mirror of
https://github.com/koala73/worldmonitor.git
synced 2026-04-25 17:14:57 +02:00
81d773d2bb
* feat(intelligence): DeepEar QW-1/QW-3/QW-4 — Polymarket context injection, dual-model routing, robust JSON parsing QW-1: Inject crowd-calibrated Polymarket/Kalshi odds into deduction prompts - Fetch prediction:markets-bootstrap:v1 from Redis and keyword-score markets against user query; top-7 matches appended as structured context block - Hash prediction context into cache key so odds movements trigger fresh LLM calls - Sanitize market titles with sanitizeHeadline() before prompt injection - deductSituation now uses callLlmReasoning (explicit reasoning-tier routing) QW-3: Dual-model LLM routing via env vars - callLlmTool (LLM_TOOL_PROVIDER/MODEL, default groq) for extraction tasks - callLlmReasoning (LLM_REASONING_PROVIDER/MODEL, default openrouter) for synthesis - Private callLlmProfile factory eliminates duplicated wrapper bodies QW-4: Shared _llm-json.mjs utility for robust LLM JSON parsing - cleanJsonText: strips C-style comments and trailing commas before JSON.parse - extractFirstJsonObject/Array: private extractFirstDelimited walker (was 2x23 LOC) - Removes duplicate function definitions from seed-forecasts.mjs - All JSON.parse call sites in tryParseImpactExpansionCandidate and tryParseStructuredCandidate now use cleanJsonText for comment/comma tolerance * fix(intelligence): address PR #2421 review findings - P1: getCachedJson(PREDICTION_BOOTSTRAP_KEY, true) — seed keys must be read raw (unprefixed); without this the prediction-odds block never activates outside production - P2: validate LLM_TOOL_PROVIDER / LLM_REASONING_PROVIDER env value against PROVIDER_SET before use; log a warning and fall back to defaultProvider on typo instead of silently rerouting - Filter: lower word-length threshold from >3 to >1 so short but critical terms like "war", "oil", "EU", "AI", "Fed", "USD" match prediction markets - Suggestion: bucket yesPrice to 5% increments before building the context string to reduce 1%-movement cache churn (20 bands vs 100) * fix(intelligence): guard empty prediction context header when all titles sanitize to empty