main

KKnockOutEZfeat(llm): optional vision/image input on the text-call adapters

93112a7b创建于 7月25日历史提交

文件	最后提交记录	最后更新时间
cloud	feat(llm): optional vision/image input on the text-call adapters	2 个月前
custom-backend.test.ts	chore: remove internal dev-process references from tracked files Strip development breadcrumbs (codename, slice/phase/wave/flaw and audit-item tags, internal design-doc pointers) from source comments, SQL migration headers, tests, CHANGELOG, and a benchmark baseline. Correct stale diff/watch tool-schema comments that described them as unimplemented stubs. Reword to preserve the real rationale; no code, control-flow, or test-assertion changes.	3 个月前
llm-retry.test.ts	fix(llm): retry transient 429/503/5xx with exponential backoff in runLlmText Bench v0.1.12 showed every Gemini consumer (search/format=answer, agent, research) silently degrading to heuristic fallback when Flash Lite returned 503 'high demand'. Two retries with 500ms/1500ms backoff recover most transient throttles without leaking into the score.	4 个月前
llm-runner.test.ts	feat(llm): ollama alias + model auto-pick for local synthesis Add resolveCustomBackend() mapping both http(s) URLs and the literal `ollama` alias (env or config.json) to a keyless OpenAI-compatible backend. Wire it at the three sites that previously only matched http(s):// prefixes — isLlmConfigured(WithKeyStore), the runLlmText custom path, and selectProviderWithKeyStore (returns null, no API key). When the backend is ollama and no model is set, auto-pick an installed synthesis model from /api/tags (priority list -> first installed -> recommended fallback), resolved once per request. Explicit WIGOLO_LLM_MODEL still wins and skips the probe. Mid-run server failure surfaces through the existing retry path unchanged, so the research pipeline falls back to its template report.	3 个月前
llm-select-seam.test.ts	feat(llm): ollama alias + model auto-pick for local synthesis Add resolveCustomBackend() mapping both http(s) URLs and the literal `ollama` alias (env or config.json) to a keyless OpenAI-compatible backend. Wire it at the three sites that previously only matched http(s):// prefixes — isLlmConfigured(WithKeyStore), the runLlmText custom path, and selectProviderWithKeyStore (returns null, no API key). When the backend is ollama and no model is set, auto-pick an installed synthesis model from /api/tags (priority list -> first installed -> recommended fallback), resolved once per request. Explicit WIGOLO_LLM_MODEL still wins and skips the probe. Mid-run server failure surfaces through the existing retry path unchanged, so the research pipeline falls back to its template report.	3 个月前
local-tier.test.ts	fix(llm): bound resolveLocalModelTier model pick with a timeout A server can pass the fast reachability probe then stall on the /api/tags model-list call, so the unbounded pickModel() could hang the resolver. Run the pick under an AbortController on the same budget as the probe (pickTimeoutMs, default DEFAULT_PROBE_TIMEOUT_MS); on abort/failure it degrades to a null tier instead of blocking. Deterministic mocked test proves the pick receives the signal and the resolver resolves promptly on a stalling server.	3 个月前