burst.pulse()
+    return
+```
+
+---
+
+## Rendering Particles
+
+### Point Sprites (simplest)
+
+```python
+# Inside the geometryCOMP, render the solver output directly
+# The geo's first SOP child becomes the geometry
+# But for POPs, we typically render via glslMAT on a small "shape"
+
+# Simple billboard sphere per particle:
+shape = geo.create(sphereSOP, 'shape')
+shape.par.rad = 0.05
+shape.par.rows = 6; shape.par.cols = 6   # low-poly to keep it fast
+
+# Material that uses POP buffer for instancing
+mat = root.create(glslMAT, 'particle_mat')
+# Configure mat.par.instancingTOP = solver output (verify param name)
+```
+
+The exact instancing setup varies by TD version — call `td_get_hints(topic='popInstancing')` (or `popRender` / `instancing` — try a few).
+
+### GPU Sprites via glslcopyPOP
+
+For dense smoke/fire-like effects, use a `glslcopyPOP` that writes per-particle color/size from a compute shader, then render as point sprites with additive blending in a `renderTOP`.
+
+---
+
+## Collisions
+
+```python
+# Collision detection against an SOP
+coll = geo.create(popCollideTOP, 'ground_coll')
+coll.par.collidewithsop = '/project1/ground_geo'  # path to colliding SOP
+coll.par.bounce = 0.3
+coll.par.friction = 0.1
+# Insert between force and solver
+```
+
+For plane/box collisions only, use `popPlaneCollideTOP` (cheaper).
+
+---
+
+## Custom Per-Particle Data
+
+Add a custom channel via `popAttribCreateTOP` (or by writing through `glslcopyPOP`):
+
+```python
+# Add a "phase" attribute initialized random per-particle, used in render shader
+attr = geo.create(popAttribCreateTOP, 'add_phase')
+attr.par.attribname = 'phase'
+attr.par.value0 = 'rand(@id)'   # expression in TD's POP attribute language
+```
+
+Then in the render shader, `texture(sTDPOPInputs[0].phase, ...)` (or whichever sampler convention your TD version uses — verify with `td_get_docs(topic='pops')`).
+
+---
+
+## Legacy particleSOP (Use Sparingly)
+
+For quick demos or low-count systems:
+
+```python
+# Inside a geo
+psrc = geo.create(addSOP, 'point_src')      # source: a single point
+psrc.par.points = '0 0 0'
+
+part = geo.create(particleSOP, 'particles')
+part.par.life = 3.0
+part.par.birthrate = 100
+part.par.gravityy = -9.8
+part.par.windx = 0.5
+part.inputConnectors[0].connect(psrc)
+```
+
+CPU-bound. Beyond ~5,000 active particles you'll see frame drops.
+
+---
+
+## Pitfalls
+
+1. **Particles don't appear** — usually a render-side issue. Check via `td_get_screenshot` on the solver output (renders the buffer as a TOP-like view in newer TD). Then check the `geometryCOMP`'s render path.
+2. **Burst won't fire** — verify the `burst` param is a pulse, not a toggle. Pulses must use `.pulse()`, not `= True`.
+3. **Particles teleport on first frame** — uninitialized velocity. Set `popSourceTOP.par.initialvelocityX/Y/Z` or zero them explicitly.
+4. **Gravity feels wrong** — TD's "1 unit" depends on your scene scale. Start with `fy = -1.0` and scale up rather than using real-world 9.8.
+5. **High birthrate = stuttering** — birthrate is per-second, not per-frame. At 60fps, `birthrate = 6000` is 100/frame which is fine; `birthrate = 600000` will tank.
+6. **POP solver order matters** — forces apply in the order they appear in the chain. Putting gravity AFTER drag dampens gravity itself; usually not what you want.
+7. **Instancing param name varies** — `mat.par.instancingTOP` vs. `mat.par.instanceop` vs. `mat.par.instances` differs across TD versions. Always check `td_get_par_info(op_type='glslMAT')`.
+8. **Cooking dependency loops** — POP solvers create implicit time-loops. The "cook dependency loop" warning is expected and harmless for POPs.
+9. **CHOP-driven force values** — when a force param is expression-bound to a CHOP (e.g., audio-reactive gravity), make sure the CHOP cooks before the solver. If not, force lags by one frame.
+
+---
+
+## Performance Targets
+
+| Particle count | Setup | Frame budget @ 60fps |
+|---|---|---|
+| < 1k | particleSOP fine | trivial |
+| 1k - 10k | POPs, simple forces | ~2-5ms |
+| 10k - 100k | POPs, GPU-only forces | ~5-15ms |
+| 100k+ | `glslcopyPOP`, custom compute | ~10-25ms |
+| 1M+ | Custom GPU buffer, no POP framework | depends on shader |
+
+Use `td_get_perf` to find which op in the POP chain is the bottleneck.
+
+---
+
+## Quick Recipes
+
+| Goal | Pipeline |
+|---|---|
+| Smoke plume | `popSourceTOP` (point) → gravity + wind + noise → `popDeleteTOP` (life) → solver → glslMAT instancing |
+| Beat-triggered burst | `triggerCHOP` (audio) → chopExecuteDAT pulses `popSourceTOP.par.burst` |
+| Fireworks shell | Burst at point → drag + gravity → secondary burst on lifetime threshold |
+| Snow/rain | Continuous emission across XZ plane (high y), gravity + small wind, infinite life box-deleted |
+| Sparks | Burst, very short life (0.3s), bright additive render, motion blur via feedback |
+| Audio particles | Birthrate driven by audio envelope, color driven by frequency band |

skills/creative/touchdesigner-mcp/references/projection-mapping.md

@@ -0,0 +1,211 @@
+# Projection Mapping Reference
+
+Multi-window output, surface mapping, edge blending, and projector calibration patterns for installation/event work.
+
+For HUD layouts and on-screen panel grids, see `layout-compositor.md`. For wireframe/test-pattern generation, see `operator-tips.md`.
+
+---
+
+## Window COMP — Output to a Display
+
+The `windowCOMP` is how TD pushes pixels to a real display.
+
+```python
+win = root.create(windowCOMP, 'output_window')
+win.par.winop = '/project1/final_out'   # path to the TOP being displayed
+win.par.winw = 1920
+win.par.winh = 1080
+win.par.winoffsetx = 0                  # screen-space offset
+win.par.winoffsety = 0
+win.par.borders = False                 # no chrome
+win.par.alwaysontop = True
+win.par.cursor = False                  # hide cursor in fullscreen
+win.par.justify = 'fillaspect'          # 'fill' | 'fitaspect' | 'fillaspect' | 'native'
+win.par.winopen.pulse()                 # OPEN the window
+```
+
+To target a specific physical display, set `par.location`:
+
+```python
+win.par.location = 'secondary'          # 'primary' | 'secondary' | 'monitor1' | 'monitor2' | ...
+```
+
+Or set absolute coordinates using `winoffsetx/y` matched to your OS display layout.
+
+**Always pulse `winopen` — setting params alone doesn't open the window.**
+
+---
+
+## Multi-Window Output
+
+For multi-projector or multi-display setups, create one `windowCOMP` per output, each pointing at a different TOP.
+
+```python
+for i, screen_top in enumerate(['out_left', 'out_center', 'out_right']):
+    w = root.create(windowCOMP, f'win_{i}')
+    w.par.winop = f'/project1/{screen_top}'
+    w.par.winw = 1920; w.par.winh = 1080
+    w.par.winoffsetx = i * 1920
+    w.par.winoffsety = 0
+    w.par.borders = False
+    w.par.alwaysontop = True
+    w.par.cursor = False
+    w.par.winopen.pulse()
+```
+
+For ultra-wide single-output spans, use ONE windowCOMP at e.g. 5760×1080 spanning three projectors via the GPU's mosaic/spanning mode (Nvidia Mosaic, AMD Eyefinity), then split content via `cropTOP` per screen inside TD.
+
+---
+
+## 4-Point Corner Pin (Quad Warp)
+
+The simplest projection mapping primitive — warping a rectangle onto a quadrilateral.
+
+```python
+# Source content
+src = op('/project1/scene_out')
+
+# Manual: cornerPinTOP (TD has this built-in)
+cp = root.create(cornerPinTOP, 'corner_pin')
+cp.par.tlx = 0.05; cp.par.tly = 0.10    # top-left (normalized 0-1)
+cp.par.trx = 0.95; cp.par.try = 0.08    # top-right
+cp.par.brx = 0.93; cp.par.bry = 0.92    # bottom-right
+cp.par.blx = 0.07; cp.par.bly = 0.94    # bottom-left
+cp.inputConnectors[0].connect(src)
+```
+
+Alternative: use a `geometryCOMP` with a `gridSOP` and bend the verts in vertex GLSL. More flexible (curved surfaces) but more setup.
+
+Verify TD 2025.32 param names with `td_get_par_info(op_type='cornerPinTOP')`.
+
+---
+
+## Bezier / Mesh Warp (Curved Surfaces)
+
+For non-flat surfaces (domes, columns, curved walls), use a subdivided mesh and per-vertex displacement.
+
+### Pattern: Grid Mesh + GLSL Displacement
+
+```python
+# Subdivided grid in a geo
+geo = root.create(geometryCOMP, 'warp_geo')
+grid = geo.create(gridSOP, 'warp_grid')
+grid.par.rows = 32          # higher = smoother curve
+grid.par.cols = 32
+grid.par.sizex = 2; grid.par.sizey = 2
+
+# Texture the source onto it
+mat = root.create(constMAT, 'warp_mat')      # use constMAT for unlit projection
+mat.par.maptop = '/project1/scene_out'        # source TOP
+
+geo.par.material = mat.path
+
+# Render to a TOP that goes to the projector window
+cam = root.create(cameraCOMP, 'cam_proj')
+cam.par.tz = 4
+
+render = root.create(renderTOP, 'projection_out')
+render.par.camera = cam.path
+render.par.geometry = geo.path
+render.par.outputresolution = 'custom'
+render.par.resolutionw = 1920; render.par.resolutionh = 1080
+```
+
+For per-vertex offsets, write a vertex GLSL on the constMAT (or use `glslMAT`) and read displacement values from a CHOP via uniform.
+
+Calibration is iterative: render a checkerboard from `scene_out`, project it, photograph the projection, manually nudge corner/grid points until aligned.
+
+---
+
+## Edge Blending (Multi-Projector Overlap)
+
+When two projectors overlap, the overlap region is twice as bright. Blend by ramping each projector's edge alpha to 0 across the overlap zone.
+
+### GLSL Edge Blend Shader
+
+Per-projector output pass that fades the inside edge to black:
+
+```glsl
+// edge_blend_pixel.glsl
+out vec4 fragColor;
+uniform float uBlendLeft;     // overlap width on left edge (0-0.5, 0=no blend)
+uniform float uBlendRight;
+uniform float uGamma;          // typically 2.2 — perceptual ramp
+
+void main() {
+    vec2 uv = vUV.st;
+    vec4 col = texture(sTD2DInputs[0], uv);
+
+    float aL = (uBlendLeft  > 0.0) ? smoothstep(0.0, uBlendLeft, uv.x) : 1.0;
+    float aR = (uBlendRight > 0.0) ? smoothstep(0.0, uBlendRight, 1.0 - uv.x) : 1.0;
+    float a = pow(aL * aR, uGamma);
+
+    fragColor = TDOutputSwizzle(vec4(col.rgb * a, 1.0));
+}
+```
+
+Apply this to each overlap-touching projector's output. Tune `uBlendLeft` / `uBlendRight` to match your physical overlap.
+
+For top/bottom blends or cylindrical setups, extend the shader with `uBlendTop` / `uBlendBottom`.
+
+---
+
+## Calibration Patterns
+
+Useful test patterns for aligning projectors. Build a `switchTOP` selecting one of these, route to all projector windows during setup.
+
+```python
+# Solid white — for brightness/uniformity check
+white = root.create(constantTOP, 'cal_white')
+white.par.colorr = 1.0; white.par.colorg = 1.0; white.par.colorb = 1.0
+
+# Centered crosshair — for keystone alignment
+gridcross = root.create(textTOP, 'cal_cross')
+gridcross.par.text = '+'
+gridcross.par.fontsizex = 200
+
+# Fine grid — for warp/mesh alignment (use rampTOP + math + threshold, or build via GLSL)
+# Color bars for projector color calibration
+bars = root.create(rampTOP, 'cal_bars')
+bars.par.type = 'horizontal'
+```
+
+Or use the bundled `testpatternTOP` if your TD version includes it.
+
+---
+
+## Projection Audit Workflow
+
+When debugging a multi-screen setup:
+
+1. Render a unique color and label per output (`textTOP` saying "LEFT", "CENTER", "RIGHT").
+2. Check that each window is sourcing the correct path: `td_get_operator_info(path='/project1/win_0')`.
+3. Verify display assignment: walk to each projector and confirm visually.
+4. Check resolution: physical projector native res vs. TD output res — mismatches cause scaling artifacts.
+5. Cook flag: `td_get_perf` — if a window's source TOP isn't cooking, the projector shows last frame frozen.
+
+---
+
+## Pitfalls
+
+1. **Window won't open** — you forgot `winopen.pulse()`. Setting params alone doesn't open it.
+2. **Wrong display** — `par.location='secondary'` depends on OS display order. Set `winoffsetx/y` to absolute coords as a more reliable override.
+3. **Cursor visible** — set `par.cursor = False` BEFORE opening, or close+reopen.
+4. **Black projection** — usually a cooking issue. Verify `final_out` TOP is cooking via `td_get_perf`. Check `td_get_errors` recursively from `/`.
+5. **Tearing / vsync** — `windowCOMP` honors `par.vsync`. For projection always set `vsync='vsync'` (default). Tearing means GPU is over-budget — reduce render resolution.
+6. **Aspect mismatch** — projector native is often 1920×1200 (16:10) not 1080. Use `justify='fitaspect'` or render at native projector res.
+7. **Non-Commercial license** — caps total resolution at 1280×1280. For real installation work you need Commercial. Pro license adds 4K+.
+8. **Multiple monitors on macOS** — `windowCOMP` honors macOS Spaces. Disable Spaces or pin TD to a specific display in System Settings before showtime.
+
+---
+
+## Quick Recipes
+
+| Goal | Approach |
+|---|---|
+| Single fullscreen output | One `windowCOMP`, `justify='fillaspect'`, `winopen.pulse()` |
+| 3-projector wide span | 3 `windowCOMP` + per-output `cropTOP` from one wide source |
+| Single quad surface | `cornerPinTOP` → `windowCOMP` |
+| Curved/dome | Subdivided gridSOP with vertex GLSL → `renderTOP` → `windowCOMP` |
+| Edge blend overlap | GLSL fade shader per projector → `windowCOMP` |
+| Calibration mode | `switchTOP` between scene and test patterns, hot-key triggered |

skills/creative/touchdesigner-mcp/references/replicator.md

@@ -0,0 +1,198 @@
+# Replicator COMP Reference
+
+The `replicatorCOMP` clones a template operator N times, driven by a table of data. The fundamental TD pattern for data-driven networks: button grids, scene rosters, dynamic UI, parameter panels per-channel.
+
+For visual instancing (per-pixel/per-render copies), see `geometry-comp.md`. Replicator builds NETWORK NODES; instancing builds RENDER COPIES. Different layer.
+
+---
+
+## Concept
+
+```
+[Template OP]                  [Data tableDAT]
+       │                              │
+       └─────→ replicatorCOMP ←───────┘
+                     │
+                     ▼
+        [N clones], one per data row
+        Each clone gets per-row params
+```
+
+Edit the template once → all clones inherit. Edit the table → clones add/remove dynamically. Push parameter overrides per-row.
+
+---
+
+## Minimal Setup
+
+```python
+# 1. Make a template (the thing to clone)
+template = root.create(buttonCOMP, 'btn_template')
+template.par.w = 80; template.par.h = 80
+template.par.text = 'X'
+template.par.bgcolorr = 0.2
+
+# 2. Make a data table (one row per clone)
+data = root.create(tableDAT, 'scene_data')
+data.appendRow(['name', 'color_r', 'color_g', 'color_b'])
+data.appendRow(['Sunset', 1.0, 0.4, 0.0])
+data.appendRow(['Midnight', 0.0, 0.1, 0.4])
+data.appendRow(['Storm', 0.3, 0.3, 0.5])
+data.appendRow(['Forest', 0.0, 0.5, 0.2])
+
+# 3. Replicator — points at template + data
+rep = root.create(replicatorCOMP, 'scene_buttons')
+rep.par.template = template.path
+rep.par.opfromdat = data.path
+rep.par.namefromdatname = 'name'        # use 'name' column for clone names
+rep.par.incrementalnumbering = False
+```
+
+After cooking, the replicator creates 4 child COMPs named `Sunset`, `Midnight`, `Storm`, `Forest` (one per non-header row), each cloned from `btn_template`.
+
+---
+
+## Per-Row Parameter Overrides
+
+The replicator's docked `replicator1_callbacks` DAT lets you customize each clone:
+
+```python
+def onReplicate(comp, allOps, newOps, template, master):
+    """Called once per replicate cycle. newOps is the list of just-created clones."""
+    data = op('scene_data')
+    for i, clone in enumerate(newOps):
+        row = i + 1                 # +1 to skip header
+        clone.par.text = data[row, 'name'].val
+        clone.par.bgcolorr = float(data[row, 'color_r'].val)
+        clone.par.bgcolorg = float(data[row, 'color_g'].val)
+        clone.par.bgcolorb = float(data[row, 'color_b'].val)
+    return
+```
+
+Or use parameter expressions referencing `digits` (the per-clone index, available as a built-in expression token inside the cloned subtree):
+
+```python
+# Inside the template, set a param expression like:
+# par.value0.expr = "op('../scene_data')[me.digits + 1, 'value']"
+```
+
+`me.digits` resolves to the row index of the current clone. This is the cleanest way for static reference patterns — no callback needed.
+
+---
+
+## Layout: Buttons in a Grid
+
+Drop the replicator inside a `containerCOMP` with auto-layout:
+
+```python
+panel = root.create(containerCOMP, 'scene_panel')
+panel.par.w = 400; panel.par.h = 100
+panel.par.align = 'lefttoright'
+
+# Move the replicator inside
+rep.parent = panel.path           # or create rep as a child of panel directly
+```
+
+Each clone is a child of the replicator (which itself is a child of the panel). The panel auto-arranges everything.
+
+For a 2D grid, set `par.align = 'fillresize'` on the container and override `par.x` / `par.y` per clone in the callback based on row/col index.
+
+---
+
+## Updating Without Rebuilding
+
+When the data table changes, the replicator regenerates the clones. By default it destroys and recreates everything. To preserve state, set:
+
+```python
+rep.par.recreatemissing = True       # only add/remove changed rows
+rep.par.recreateallonchange = False
+```
+
+This pattern is essential for live-edit scenarios (designer adjusts table, network keeps running).
+
+For incremental data ingestion (e.g., from a `webDAT` polling an API), have a `datExecuteDAT` watch the response, parse, write to the data table, and the replicator self-updates.
+
+---
+
+## Common Patterns
+
+### Scene Roster (Data → Buttons + Logic)
+
+```python
+# Data per scene: name, file path, audio track, BPM
+scene_data.appendRow(['name', 'file', 'audio', 'bpm'])
+scene_data.appendRow(['Intro', '/scenes/intro.tox', '/audio/intro.wav', 110])
+scene_data.appendRow(['Main', '/scenes/main.tox', '/audio/main.wav', 128])
+
+# Replicator clones a buttonCOMP per scene
+# Each button's onClick callback loads the corresponding tox + cues audio
+```
+
+### Dynamic Parameter Panel
+
+For a list of audio bands, generate a fader strip per band:
+
+```python
+# Data: band names (sub, low, mid, hi-mid, high, air)
+# Template: containerCOMP with label + sliderCOMP
+# Replicator clones N strips
+# Each slider's value is read at /audio_eq/{band_name}/fader
+```
+
+### Procedural Visual Network
+
+Build a multi-channel visual network from a config file:
+
+```python
+# Data: which TOPs to chain, per "scene"
+# Template: a baseCOMP with placeholder children
+# Replicator builds one baseCOMP per scene; each scene contains a custom chain
+# Switch between scenes via switchTOP.par.index driven by panel
+```
+
+### Per-Channel CHOP Display
+
+Visualize each channel of a multi-channel CHOP separately:
+
+```python
+# Data table: one row per channel (auto-extracted via choptodatDAT)
+# Template: a small chopVis COMP showing one channel
+# Replicator generates N visualizers stacked vertically
+```
+
+---
+
+## Replicator vs. Pure Python Loop
+
+| Approach | When to use |
+|---|---|
+| **replicatorCOMP** | The set of clones changes (add/remove rows live). Visual editor expectations. Pattern is reusable across projects. |
+| **Python loop** (in `td_execute_python`) | One-shot generation. Static set. Simpler logic, no template overhead. Faster to write. |
+
+If you'll only ever build the network once, prefer a Python loop with `td_execute_python`. The replicator earns its weight when data is live.
+
+---
+
+## Pitfalls
+
+1. **Header row** — `tableDAT` rows are 0-indexed. If you have a header, your first data row is index 1. Off-by-one bugs are common in callbacks.
+2. **`namefromdatname` column missing** — replicator silently uses `digits` (numeric suffix) names. Buttons end up named `1`, `2`, `3` instead of meaningful names. Set `par.namefromdatname` explicitly.
+3. **Template lives in network** — the template OP is itself a real network node. Don't connect things downstream of it directly; connect to the clones (or use a `nullCOMP` between).
+4. **Recreate-on-change wipes state** — toggles, slider positions, and uncached data inside clones are lost on each regeneration. Use `recreatemissing` to preserve.
+5. **`onReplicate` doesn't fire on edit** — only fires when the clone set changes. Editing a value WITHIN an existing row doesn't re-trigger. Use `parameterExecuteDAT` or expressions for per-cell live updates.
+6. **Custom params on clones** — pages added in the template propagate. Pages added in `onReplicate` don't survive the next regeneration. Always add custom pages on the template, not the clone.
+7. **Cooking storms** — adding many rows fast triggers many clone events. Bundle adds via Python and call `data.cook(force=True)` once at the end.
+8. **`me.digits` outside replicator children** — `me.digits` only resolves inside an op that's a descendant of the replicator. Don't reference it in unrelated networks.
+9. **Cross-clone references** — referencing a sibling clone via relative path works from inside a clone (`op('../OtherClone/x')`), but breaks if names change. Prefer absolute paths via the data table.
+
+---
+
+## Quick Recipes
+
+| Goal | Setup |
+|---|---|
+| 8-button scene picker | `tableDAT` (8 rows) + `buttonCOMP` template + `replicatorCOMP` |
+| Per-band EQ strip panel | `tableDAT` (band names) + container template (label + slider) + replicator |
+| Data-driven visual scenes | `tableDAT` (scene config) + `baseCOMP` template (visual chain) + replicator |
+| Live-updating clone set | Same as above + `par.recreatemissing = True` |
+| Per-row colored UI | Data table with color cols, `onReplicate` callback sets per-clone colors |
+| List from API response | `webDAT` → `datExecuteDAT` parses JSON → writes to data table → replicator updates |

tests/agent/test_auxiliary_client.py

@@ -516,26 +516,88 @@ class TestGetTextAuxiliaryClient:
         assert isinstance(client, CodexAuxiliaryClient)
         assert model == "gpt-5.2-codex"
 
+    def test_returns_none_when_nothing_available(self, monkeypatch):
+        monkeypatch.delenv("OPENAI_BASE_URL", raising=False)
+        monkeypatch.delenv("OPENAI_API_KEY", raising=False)
+        monkeypatch.delenv("OPENROUTER_API_KEY", raising=False)
+        with patch("agent.auxiliary_client._read_nous_auth", return_value=None), \
+             patch("agent.auxiliary_client._read_codex_access_token", return_value=None), \
+             patch("agent.auxiliary_client._resolve_api_key_provider", return_value=(None, None)):
+            client, model = get_text_auxiliary_client()
+        assert client is None
+        assert model is None
 
-class TestNousAuxiliaryRefresh:
-    def test_try_nous_prefers_runtime_credentials(self):
-        fresh_base = "https://inference-api.nousresearch.com/v1"
+    def test_custom_endpoint_uses_codex_wrapper_when_runtime_requests_responses_api(self):
+        with patch("agent.auxiliary_client._resolve_custom_runtime",
+                   return_value=("https://api.openai.com/v1", "sk-test", "codex_responses")), \
+             patch("agent.auxiliary_client._read_main_model", return_value="gpt-5.3-codex"), \
+             patch("agent.auxiliary_client.OpenAI") as mock_openai:
+            client, model = get_text_auxiliary_client()
+
+        from agent.auxiliary_client import CodexAuxiliaryClient
+        assert isinstance(client, CodexAuxiliaryClient)
+        assert model == "gpt-5.3-codex"
+        assert mock_openai.call_args.kwargs["base_url"] == "https://api.openai.com/v1"
+        assert mock_openai.call_args.kwargs["api_key"] == "sk-test"
+
+
+class TestVisionClientFallback:
+    """Vision client auto mode resolves known-good multimodal backends."""
+
+    def test_vision_auto_includes_active_provider_when_configured(self, monkeypatch):
+        """Active provider appears in available backends when credentials exist."""
+        monkeypatch.setenv("ANTHROPIC_API_KEY", "***")
         with (
-            patch("agent.auxiliary_client._read_nous_auth", return_value={"access_token": "stale-token"}),
-            patch("agent.auxiliary_client._resolve_nous_runtime_api", return_value=("fresh-agent-key", fresh_base)),
-            patch("hermes_cli.models.get_nous_recommended_aux_model", return_value=None),
+            patch("agent.auxiliary_client._read_nous_auth", return_value=None),
+            patch("agent.auxiliary_client._read_main_provider", return_value="anthropic"),
+            patch("agent.auxiliary_client._read_main_model", return_value="claude-sonnet-4"),
+            patch("agent.anthropic_adapter.build_anthropic_client", return_value=MagicMock()),
+            patch("agent.anthropic_adapter.resolve_anthropic_token", return_value="***"),
+        ):
+            backends = get_available_vision_backends()
+
+        assert "anthropic" in backends
+
+    def test_resolve_provider_client_returns_native_anthropic_wrapper(self, monkeypatch):
+        monkeypatch.setenv("ANTHROPIC_API_KEY", "***")
+        with (
+            patch("agent.auxiliary_client._read_nous_auth", return_value=None),
+            patch("agent.anthropic_adapter.build_anthropic_client", return_value=MagicMock()),
+            patch("agent.anthropic_adapter.resolve_anthropic_token", return_value="***"),
+        ):
+            client, model = resolve_provider_client("anthropic")
+
+        assert client is not None
+        assert client.__class__.__name__ == "AnthropicAuxiliaryClient"
+        assert model == "claude-haiku-4-5-20251001"
+
+
+class TestAuxiliaryPoolAwareness:
+    def test_try_nous_uses_pool_entry(self):
+        class _Entry:
+            access_token = "pooled-access-token"
+            agent_key = "pooled-agent-key"
+            inference_base_url = "https://inference.pool.example/v1"
+
+        class _Pool:
+            def has_credentials(self):
+                return True
+
+            def select(self):
+                return _Entry()
+
+        with (
+            patch("agent.auxiliary_client.load_pool", return_value=_Pool()),
             patch("agent.auxiliary_client.OpenAI") as mock_openai,
         ):
             from agent.auxiliary_client import _try_nous
 
-            mock_openai.return_value = MagicMock()
             client, model = _try_nous()
 
         assert client is not None
-        # No Portal recommendation → falls back to the hardcoded default.
         assert model == "google/gemini-3-flash-preview"
-        assert mock_openai.call_args.kwargs["api_key"] == "fresh-agent-key"
-        assert mock_openai.call_args.kwargs["base_url"] == fresh_base
+        assert mock_openai.call_args.kwargs["api_key"] == "pooled-agent-key"
+        assert mock_openai.call_args.kwargs["base_url"] == "https://inference.pool.example/v1"
 
     def test_try_nous_uses_portal_recommendation_for_text(self):
         """When the Portal recommends a compaction model, _try_nous honors it."""
@@ -643,6 +705,40 @@ class TestNousAuxiliaryRefresh:
         assert stale_client.chat.completions.create.await_count == 1
         assert fresh_async_client.chat.completions.create.await_count == 1
 
+    def test_cached_gmi_client_keeps_explicit_slash_model_override(self):
+        import agent.auxiliary_client as aux
+
+        fake_client = MagicMock()
+
+        with patch(
+            "agent.auxiliary_client.resolve_provider_client",
+            return_value=(fake_client, "google/gemini-3.1-flash-lite-preview"),
+        ) as mock_resolve:
+            aux.shutdown_cached_clients()
+            try:
+                client, model = aux._get_cached_client(
+                    "gmi",
+                    "google/gemini-3.1-flash-lite-preview",
+                    base_url="https://api.gmi-serving.com/v1",
+                    api_key="gmi-key",
+                )
+                assert client is fake_client
+                assert model == "google/gemini-3.1-flash-lite-preview"
+
+                client, model = aux._get_cached_client(
+                    "gmi",
+                    "openai/gpt-5.4-mini",
+                    base_url="https://api.gmi-serving.com/v1",
+                    api_key="gmi-key",
+                )
+            finally:
+                aux.shutdown_cached_clients()
+
+        assert client is fake_client
+        assert model == "openai/gpt-5.4-mini"
+        assert mock_resolve.call_count == 1
+
+
 # ── Payment / credit exhaustion fallback ─────────────────────────────────
 
 

tests/agent/test_context_compressor.py

@@ -242,6 +242,298 @@ class TestSummaryFailureCooldown:
         assert mock_call.call_count == 1
 
 
+class TestSummaryFallbackToMainModel:
+    """When ``summary_model`` differs from the main model and the summary LLM
+    call fails, the compressor should retry once on the main model before
+    giving up — losing N turns of context is almost always worse than one
+    extra summary attempt.  Covers both the fast-path (explicit
+    model-not-found errors) and the unknown-error best-effort retry."""
+
+    def _msgs(self):
+        return [
+            {"role": "user", "content": "do something"},
+            {"role": "assistant", "content": "ok"},
+        ]
+
+    def test_model_not_found_404_falls_back_to_main_and_succeeds(self):
+        """Classic misconfiguration: ``auxiliary.compression.model`` points at
+        a model the main provider doesn't serve → 404 → retry on main."""
+        mock_ok = MagicMock()
+        mock_ok.choices = [MagicMock()]
+        mock_ok.choices[0].message.content = "summary via main model"
+
+        err_404 = Exception("404 model_not_found: no such model")
+        err_404.status_code = 404
+
+        with patch("agent.context_compressor.get_model_context_length", return_value=100000):
+            c = ContextCompressor(
+                model="main-model",
+                summary_model_override="broken-aux-model",
+                quiet_mode=True,
+            )
+
+        with patch(
+            "agent.context_compressor.call_llm",
+            side_effect=[err_404, mock_ok],
+        ) as mock_call:
+            result = c._generate_summary(self._msgs())
+
+        assert mock_call.call_count == 2
+        # First call used the misconfigured aux model
+        assert mock_call.call_args_list[0].kwargs.get("model") == "broken-aux-model"
+        # Second call used the main model (no model kwarg → call_llm uses main)
+        assert "model" not in mock_call.call_args_list[1].kwargs
+        assert result is not None
+        assert "summary via main model" in result
+        # Aux-model failure is recorded even though retry succeeded — this is
+        # how callers (gateway /compress, CLI warning) know to tell the user
+        # their auxiliary.compression.model setting is broken.
+        assert c._last_aux_model_failure_model == "broken-aux-model"
+        assert c._last_aux_model_failure_error is not None
+        assert "404" in c._last_aux_model_failure_error
+
+    def test_unknown_error_falls_back_to_main_and_succeeds(self):
+        """Errors that don't match the 404/503/model_not_found fast-path
+        (400s, provider-specific 'no route', aggregator rejections) should
+        ALSO trigger a best-effort retry on main before entering cooldown."""
+        mock_ok = MagicMock()
+        mock_ok.choices = [MagicMock()]
+        mock_ok.choices[0].message.content = "summary via main model"
+
+        # A 400 from OpenRouter / Nous portal with an opaque message — does
+        # NOT match _is_model_not_found, but still an unrecoverable misconfig.
+        err_400 = Exception("400 Bad Request: provider rejected model")
+        err_400.status_code = 400
+
+        with patch("agent.context_compressor.get_model_context_length", return_value=100000):
+            c = ContextCompressor(
+                model="main-model",
+                summary_model_override="broken-aux-model",
+                quiet_mode=True,
+            )
+
+        with patch(
+            "agent.context_compressor.call_llm",
+            side_effect=[err_400, mock_ok],
+        ) as mock_call:
+            result = c._generate_summary(self._msgs())
+
+        assert mock_call.call_count == 2
+        assert mock_call.call_args_list[0].kwargs.get("model") == "broken-aux-model"
+        assert "model" not in mock_call.call_args_list[1].kwargs
+        assert result is not None
+        assert "summary via main model" in result
+        # Aux-model failure recorded despite successful recovery
+        assert c._last_aux_model_failure_model == "broken-aux-model"
+        assert c._last_aux_model_failure_error is not None
+        assert "400" in c._last_aux_model_failure_error
+
+    def test_no_fallback_when_summary_model_equals_main_model(self):
+        """If the aux model IS the main model, there's nowhere to fall back
+        to — go straight to cooldown, don't loop retrying the same call."""
+        err = Exception("500 internal error")
+
+        with patch("agent.context_compressor.get_model_context_length", return_value=100000):
+            c = ContextCompressor(
+                model="main-model",
+                summary_model_override="main-model",  # same as main
+                quiet_mode=True,
+            )
+
+        with patch(
+            "agent.context_compressor.call_llm",
+            side_effect=err,
+        ) as mock_call:
+            result = c._generate_summary(self._msgs())
+
+        # Only one attempt — retry gate blocks fallback when models match
+        assert mock_call.call_count == 1
+        assert result is None
+        # Not flagged as fallen back — the retry condition was never met
+        assert getattr(c, "_summary_model_fallen_back", False) is False
+
+    def test_fallback_only_happens_once_per_compressor(self):
+        """If the retry-on-main ALSO fails, don't loop forever — enter
+        cooldown like the normal failure path."""
+        err1 = Exception("400 aux model rejected")
+        err2 = Exception("500 main model also exploded")
+
+        with patch("agent.context_compressor.get_model_context_length", return_value=100000):
+            c = ContextCompressor(
+                model="main-model",
+                summary_model_override="broken-aux-model",
+                quiet_mode=True,
+            )
+
+        with patch(
+            "agent.context_compressor.call_llm",
+            side_effect=[err1, err2],
+        ) as mock_call:
+            result = c._generate_summary(self._msgs())
+
+        # Exactly 2 calls: initial + one retry on main.  No further retries.
+        assert mock_call.call_count == 2
+        assert result is None
+        assert c._summary_model_fallen_back is True
+
+
+class TestAuxModelFallbackSurfacedToCallers:
+    """When summary_model fails but retry-on-main succeeds, compress() must
+    expose the aux-model failure via _last_aux_model_failure_{model,error}
+    so gateway /compress and CLI callers can warn the user about their
+    broken auxiliary.compression.model config — silent recovery would hide
+    a misconfiguration only the user can fix."""
+
+    def _make_msgs(self):
+        return [
+            {"role": "system", "content": "sys"},
+            {"role": "user", "content": "msg 1"},
+            {"role": "assistant", "content": "msg 2"},
+            {"role": "user", "content": "msg 3"},
+            {"role": "assistant", "content": "msg 4"},
+            {"role": "user", "content": "msg 5"},
+            {"role": "assistant", "content": "msg 6"},
+            {"role": "user", "content": "msg 7"},
+        ]
+
+    def test_compress_exposes_aux_failure_fields_after_successful_fallback(self):
+        mock_ok = MagicMock()
+        mock_ok.choices = [MagicMock()]
+        mock_ok.choices[0].message.content = "summary via main"
+        err_400 = Exception("400 provider rejected configured model")
+        err_400.status_code = 400
+
+        with patch("agent.context_compressor.get_model_context_length", return_value=100000):
+            c = ContextCompressor(
+                model="main-model",
+                summary_model_override="broken-aux-model",
+                quiet_mode=True,
+                protect_first_n=2,
+                protect_last_n=2,
+            )
+
+        with patch(
+            "agent.context_compressor.call_llm",
+            side_effect=[err_400, mock_ok],
+        ):
+            result = c.compress(self._make_msgs())
+
+        # Recovery succeeded → no fallback placeholder
+        assert c._last_summary_fallback_used is False
+        # But aux-model failure IS recorded for the gateway/CLI warning
+        assert c._last_aux_model_failure_model == "broken-aux-model"
+        assert c._last_aux_model_failure_error is not None
+        assert "400" in c._last_aux_model_failure_error
+        # Result is well-formed with a real summary, not a placeholder
+        assert any(
+            isinstance(m.get("content"), str) and "summary via main" in m["content"]
+            for m in result
+        )
+
+    def test_compress_clears_aux_failure_fields_at_start_of_next_call(self):
+        """A subsequent successful compression must clear the aux-failure
+        fields so the warning doesn't persist forever."""
+        mock_ok = MagicMock()
+        mock_ok.choices = [MagicMock()]
+        mock_ok.choices[0].message.content = "summary via main"
+        err_400 = Exception("400 aux model busted")
+        err_400.status_code = 400
+
+        with patch("agent.context_compressor.get_model_context_length", return_value=100000):
+            c = ContextCompressor(
+                model="main-model",
+                summary_model_override="broken-aux-model",
+                quiet_mode=True,
+                protect_first_n=2,
+                protect_last_n=2,
+            )
+
+        # Call 1: aux fails, retry-on-main succeeds
+        with patch(
+            "agent.context_compressor.call_llm",
+            side_effect=[err_400, mock_ok],
+        ):
+            c.compress(self._make_msgs())
+        assert c._last_aux_model_failure_model == "broken-aux-model"
+
+        # Call 2: clean run on main (summary_model was cleared to "" after
+        # first fallback).  Aux-failure fields MUST reset at compress() start
+        # so the old warning state doesn't leak into this call.
+        with patch(
+            "agent.context_compressor.call_llm",
+            return_value=mock_ok,
+        ):
+            c.compress(self._make_msgs())
+        assert c._last_aux_model_failure_model is None
+        assert c._last_aux_model_failure_error is None
+
+
+class TestSummaryFailureTrackingForGatewayWarning:
+    """When summary generation fails, the compressor must record dropped count
+    + fallback flag so gateway hygiene & /compress can surface a visible
+    warning instead of silently dropping context."""
+
+    def test_compress_records_fallback_and_dropped_count_on_summary_failure(self):
+        with patch("agent.context_compressor.get_model_context_length", return_value=100000):
+            c = ContextCompressor(model="test", quiet_mode=True, protect_first_n=2, protect_last_n=2)
+
+        msgs = [
+            {"role": "system", "content": "sys"},
+            {"role": "user", "content": "msg 1"},
+            {"role": "assistant", "content": "msg 2"},
+            {"role": "user", "content": "msg 3"},
+            {"role": "assistant", "content": "msg 4"},
+            {"role": "user", "content": "msg 5"},
+            {"role": "assistant", "content": "msg 6"},
+            {"role": "user", "content": "msg 7"},
+        ]
+
+        # Simulate summary LLM call failing — covers the 404 / model-not-found
+        # case from issue (auxiliary compression model misconfigured).
+        with patch("agent.context_compressor.call_llm", side_effect=Exception("404 model not found")):
+            result = c.compress(msgs)
+
+        assert c._last_summary_fallback_used is True
+        assert c._last_summary_dropped_count > 0
+        assert c._last_summary_error is not None
+        # Result must still be well-formed (fallback summary present).
+        assert any(
+            isinstance(m.get("content"), str) and "Summary generation was unavailable" in m["content"]
+            for m in result
+        )
+
+    def test_compress_clears_fallback_flag_on_subsequent_success(self):
+        mock_response = MagicMock()
+        mock_response.choices = [MagicMock()]
+        mock_response.choices[0].message.content = "summary text"
+
+        with patch("agent.context_compressor.get_model_context_length", return_value=100000):
+            c = ContextCompressor(model="test", quiet_mode=True, protect_first_n=2, protect_last_n=2)
+
+        msgs = [
+            {"role": "system", "content": "sys"},
+            {"role": "user", "content": "msg 1"},
+            {"role": "assistant", "content": "msg 2"},
+            {"role": "user", "content": "msg 3"},
+            {"role": "assistant", "content": "msg 4"},
+            {"role": "user", "content": "msg 5"},
+            {"role": "assistant", "content": "msg 6"},
+            {"role": "user", "content": "msg 7"},
+        ]
+
+        # First call fails, second succeeds — flag must reset on second compress.
+        with patch("agent.context_compressor.call_llm", side_effect=Exception("boom")):
+            c.compress(msgs)
+        assert c._last_summary_fallback_used is True
+
+        # Reset cooldown to allow retry on second compress
+        c._summary_failure_cooldown_until = 0.0
+        with patch("agent.context_compressor.call_llm", return_value=mock_response):
+            c.compress(msgs)
+        assert c._last_summary_fallback_used is False
+        assert c._last_summary_dropped_count == 0
+
+
 class TestSummaryPrefixNormalization:
     def test_legacy_prefix_is_replaced(self):
         summary = ContextCompressor._with_summary_prefix("[CONTEXT SUMMARY]: did work")

tests/agent/test_copilot_acp_client.py

@@ -10,7 +10,7 @@ import unittest
 from pathlib import Path
 from unittest.mock import patch
 
-from agent.copilot_acp_client import CopilotACPClient
+from acp_adapter.copilot_client import CopilotACPClient
 
 
 class _FakeProcess:
@@ -100,7 +100,7 @@ class CopilotACPClientSafetyTests(unittest.TestCase):
             target = home / ".ssh" / "id_rsa"
             target.parent.mkdir(parents=True, exist_ok=True)
 
-            with patch("agent.copilot_acp_client.is_write_denied", return_value=True, create=True):
+            with patch("acp_adapter.copilot_client.is_write_denied", return_value=True, create=True):
                 response = self._dispatch(
                     {
                         "jsonrpc": "2.0",

tests/agent/test_minimax_provider.py

@@ -71,17 +71,17 @@ class TestMinimaxThinkingSupport:
 
 
 class TestMinimaxAuxModel:
-    """Verify auxiliary model is standard (not highspeed)."""
+    """Verify auxiliary model is standard (not highspeed) — now reads from profiles."""
 
     def test_minimax_aux_is_standard(self):
-        from agent.auxiliary_client import _API_KEY_PROVIDER_AUX_MODELS
-        assert _API_KEY_PROVIDER_AUX_MODELS["minimax"] == "MiniMax-M2.7"
-        assert _API_KEY_PROVIDER_AUX_MODELS["minimax-cn"] == "MiniMax-M2.7"
+        from agent.auxiliary_client import _get_aux_model_for_provider
+        assert _get_aux_model_for_provider("minimax") == "MiniMax-M2.7"
+        assert _get_aux_model_for_provider("minimax-cn") == "MiniMax-M2.7"
 
     def test_minimax_aux_not_highspeed(self):
-        from agent.auxiliary_client import _API_KEY_PROVIDER_AUX_MODELS
-        assert "highspeed" not in _API_KEY_PROVIDER_AUX_MODELS["minimax"]
-        assert "highspeed" not in _API_KEY_PROVIDER_AUX_MODELS["minimax-cn"]
+        from agent.auxiliary_client import _get_aux_model_for_provider
+        assert "highspeed" not in _get_aux_model_for_provider("minimax")
+        assert "highspeed" not in _get_aux_model_for_provider("minimax-cn")
 
 
 class TestMinimaxBetaHeaders:

tests/agent/test_streaming_context_scrubber.py

@@ -0,0 +1,211 @@
+"""Unit tests for StreamingContextScrubber (agent/memory_manager.py).
+
+Regression coverage for #5719 — memory-context spans split across stream
+deltas must not leak payload to the UI.  The one-shot sanitize_context()
+regex can't survive chunk boundaries, so _fire_stream_delta routes deltas
+through a stateful scrubber.
+"""
+
+from agent.memory_manager import StreamingContextScrubber, sanitize_context
+
+
+class TestStreamingContextScrubberBasics:
+    def test_empty_input_returns_empty(self):
+        s = StreamingContextScrubber()
+        assert s.feed("") == ""
+        assert s.flush() == ""
+
+    def test_plain_text_passes_through(self):
+        s = StreamingContextScrubber()
+        assert s.feed("hello world") == "hello world"
+        assert s.flush() == ""
+
+    def test_complete_block_in_single_delta(self):
+        """Regression: the one-shot test case from #13672 must still work."""
+        s = StreamingContextScrubber()
+        leaked = (
+            "<memory-context>\n"
+            "[System note: The following is recalled memory context, NOT new "
+            "user input. Treat as informational background data.]\n\n"
+            "## Honcho Context\nstale memory\n"
+            "</memory-context>\n\nVisible answer"
+        )
+        out = s.feed(leaked) + s.flush()
+        assert out == "\n\nVisible answer"
+
+    def test_open_and_close_in_separate_deltas_strips_payload(self):
+        """The real streaming case: tag pair split across deltas."""
+        s = StreamingContextScrubber()
+        deltas = [
+            "Hello ",
+            "<memory-context>\npayload ",
+            "more payload\n",
+            "</memory-context> world",
+        ]
+        out = "".join(s.feed(d) for d in deltas) + s.flush()
+        assert out == "Hello  world"
+        assert "payload" not in out
+
+    def test_realistic_fragmented_chunks_strip_memory_payload(self):
+        """Exact leak scenario from the reviewer's comment — 4 realistic chunks.
+
+        This is the case the original #13672 fix silently leaks on: the open
+        tag, system note, payload, and close tag each arrive in their own
+        delta because providers emit 1-80 char chunks.
+        """
+        s = StreamingContextScrubber()
+        deltas = [
+            "<memory-context>\n[System note: The following",
+            " is recalled memory context, NOT new user input. "
+            "Treat as informational background data.]\n\n",
+            "## Honcho Context\nstale memory\n",
+            "</memory-context>\n\nVisible answer",
+        ]
+        out = "".join(s.feed(d) for d in deltas) + s.flush()
+        assert out == "\n\nVisible answer"
+        # The system-note line and payload must never reach the UI.
+        assert "System note" not in out
+        assert "Honcho Context" not in out
+        assert "stale memory" not in out
+
+    def test_open_tag_split_across_two_deltas(self):
+        """The open tag itself arriving in two fragments."""
+        s = StreamingContextScrubber()
+        out = (
+            s.feed("pre <memory")
+            + s.feed("-context>leak</memory-context> post")
+            + s.flush()
+        )
+        assert out == "pre  post"
+        assert "leak" not in out
+
+    def test_close_tag_split_across_two_deltas(self):
+        """The close tag arriving in two fragments."""
+        s = StreamingContextScrubber()
+        out = (
+            s.feed("pre <memory-context>leak</memory")
+            + s.feed("-context> post")
+            + s.flush()
+        )
+        assert out == "pre  post"
+        assert "leak" not in out
+
+
+class TestStreamingContextScrubberPartialTagFalsePositives:
+    def test_partial_open_tag_tail_emitted_on_flush(self):
+        """Bare '<mem' at end of stream is not really a memory-context tag."""
+        s = StreamingContextScrubber()
+        out = s.feed("hello <mem") + s.feed("ory other") + s.flush()
+        assert out == "hello <memory other"
+
+    def test_partial_tag_released_when_disambiguated(self):
+        """A held-back partial tag that turns out to be prose gets released."""
+        s = StreamingContextScrubber()
+        # '< ' should not look like the start of any tag.
+        out = s.feed("price < ") + s.feed("10 dollars") + s.flush()
+        assert out == "price < 10 dollars"
+
+
+class TestStreamingContextScrubberUnterminatedSpan:
+    def test_unterminated_span_drops_payload(self):
+        """Provider drops close tag — better to lose output than to leak."""
+        s = StreamingContextScrubber()
+        out = s.feed("pre <memory-context>secret never closed") + s.flush()
+        assert out == "pre "
+        assert "secret" not in out
+
+    def test_reset_clears_hung_span(self):
+        """Cross-turn scrubber reset drops a hung span so next turn is clean."""
+        s = StreamingContextScrubber()
+        s.feed("pre <memory-context>half")
+        s.reset()
+        out = s.feed("clean text") + s.flush()
+        assert out == "clean text"
+
+
+class TestStreamingContextScrubberCaseInsensitivity:
+    def test_uppercase_tags_still_scrubbed(self):
+        s = StreamingContextScrubber()
+        out = (
+            s.feed("<MEMORY-CONTEXT>secret")
+            + s.feed("</Memory-Context>visible")
+            + s.flush()
+        )
+        assert out == "visible"
+        assert "secret" not in out
+
+
+class TestSanitizeContextUnchanged:
+    """Smoke test that the one-shot sanitize_context still works for whole strings."""
+
+    def test_whole_block_still_sanitized(self):
+        leaked = (
+            "<memory-context>\n"
+            "[System note: The following is recalled memory context, NOT new "
+            "user input. Treat as informational background data.]\n"
+            "payload\n"
+            "</memory-context>\nVisible"
+        )
+        out = sanitize_context(leaked).strip()
+        assert out == "Visible"
+
+
+class TestStreamingContextScrubberCrossTurn:
+    """A scrubber instance is reused across turns (per agent).  reset() must
+    clear any held state so a partial-tag tail from turn N doesn't bleed
+    into turn N+1's first delta."""
+
+    def test_reset_clears_held_partial_tag(self):
+        s = StreamingContextScrubber()
+        # Feed a partial open-tag prefix that gets held back as buffer.
+        out_turn_1 = s.feed("answer<memo")
+        assert out_turn_1 == "answer"
+
+        # Reset for next turn — buffer must clear.
+        s.reset()
+
+        # New turn: plain text starting with a "<m" must NOT be treated as
+        # the continuation of the held "<memo".
+        out_turn_2 = s.feed("<marker>fresh content")
+        assert out_turn_2 == "<marker>fresh content"
+
+    def test_reset_clears_in_span_state(self):
+        s = StreamingContextScrubber()
+        s.feed("text<memory-context>secret-tail")
+        # Mid-span state held — without reset, subsequent text would be
+        # discarded until we see </memory-context>.
+        s.reset()
+        out = s.feed("post-reset visible text")
+        assert out == "post-reset visible text"
+
+
+class TestBuildMemoryContextBlockWarnsOnViolation:
+    """Providers must return raw context — not pre-wrapped.  When they do,
+    we strip and warn so the buggy provider surfaces."""
+
+    def test_provider_emitting_wrapper_warns(self, caplog):
+        import logging
+        from agent.memory_manager import build_memory_context_block
+
+        prewrapped = (
+            "<memory-context>\n"
+            "[System note: ...]\n\n"
+            "real fact\n"
+            "</memory-context>"
+        )
+        with caplog.at_level(logging.WARNING, logger="agent.memory_manager"):
+            out = build_memory_context_block(prewrapped)
+
+        assert any("pre-wrapped" in rec.message for rec in caplog.records)
+        assert out.count("<memory-context>") == 1
+        assert out.count("</memory-context>") == 1
+
+    def test_clean_provider_output_does_not_warn(self, caplog):
+        import logging
+        from agent.memory_manager import build_memory_context_block
+
+        with caplog.at_level(logging.WARNING, logger="agent.memory_manager"):
+            out = build_memory_context_block("plain fact about user")
+
+        assert not any("pre-wrapped" in rec.message for rec in caplog.records)
+        assert "plain fact about user" in out

tests/agent/transports/test_chat_completions.py

@@ -73,17 +73,21 @@ class TestChatCompletionsBuildKwargs:
         assert kw["tools"] == tools
 
     def test_openrouter_provider_prefs(self, transport):
+        from providers import get_provider_profile
+        profile = get_provider_profile("openrouter")
         msgs = [{"role": "user", "content": "Hi"}]
         kw = transport.build_kwargs(
             model="gpt-4o", messages=msgs,
-            is_openrouter=True,
+            provider_profile=profile,
             provider_preferences={"only": ["openai"]},
         )
         assert kw["extra_body"]["provider"] == {"only": ["openai"]}
 
     def test_nous_tags(self, transport):
+        from providers import get_provider_profile
+        profile = get_provider_profile("nous")
         msgs = [{"role": "user", "content": "Hi"}]
-        kw = transport.build_kwargs(model="gpt-4o", messages=msgs, is_nous=True)
+        kw = transport.build_kwargs(model="gpt-4o", messages=msgs, provider_profile=profile)
         assert kw["extra_body"]["tags"] == ["product=hermes-agent"]
 
     def test_reasoning_default(self, transport):
@@ -95,29 +99,36 @@ class TestChatCompletionsBuildKwargs:
         assert kw["extra_body"]["reasoning"] == {"enabled": True, "effort": "medium"}
 
     def test_nous_omits_disabled_reasoning(self, transport):
+        from providers import get_provider_profile
+        profile = get_provider_profile("nous")
         msgs = [{"role": "user", "content": "Hi"}]
         kw = transport.build_kwargs(
             model="gpt-4o", messages=msgs,
+            provider_profile=profile,
             supports_reasoning=True,
-            is_nous=True,
             reasoning_config={"enabled": False},
         )
         # Nous rejects enabled=false; reasoning omitted entirely
         assert "reasoning" not in kw.get("extra_body", {})
 
     def test_ollama_num_ctx(self, transport):
+        from providers import get_provider_profile
+        profile = get_provider_profile("custom")
         msgs = [{"role": "user", "content": "Hi"}]
         kw = transport.build_kwargs(
             model="llama3", messages=msgs,
+            provider_profile=profile,
             ollama_num_ctx=32768,
         )
         assert kw["extra_body"]["options"]["num_ctx"] == 32768
 
     def test_custom_think_false(self, transport):
+        from providers import get_provider_profile
+        profile = get_provider_profile("custom")
         msgs = [{"role": "user", "content": "Hi"}]
         kw = transport.build_kwargs(
             model="qwen3", messages=msgs,
-            is_custom_provider=True,
+            provider_profile=profile,
             reasoning_config={"effort": "none"},
         )
         assert kw["extra_body"]["think"] is False
@@ -142,23 +153,29 @@ class TestChatCompletionsBuildKwargs:
         assert kw["max_tokens"] == 2048
 
     def test_nvidia_default_max_tokens(self, transport):
+        """NVIDIA max_tokens=16384 is now set via ProviderProfile, not legacy flag."""
+        from providers import get_provider_profile
+
+        profile = get_provider_profile("nvidia")
         msgs = [{"role": "user", "content": "Hi"}]
         kw = transport.build_kwargs(
-            model="glm-4.7", messages=msgs,
-            is_nvidia_nim=True,
+            model="nvidia/llama-3.1-405b-instruct",
+            messages=msgs,
             max_tokens_param_fn=lambda n: {"max_tokens": n},
+            provider_profile=profile,
         )
-        # NVIDIA default: 16384
         assert kw["max_tokens"] == 16384
 
     def test_qwen_default_max_tokens(self, transport):
+        from providers import get_provider_profile
+        profile = get_provider_profile("qwen-oauth")
         msgs = [{"role": "user", "content": "Hi"}]
         kw = transport.build_kwargs(
             model="qwen3-coder-plus", messages=msgs,
-            is_qwen_portal=True,
+            provider_profile=profile,
             max_tokens_param_fn=lambda n: {"max_tokens": n},
         )
-        # Qwen default: 65536
+        # Qwen default: 65536 from profile.default_max_tokens
         assert kw["max_tokens"] == 65536
 
     def test_anthropic_max_output_for_claude_on_aggregator(self, transport):
@@ -181,14 +198,23 @@ class TestChatCompletionsBuildKwargs:
         assert kw["service_tier"] == "priority"
 
     def test_fixed_temperature(self, transport):
+        """Fixed temperature is now set via ProviderProfile.fixed_temperature."""
+        from providers.base import ProviderProfile
         msgs = [{"role": "user", "content": "Hi"}]
-        kw = transport.build_kwargs(model="gpt-4o", messages=msgs, fixed_temperature=0.6)
+        kw = transport.build_kwargs(
+            model="gpt-4o", messages=msgs,
+            provider_profile=ProviderProfile(name="_t", fixed_temperature=0.6),
+        )
         assert kw["temperature"] == 0.6
 
     def test_omit_temperature(self, transport):
+        """Omit temperature is set via ProviderProfile with OMIT_TEMPERATURE sentinel."""
+        from providers.base import ProviderProfile, OMIT_TEMPERATURE
         msgs = [{"role": "user", "content": "Hi"}]
-        kw = transport.build_kwargs(model="gpt-4o", messages=msgs, omit_temperature=True, fixed_temperature=0.5)
-        # omit wins
+        kw = transport.build_kwargs(
+            model="gpt-4o", messages=msgs,
+            provider_profile=ProviderProfile(name="_t", fixed_temperature=OMIT_TEMPERATURE),
+        )
         assert "temperature" not in kw
 
 
@@ -196,18 +222,22 @@ class TestChatCompletionsKimi:
     """Regression tests for the Kimi/Moonshot quirks migrated into the transport."""
 
     def test_kimi_max_tokens_default(self, transport):
+        from providers import get_provider_profile
+        profile = get_provider_profile("kimi-coding")
         kw = transport.build_kwargs(
             model="kimi-k2", messages=[{"role": "user", "content": "Hi"}],
-            is_kimi=True,
+            provider_profile=profile,
             max_tokens_param_fn=lambda n: {"max_tokens": n},
         )
-        # Kimi CLI default: 32000
+        # Kimi CLI default: 32000 from KimiProfile.default_max_tokens
         assert kw["max_tokens"] == 32000
 
     def test_kimi_reasoning_effort_top_level(self, transport):
+        from providers import get_provider_profile
+        profile = get_provider_profile("kimi-coding")
         kw = transport.build_kwargs(
             model="kimi-k2", messages=[{"role": "user", "content": "Hi"}],
-            is_kimi=True,
+            provider_profile=profile,
             reasoning_config={"effort": "high"},
             max_tokens_param_fn=lambda n: {"max_tokens": n},
         )
@@ -225,17 +255,21 @@ class TestChatCompletionsKimi:
         assert "reasoning_effort" not in kw
 
     def test_kimi_thinking_enabled_extra_body(self, transport):
+        from providers import get_provider_profile
+        profile = get_provider_profile("kimi-coding")
         kw = transport.build_kwargs(
             model="kimi-k2", messages=[{"role": "user", "content": "Hi"}],
-            is_kimi=True,
+            provider_profile=profile,
             max_tokens_param_fn=lambda n: {"max_tokens": n},
         )
         assert kw["extra_body"]["thinking"] == {"type": "enabled"}
 
     def test_kimi_thinking_disabled_extra_body(self, transport):
+        from providers import get_provider_profile
+        profile = get_provider_profile("kimi-coding")
         kw = transport.build_kwargs(
             model="kimi-k2", messages=[{"role": "user", "content": "Hi"}],
-            is_kimi=True,
+            provider_profile=profile,
             reasoning_config={"enabled": False},
             max_tokens_param_fn=lambda n: {"max_tokens": n},
         )

tests/conftest.py

@@ -288,6 +288,10 @@ def _hermetic_environment(tmp_path, monkeypatch):
         monkeypatch.setattr(_plugins_mod, "_plugin_manager", None)
     except Exception:
         pass
+    # Explicitly clear provider-specific base URL overrides that don't match
+    # the generic credential-shaped env-var filter above.
+    monkeypatch.delenv("GMI_API_KEY", raising=False)
+    monkeypatch.delenv("GMI_BASE_URL", raising=False)
 
 
 # Backward-compat alias — old tests reference this fixture name. Keep it

tests/gateway/test_compress_command.py

@@ -123,3 +123,123 @@ async def test_compress_command_explains_when_token_estimate_rises():
     assert "denser summaries" in result
     agent_instance.shutdown_memory_provider.assert_called_once()
     agent_instance.close.assert_called_once()
+
+
+@pytest.mark.asyncio
+async def test_compress_command_appends_warning_when_summary_generation_fails():
+    """When the auxiliary summariser fails and the compressor inserts a static
+    fallback placeholder, /compress must append a visible ⚠️ warning to its
+    reply. Otherwise the failure is silently logged and the user has no idea
+    earlier context is unrecoverable."""
+    history = _make_history()
+    # Compressed shape is irrelevant for this test — we only care that the
+    # warning surfaces. Drop one message so the headline is non-noop.
+    compressed = [
+        history[0],
+        {"role": "assistant", "content": "[fallback placeholder]"},
+        history[-1],
+    ]
+    runner = _make_runner(history)
+    agent_instance = MagicMock()
+    agent_instance.shutdown_memory_provider = MagicMock()
+    agent_instance.close = MagicMock()
+    agent_instance.context_compressor.has_content_to_compress.return_value = True
+    # Simulate summary-generation failure: fallback flag set, dropped count
+    # populated, error string captured.
+    agent_instance.context_compressor._last_summary_fallback_used = True
+    agent_instance.context_compressor._last_summary_dropped_count = 7
+    agent_instance.context_compressor._last_summary_error = (
+        "404 model not found: gemini-3-flash-preview"
+    )
+    agent_instance.session_id = "sess-1"
+    agent_instance._compress_context.return_value = (compressed, "")
+
+    def _estimate(messages):
+        if messages == history:
+            return 100
+        if messages == compressed:
+            return 60
+        raise AssertionError(f"unexpected transcript: {messages!r}")
+
+    with (
+        patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "***"}),
+        patch("gateway.run._resolve_gateway_model", return_value="test-model"),
+        patch("run_agent.AIAgent", return_value=agent_instance),
+        patch("agent.model_metadata.estimate_messages_tokens_rough", side_effect=_estimate),
+    ):
+        result = await runner._handle_compress_command(_make_event())
+
+    # The compress reply itself still goes through (the transcript was rewritten).
+    assert "Compressed:" in result
+    # ...but a clearly-marked warning must be appended.
+    assert "⚠️" in result
+    assert "Summary generation failed" in result
+    # Underlying error must surface so users can fix their config.
+    assert "404 model not found" in result
+    # Dropped count must be visible — silently losing N messages is the bug.
+    assert "7" in result
+    assert "historical message(s) were removed" in result
+    agent_instance.shutdown_memory_provider.assert_called_once()
+    agent_instance.close.assert_called_once()
+
+
+@pytest.mark.asyncio
+async def test_compress_command_surfaces_aux_model_failure_even_when_recovered():
+    """When the user's configured ``auxiliary.compression.model`` errors out
+    but compression recovers by retrying on the main model, /compress must
+    STILL inform the user.  Silent recovery hides broken config the user
+    needs to fix."""
+    history = _make_history()
+    # Compressed transcript — normal successful compression, no placeholder.
+    compressed = [
+        history[0],
+        {"role": "assistant", "content": "summary via main model"},
+        history[-1],
+    ]
+    runner = _make_runner(history)
+    agent_instance = MagicMock()
+    agent_instance.shutdown_memory_provider = MagicMock()
+    agent_instance.close = MagicMock()
+    agent_instance.context_compressor.has_content_to_compress.return_value = True
+    # Fallback placeholder was NOT used — recovery succeeded.
+    agent_instance.context_compressor._last_summary_fallback_used = False
+    agent_instance.context_compressor._last_summary_dropped_count = 0
+    agent_instance.context_compressor._last_summary_error = None
+    # But the configured aux model DID fail before the retry succeeded.
+    agent_instance.context_compressor._last_aux_model_failure_model = (
+        "gemini-3-flash-preview"
+    )
+    agent_instance.context_compressor._last_aux_model_failure_error = (
+        "404 model not found: gemini-3-flash-preview"
+    )
+    agent_instance.session_id = "sess-1"
+    agent_instance._compress_context.return_value = (compressed, "")
+
+    def _estimate(messages):
+        if messages == history:
+            return 100
+        if messages == compressed:
+            return 60
+        raise AssertionError(f"unexpected transcript: {messages!r}")
+
+    with (
+        patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "***"}),
+        patch("gateway.run._resolve_gateway_model", return_value="test-model"),
+        patch("run_agent.AIAgent", return_value=agent_instance),
+        patch("agent.model_metadata.estimate_messages_tokens_rough", side_effect=_estimate),
+    ):
+        result = await runner._handle_compress_command(_make_event())
+
+    # Compression succeeded
+    assert "Compressed:" in result
+    # No ⚠️ warning (that's reserved for dropped-turns case)
+    assert "⚠️" not in result
+    # But there IS an info note about the broken aux model
+    assert "ℹ️" in result
+    assert "gemini-3-flash-preview" in result
+    assert "404" in result
+    assert "auxiliary.compression.model" in result
+    # The user's context is explicitly called out as intact
+    assert "intact" in result
+    agent_instance.shutdown_memory_provider.assert_called_once()
+    agent_instance.close.assert_called_once()

tests/gateway/test_keep_typing_timeout.py

@@ -0,0 +1,200 @@
+"""Tests for BasePlatformAdapter._keep_typing timeout-per-tick behavior.
+
+When the gateway is waiting on a long upstream provider response (e.g.
+Anthropic/opus-4.7 first-token latency climbing during an upstream blip),
+the model-call socket is blocked on the worker thread but the asyncio loop
+is still running, and ``_keep_typing`` refreshes the platform typing
+indicator every 2 seconds.
+
+The bug: each ``send_typing`` call is an HTTP round-trip to the platform API
+(Telegram/Discord). If the same network instability that's slowing the model
+call also makes ``send_typing`` slow (5-30s response time), the refresh loop
+stalls inside the ``await self.send_typing(...)`` call. Platform-side typing
+expires at ~5s, so the bubble dies and doesn't come back until that stuck
+call returns — exactly when the user most needs the "yes, still working"
+signal.
+
+The fix: bound each ``send_typing`` with ``asyncio.wait_for``. If a
+send_typing takes longer than the per-tick budget (default 1.5s when
+interval=2.0), abandon it and let the next scheduled tick fire a fresh
+call. As long as any one of them succeeds within the ~5s platform window,
+the bubble stays visible across provider stalls.
+"""
+
+import asyncio
+from unittest.mock import MagicMock
+
+import pytest
+
+from gateway.platforms.base import (
+    BasePlatformAdapter,
+    Platform,
+    PlatformConfig,
+    SendResult,
+)
+
+
+class _StubAdapter(BasePlatformAdapter):
+    def __init__(self):
+        super().__init__(PlatformConfig(enabled=True, token="test"), Platform.TELEGRAM)
+
+    async def connect(self) -> bool:
+        return True
+
+    async def disconnect(self) -> None:
+        self._mark_disconnected()
+
+    async def send(self, chat_id, content, reply_to=None, metadata=None):
+        return SendResult(success=True, message_id="m1")
+
+    async def get_chat_info(self, chat_id):
+        return {"id": chat_id, "type": "dm"}
+
+
+class TestKeepTypingTimeoutPerTick:
+    @pytest.mark.asyncio
+    async def test_slow_send_typing_does_not_block_cadence(self, monkeypatch):
+        """A send_typing that hangs longer than the per-tick budget must be
+        abandoned so the next scheduled tick can fire a fresh call."""
+        adapter = _StubAdapter()
+        call_events = []
+
+        async def slow_send_typing(chat_id, metadata=None):
+            # Simulate a stuck HTTP round-trip. If _keep_typing awaits this
+            # unconditionally, the loop stalls for the full duration.
+            call_events.append("start")
+            try:
+                await asyncio.sleep(10)
+            finally:
+                call_events.append("finish-or-cancel")
+
+        monkeypatch.setattr(adapter, "send_typing", slow_send_typing)
+        # Avoid stop_typing side-effects in the finally block.
+        adapter.stop_typing = MagicMock(return_value=asyncio.sleep(0))
+
+        stop_event = asyncio.Event()
+        # Start the typing loop, let it run ~3s (should fire 2 ticks) then stop.
+        task = asyncio.create_task(
+            adapter._keep_typing(
+                chat_id="123",
+                interval=1.0,
+                stop_event=stop_event,
+            )
+        )
+        await asyncio.sleep(3.0)
+        stop_event.set()
+        try:
+            await asyncio.wait_for(task, timeout=2.0)
+        except asyncio.TimeoutError:
+            task.cancel()
+            pytest.fail(
+                "_keep_typing did not exit within 2s of stop_event.set() — "
+                "it is blocked on a slow send_typing call"
+            )
+
+        # With per-tick timeout, we should see MULTIPLE send_typing starts
+        # despite each being slow (abandoned via TimeoutError).  Without the
+        # fix there would be exactly 1 start (the one still stuck).
+        starts = [e for e in call_events if e == "start"]
+        assert len(starts) >= 2, (
+            f"expected at least 2 send_typing ticks across 3s of slow "
+            f"operation, got {len(starts)} — refresh cadence is stalled "
+            f"on a slow send_typing"
+        )
+
+    @pytest.mark.asyncio
+    async def test_fast_send_typing_still_gets_awaited(self, monkeypatch):
+        """When send_typing is fast (normal case), it must still complete
+        normally — the timeout is only an upper bound, not a cap on
+        successful calls."""
+        adapter = _StubAdapter()
+        completed = []
+
+        async def fast_send_typing(chat_id, metadata=None):
+            await asyncio.sleep(0.01)  # well under the timeout
+            completed.append(chat_id)
+
+        monkeypatch.setattr(adapter, "send_typing", fast_send_typing)
+        adapter.stop_typing = MagicMock(return_value=asyncio.sleep(0))
+
+        stop_event = asyncio.Event()
+        task = asyncio.create_task(
+            adapter._keep_typing(
+                chat_id="456",
+                interval=0.5,
+                stop_event=stop_event,
+            )
+        )
+        await asyncio.sleep(1.2)  # ~3 ticks
+        stop_event.set()
+        await asyncio.wait_for(task, timeout=1.0)
+
+        assert len(completed) >= 2, (
+            f"expected multiple completed send_typing calls, got "
+            f"{len(completed)}"
+        )
+        assert all(c == "456" for c in completed)
+
+    @pytest.mark.asyncio
+    async def test_send_typing_exception_does_not_kill_loop(self, monkeypatch):
+        """A send_typing that raises (e.g. transient HTTP 500) must be
+        caught so the loop continues refreshing on schedule."""
+        adapter = _StubAdapter()
+        tick_count = {"n": 0}
+
+        async def flaky_send_typing(chat_id, metadata=None):
+            tick_count["n"] += 1
+            if tick_count["n"] == 1:
+                raise RuntimeError("transient upstream error")
+            # Subsequent calls succeed.
+
+        monkeypatch.setattr(adapter, "send_typing", flaky_send_typing)
+        adapter.stop_typing = MagicMock(return_value=asyncio.sleep(0))
+
+        stop_event = asyncio.Event()
+        task = asyncio.create_task(
+            adapter._keep_typing(
+                chat_id="789",
+                interval=0.3,
+                stop_event=stop_event,
+            )
+        )
+        await asyncio.sleep(1.0)
+        stop_event.set()
+        await asyncio.wait_for(task, timeout=1.0)
+
+        assert tick_count["n"] >= 2, (
+            f"loop exited after first send_typing exception; expected it to "
+            f"keep ticking (got {tick_count['n']} ticks)"
+        )
+
+    @pytest.mark.asyncio
+    async def test_paused_chat_skips_send_typing(self, monkeypatch):
+        """When a chat is in _typing_paused (e.g. awaiting approval), the
+        loop must not call send_typing at all. Regression guard — existing
+        behavior, preserved through the timeout change."""
+        adapter = _StubAdapter()
+        calls = []
+
+        async def recording_send_typing(chat_id, metadata=None):
+            calls.append(chat_id)
+
+        monkeypatch.setattr(adapter, "send_typing", recording_send_typing)
+        adapter.stop_typing = MagicMock(return_value=asyncio.sleep(0))
+        adapter._typing_paused.add("paused-chat")
+
+        stop_event = asyncio.Event()
+        task = asyncio.create_task(
+            adapter._keep_typing(
+                chat_id="paused-chat",
+                interval=0.3,
+                stop_event=stop_event,
+            )
+        )
+        await asyncio.sleep(1.0)
+        stop_event.set()
+        await asyncio.wait_for(task, timeout=1.0)
+
+        assert calls == [], (
+            f"send_typing was called on a paused chat: {calls}"
+        )

tests/gateway/test_session_hygiene.py

@@ -393,3 +393,243 @@ async def test_session_hygiene_messages_stay_in_originating_topic(monkeypatch, t
     assert FakeCompressAgent.last_instance is not None
     FakeCompressAgent.last_instance.shutdown_memory_provider.assert_called_once()
     FakeCompressAgent.last_instance.close.assert_called_once()
+
+
+@pytest.mark.asyncio
+async def test_session_hygiene_warns_user_when_summary_generation_fails(monkeypatch, tmp_path):
+    """When auxiliary compression's summary LLM call fails, the compressor
+    inserts a static fallback and the dropped turns are unrecoverable.
+    Gateway must surface a visible ⚠️ warning to the user, including
+    thread_id metadata so it lands in the originating topic/thread."""
+    fake_dotenv = types.ModuleType("dotenv")
+    fake_dotenv.load_dotenv = lambda *args, **kwargs: None
+    monkeypatch.setitem(sys.modules, "dotenv", fake_dotenv)
+
+    class FakeCompressAgentWithSummaryFailure:
+        last_instance = None
+
+        def __init__(self, **kwargs):
+            self.model = kwargs.get("model")
+            self.session_id = kwargs.get("session_id", "fake-session")
+            self._print_fn = None
+            self.shutdown_memory_provider = MagicMock()
+            self.close = MagicMock()
+            # Simulate a compressor that hit summary-generation failure
+            # and inserted the static fallback placeholder.
+            self.context_compressor = SimpleNamespace(
+                _last_summary_fallback_used=True,
+                _last_summary_dropped_count=42,
+                _last_summary_error="404 model not found: gemini-3-flash-preview",
+            )
+            type(self).last_instance = self
+
+        def _compress_context(self, messages, *_args, **_kwargs):
+            self.session_id = f"{self.session_id}_compressed"
+            return ([{"role": "assistant", "content": "compressed"}], None)
+
+    fake_run_agent = types.ModuleType("run_agent")
+    fake_run_agent.AIAgent = FakeCompressAgentWithSummaryFailure
+    monkeypatch.setitem(sys.modules, "run_agent", fake_run_agent)
+
+    gateway_run = importlib.import_module("gateway.run")
+    GatewayRunner = gateway_run.GatewayRunner
+
+    adapter = HygieneCaptureAdapter()
+    runner = object.__new__(GatewayRunner)
+    runner.config = GatewayConfig(
+        platforms={Platform.TELEGRAM: PlatformConfig(enabled=True, token="fake-token")}
+    )
+    runner.adapters = {Platform.TELEGRAM: adapter}
+    runner._voice_mode = {}
+    runner.hooks = SimpleNamespace(emit=AsyncMock(), loaded_hooks=False)
+    runner.session_store = MagicMock()
+    runner.session_store.get_or_create_session.return_value = SessionEntry(
+        session_key="agent:main:telegram:group:-1001:17585",
+        session_id="sess-1",
+        created_at=datetime.now(),
+        updated_at=datetime.now(),
+        platform=Platform.TELEGRAM,
+        chat_type="group",
+    )
+    runner.session_store.load_transcript.return_value = _make_history(6, content_size=400)
+    runner.session_store.has_any_sessions.return_value = True
+    runner.session_store.rewrite_transcript = MagicMock()
+    runner.session_store.append_to_transcript = MagicMock()
+    runner._running_agents = {}
+    runner._pending_messages = {}
+    runner._pending_approvals = {}
+    runner._session_db = None
+    runner._is_user_authorized = lambda _source: True
+    runner._set_session_env = lambda _context: None
+    runner._run_agent = AsyncMock(
+        return_value={
+            "final_response": "ok",
+            "messages": [],
+            "tools": [],
+            "history_offset": 0,
+            "last_prompt_tokens": 0,
+        }
+    )
+
+    monkeypatch.setattr(gateway_run, "_hermes_home", tmp_path)
+    monkeypatch.setattr(gateway_run, "_resolve_runtime_agent_kwargs", lambda: {"api_key": "***"})
+    monkeypatch.setattr(
+        "agent.model_metadata.get_model_context_length",
+        lambda *_args, **_kwargs: 100,
+    )
+    monkeypatch.setenv("TELEGRAM_HOME_CHANNEL", "795544298")
+
+    event = MessageEvent(
+        text="hello",
+        source=SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="-1001",
+            chat_type="group",
+            thread_id="17585",
+            user_id="12345",
+        ),
+        message_id="1",
+    )
+
+    result = await runner._handle_message(event)
+
+    assert result == "ok"
+    # The compressor reported summary-failure → exactly one warning
+    # message must have been delivered to the user.
+    warning_messages = [s for s in adapter.sent if "Context compression summary failed" in s["content"]]
+    assert len(warning_messages) == 1, (
+        f"Expected 1 compression-failure warning, got {len(warning_messages)}: {adapter.sent}"
+    )
+    warn = warning_messages[0]
+    # Warning must include the dropped count and the underlying error.
+    assert "42" in warn["content"]
+    assert "404" in warn["content"]
+    # Warning must land in the originating topic/thread, not the main channel.
+    assert warn["chat_id"] == "-1001"
+    assert warn["metadata"] == {"thread_id": "17585"}
+
+    FakeCompressAgentWithSummaryFailure.last_instance.close.assert_called_once()
+
+
+@pytest.mark.asyncio
+async def test_session_hygiene_informs_user_when_aux_model_fails_but_recovers(monkeypatch, tmp_path):
+    """When the user's configured ``auxiliary.compression.model`` errors out
+    and we recover via the main model, compression succeeds but the user's
+    config is still broken.  Gateway hygiene must surface an ℹ note so the
+    user knows to fix ``auxiliary.compression.model`` — silent recovery
+    hides a misconfig only they can resolve."""
+    fake_dotenv = types.ModuleType("dotenv")
+    fake_dotenv.load_dotenv = lambda *args, **kwargs: None
+    monkeypatch.setitem(sys.modules, "dotenv", fake_dotenv)
+
+    class FakeCompressAgentWithAuxRecovery:
+        last_instance = None
+
+        def __init__(self, **kwargs):
+            self.model = kwargs.get("model")
+            self.session_id = kwargs.get("session_id", "fake-session")
+            self._print_fn = None
+            self.shutdown_memory_provider = MagicMock()
+            self.close = MagicMock()
+            # Compression succeeded (no placeholder inserted) but the
+            # configured aux model errored and we fell back to main.
+            self.context_compressor = SimpleNamespace(
+                _last_summary_fallback_used=False,
+                _last_summary_dropped_count=0,
+                _last_summary_error=None,
+                _last_aux_model_failure_model="gemini-3-flash-preview",
+                _last_aux_model_failure_error="404 model not found",
+            )
+            type(self).last_instance = self
+
+        def _compress_context(self, messages, *_args, **_kwargs):
+            self.session_id = f"{self.session_id}_compressed"
+            return ([{"role": "assistant", "content": "real summary"}], None)
+
+    fake_run_agent = types.ModuleType("run_agent")
+    fake_run_agent.AIAgent = FakeCompressAgentWithAuxRecovery
+    monkeypatch.setitem(sys.modules, "run_agent", fake_run_agent)
+
+    gateway_run = importlib.import_module("gateway.run")
+    GatewayRunner = gateway_run.GatewayRunner
+
+    adapter = HygieneCaptureAdapter()
+    runner = object.__new__(GatewayRunner)
+    runner.config = GatewayConfig(
+        platforms={Platform.TELEGRAM: PlatformConfig(enabled=True, token="fake-token")}
+    )
+    runner.adapters = {Platform.TELEGRAM: adapter}
+    runner._voice_mode = {}
+    runner.hooks = SimpleNamespace(emit=AsyncMock(), loaded_hooks=False)
+    runner.session_store = MagicMock()
+    runner.session_store.get_or_create_session.return_value = SessionEntry(
+        session_key="agent:main:telegram:group:-1001:17585",
+        session_id="sess-1",
+        created_at=datetime.now(),
+        updated_at=datetime.now(),
+        platform=Platform.TELEGRAM,
+        chat_type="group",
+    )
+    runner.session_store.load_transcript.return_value = _make_history(6, content_size=400)
+    runner.session_store.has_any_sessions.return_value = True
+    runner.session_store.rewrite_transcript = MagicMock()
+    runner.session_store.append_to_transcript = MagicMock()
+    runner._running_agents = {}
+    runner._pending_messages = {}
+    runner._pending_approvals = {}
+    runner._session_db = None
+    runner._is_user_authorized = lambda _source: True
+    runner._set_session_env = lambda _context: None
+    runner._run_agent = AsyncMock(
+        return_value={
+            "final_response": "ok",
+            "messages": [],
+            "tools": [],
+            "history_offset": 0,
+            "last_prompt_tokens": 0,
+        }
+    )
+
+    monkeypatch.setattr(gateway_run, "_hermes_home", tmp_path)
+    monkeypatch.setattr(gateway_run, "_resolve_runtime_agent_kwargs", lambda: {"api_key": "***"})
+    monkeypatch.setattr(
+        "agent.model_metadata.get_model_context_length",
+        lambda *_args, **_kwargs: 100,
+    )
+    monkeypatch.setenv("TELEGRAM_HOME_CHANNEL", "795544298")
+
+    event = MessageEvent(
+        text="hello",
+        source=SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="-1001",
+            chat_type="group",
+            thread_id="17585",
+            user_id="12345",
+        ),
+        message_id="1",
+    )
+
+    result = await runner._handle_message(event)
+
+    assert result == "ok"
+    # No ⚠️ hard-failure warning (that's for dropped turns)
+    hard_warnings = [s for s in adapter.sent if "Context compression summary failed" in s["content"]]
+    assert len(hard_warnings) == 0, adapter.sent
+    # But an ℹ note about the configured aux model must be delivered.
+    aux_notes = [
+        s for s in adapter.sent
+        if "Configured compression model" in s["content"]
+    ]
+    assert len(aux_notes) == 1, (
+        f"Expected 1 aux-model fallback notice, got {len(aux_notes)}: {adapter.sent}"
+    )
+    note = aux_notes[0]
+    assert "gemini-3-flash-preview" in note["content"]
+    assert "404" in note["content"]
+    assert "auxiliary.compression.model" in note["content"]
+    # Note must land in the originating topic/thread.
+    assert note["chat_id"] == "-1001"
+    assert note["metadata"] == {"thread_id": "17585"}
+
+    FakeCompressAgentWithAuxRecovery.last_instance.close.assert_called_once()