github-featured / wanshuiyin--auto-claude-code-research-in-sleep

发布

0 发布 41 Git 标签

RSS 订阅源
  • v0.4.17 65a925d50a
    比较

    fix(v0.4.17 push-gate): silence MCP notification spam, stop Codex model override, fix misleading LlmReview error

    frostbyte_neo 发布于 2026-06-13 11:43:00 +00:00

    Real-machine dogfooding of the zero-API-key Codex MCP reviewer surfaced
    three first-impression UX edges (all UX, no protocol change):

    • BUG A (mcp_stdio.rs): the per-notification aris mcp: notification skipped stderr line flooded the REPL — codex emits dozens-to-hundreds
      of codex/event notifications per call. Gated behind the existing
      ARIS_MCP_STDERR=inherit debug flag (read once per round-trip); silent by
      default. Control flow unchanged: notifications still skipped, read loop
      still waits for the id-bearing response.
    • BUG B (main.rs reviewer_routing_nudge): both codex-mcp states now tell
      the model NOT to pass a model parameter — a ChatGPT-account Codex
      rejects arbitrary names (e.g. gpt-5.2) and the first call fails until
      retried without it. Account default = gpt-5.5 + xhigh. The no-fallback
      state flips from silent to one guidance line (deliberate, annotated).
    • BUG C (tools/lib.rs run_llm_review): codex-mcp + no fallback no longer
      falls through to the OpenAI-compat path (which complained OPENAI_API_KEY
      was unset for gpt-5.5 — a credential and model the user never opted
      into). Returns a clear message directing to mcp__codex__codex; leaks no
      credential env name. (deliberate flip, annotated)

    Two characterization tests flipped in place (annotated). Full suite green
    (--test-threads=1): runtime 199 / tools 67 / aris-cli 165 / commands 5.
    Codex MCP R17 (gpt-5.5 xhigh, read-only): VERDICT GO, zero-regression
    confirmed (a)-(i).

    Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com

    下载附件