Skip to content

fix(benchmarks): restore verified dependency caches for restricted runs - #58

Open
minixalpha wants to merge 5 commits into
mainfrom
bench/pi-restricted-integrity-rerun-20261104
Open

minixalpha wants to merge 5 commits into
mainfrom
bench/pi-restricted-integrity-rerun-20261104

Conversation

@minixalpha

@minixalpha minixalpha commented Nov 4, 2026 •

Copy link
Copy Markdown
Owner

Restricted Harbor runs could not use the historical dependency-cache Compose overlays, leaving Pipeline's multi-gigabyte setup to time out. Add an opt-in, controller-owned verified cache for PyPI wheels and original uv/Python archives. Helpers exist only during setup and official verification, and their removal is checked before model work. Task images, package sources, official scripts, native deadlines, and the API-only model allowlist remain unchanged.

The rerun also exposed Harbor's 64 KiB streamed-line limit: Pi's final 209,940-byte JSON event aborted collection after the agent had completed. Use UTF-8-aware chunked reads while preserving deadlines, stdin handling, cancellation, and process cleanup. Replaying the frozen native output reproduces the original failure and confirms exact output preservation after the fix.

Recorded outcomes:

  • WAL passed all 7 official tests in the November 4 restricted run, with no integrity violation observed in the reviewed evidence.
  • Pipeline's corrected dependency preflight completed in 205.670052 seconds; setup in a fresh formal container completed in 142.791023 seconds. Both used 45 verified wheel-cache hits and both original release archives.
  • Pipeline completed 35 tool calls, but the collector error prevented official verification. Its reward and effective score remain absent. No external answer retrieval was observed; broad local test/function searches and review limitations are documented. No extra model attempt was launched after the collector failure.
  • Preserve the original setup timeout, the subsequent truncated uv download, and the unscored collector failure. Raw model logs remain private because an environment dump contains the provider credential.

Validation: 410 Harbor tests passed, plus all 3 explicitly enabled Docker tests. New output freezes cover 71 files; the earlier 49 restricted-run files and 326 historical files remain unchanged. All 7 initial Docker images are preserved. English and Simplified Chinese reports are complete, synchronized, and link-checked.

Reports: restricted WAL/Pipeline check · Pipeline dependency recovery and collector failure · 中文报告.

Pipeline's official outcome is still unresolved. A further model attempt requires an explicit maintainer exception to the repository's verifier-timeout-only retry rule.

@minixalpha minixalpha changed the title docs(benchmarks): record restricted pi integrity rerun fix(benchmarks): restore verified dependency caches for restricted runs Nov 5, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant