- stream=true now returns SSE (role chunk → content chunk → finish → [DONE])
instead of 501. Not token-by-token but compatible with frontends like
Open WebUI that require SSE format.
- Add conversation parameter for named session chaining (like /title).
Mutually exclusive with previous_response_id.
Like /title for sessions — clients can name conversations instead of
tracking response IDs manually:
POST /v1/responses {input: 'hi', conversation: 'my-project'}
POST /v1/responses {input: 'next step', conversation: 'my-project'}
Server automatically chains to the latest response in that conversation.
Mutually exclusive with previous_response_id (returns 400 if both set).
Not stored if store=false.
5 new tests (72 total).
Adds a new gateway platform adapter that exposes an HTTP server with
OpenAI-compatible endpoints, allowing any OpenAI-compatible frontend
(Open WebUI, LobeChat, etc.) to use hermes-agent as a backend.
Endpoints:
- POST /v1/chat/completions - OpenAI Chat Completions format (stateless)
- POST /v1/responses - OpenAI Responses API format (stateful via previous_response_id)
- GET /v1/models - lists hermes-agent as an available model
- GET /health - health check
Features:
- Bearer token auth via API_SERVER_KEY env var (unauthenticated when no key set)
- System messages/instructions become ephemeral system prompt (layered on top of core prompt)
- In-memory LRU response store for Responses API conversation chaining
- Agent runs in thread executor (run_conversation is synchronous)
- Streaming returns 501 (not yet implemented)
New files:
- gateway/platforms/api_server.py - APIServerAdapter class (~470 lines)
- tests/gateway/test_api_server.py - 51 tests covering all endpoints, auth, config
Modified files:
- gateway/config.py - Added Platform.API_SERVER enum, env var overrides, connected platforms
- gateway/run.py - Added _create_adapter() case, auth map entries, auth bypass
teknium1