github-featured / learningcircuit--local-deep-research

发布

0 发布 174 Git 标签

RSS 订阅源
  • v1.6.3 fc4a3b42cb
    比较

    fix(ci): switch e2e LLM to gemini-2.5-flash-lite, tolerate transient errors in markdown export check (#3696)

    frostbyte_neo 发布于 2026-04-27 06:49:28 +00:00

    The Puppeteer "should export and display research output" test failed
    with a 429 from google/gemini-2.0-flash-001 ("temporarily rate-limited
    upstream"). The results-page check already tolerates rate-limit / 5xx
    errors via isTransientLlmError, so it correctly let that failure pass;
    but the later markdown-export check was missing the same tolerance,
    which is what tripped the assertion.

    Two changes:

    1. Switch the e2e model from google/gemini-2.0-flash-001 to
      google/gemini-2.5-flash-lite. 2.5 Flash Lite is GA, ~3x cheaper,
      and on a less-contended OpenRouter pool. Gemini 2.0 Flash has been
      superseded for ~10 months. Updated all four references (workflow
      settings.toml, env var, PR-comment template, JS test fallback).

    2. Mirror the isTransientLlmError tolerance in the markdown-export
      check, matching the existing pattern on the results page. Even with
      the new model, any provider can rate-limit; the asymmetry was a
      latent bug.

    下载附件