Constants were purely in-memory; any server restart reverted them
to code defaults. Now saved to metadata table as JSON on every
autosave and restored on startup.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Default 200→400, invention 300→600, desireGeneration 500→800,
backstoryAndDesires 800→1200. Can trim back once we have usage data.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
MODEL_PRICING keys were 'gpt-oss-120b' but the actual model string
from env is 'openai/gpt-oss-120b'. The lookup always missed, so
calculateCost returned 0 for every call. Updated keys to match the
full OpenRouter model identifiers.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Templates with jsonResponse:true automatically get the compact
single-line JSON instruction appended to their system prompt via
renderTemplate. Removes manually duplicated JSON formatting
instructions from invention, backstoryAndDesires, and desireGeneration
templates.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Model sometimes returns pretty-printed JSON with newlines/indentation,
wasting tokens and causing truncation. Now requests compact single-line
JSON. When JSON is truncated, regex-extracts the backstory field rather
than setting raw JSON as the backstory text.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Post-restart logs show model still hitting 600-token ceiling on verbose
responses — truncated JSON at exactly 599-600 output tokens. Average
successful response was ~418 tokens but outliers reach 600+. Added
explicit conciseness instructions to backstory prompt to reduce variance.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Every single LLM failure was finish_reason='length' — the model was
running out of output tokens before completing the JSON response.
backstoryAndDesires at 350 tokens couldn't fit a backstory + 2 desires
with fulfillment criteria; desireGeneration at 250 was too tight for
a single desire with reasoning.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Captures full prompt content, response text, HTTP status, and finish
reason on every LLM failure path. Replaces silent null returns in
openRouterClient with structured CompletionFailure objects so error
details propagate up through the queue and service layers.
Also removes recipeList from backstoryAndDesires prompt (not useful
for backstory generation) and adds schema migration v3.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Every failure path was silent — LLM nulls, parse failures, validation
rejections, and promise errors were all swallowed. Adds per-attempt
logging at each gate plus a periodic summary of NPC/inventory/item state.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix LLM stats panel stuck on "Loading stats..." by requesting stats
immediately after admin authentication succeeds
- Add retry mechanism for backstory generation (up to 3 attempts with
increasing delays) so transient failures don't permanently skip NPCs
- Fix stale model bug: queued LLM requests now use the current model at
dequeue time instead of capturing it at enqueue time, so fallback
model switches take effect for pending requests
- Add per-template maxTokens (backstory: 350, invention: 300, desire:
250) to prevent "length" finish truncation on structured JSON responses
- Remove unused "reasoning" field from backstory prompt to reduce token usage
- Add failure logging throughout LLM pipeline (backstory null results,
queue errors) for better diagnostics
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Extract shared WORLD_PREAMBLE and prepend to all system prompts
- Remove duplicated world-setting text from individual templates
- Add server/saves/ and treesnstone.zip to .gitignore
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Wire npcBrainSystem to RuntimeConstants (was using hardcoded shared imports)
- Add auth gating on admin-update-constant and admin-reset-defaults sockets
- Unsubscribe from previous tab's data sources on left-panel tab switch
- Use per-key debounce timers in AdminPanel to avoid dropped changes
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace direct constant imports with rc.get() calls in needsDecaySystem,
socialSystem, statModifierSystem, movementSystem, gatheringSystem,
craftingSystem, buildingSystem, and SocketServer. All system functions
now accept a RuntimeConstants parameter, enabling live-tuning of game
constants via the admin panel.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace direct constant imports (TICK_RATE, BROADCAST_EVERY_N_TICKS,
ENERGY_DECAY_PER_TICK, DAY_NIGHT_RATIO) with RuntimeConstants lookups.
Add LogService for autosave logging and tick overrun detection.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
root