claude-code - 💡(How to fix) Fix [Bug] /ultrareview crashes mid-stream on a normal-sized PR (no findings) [1 comments, 1 participants]

Summary

/ultrareview crashed twice in a row against a normal-sized PR with no findings produced either time. /review against the same PR completed cleanly, suggesting the issue is specific to the ultrareview pipeline rather than the input.

Environment

• Claude Code: 2.1.126
• Platform: Windows (git for Windows, native paths)
• Shell: bash (via git for Windows)
• Model: default (whatever /ultrareview picks internally)

Repro

/ultrareview 103

against this PR: https://github.com/dotcal604/worthcall/pull/103

PR shape (nothing exotic):

• 2 commits, 17 files, 1034 additions / 180 deletions
• ~772 of the additions are auto-generated package-lock.json (husky + lint-staged install)
• Real code: shared/pricing.ts header rewrite, web/eslint.config.js token-enforcement rule, .husky/pre-commit (new), 7 marketing .tsx files migrated to brand tokens, 2 doc-prompt updates
• No binaries, no submodules, no LFS, no unusual file types

(The PR is in a private repo so Anthropic eng may not have access; happy to share specifics if useful.)

Expected

/ultrareview produces a findings report.

Actual

Both runs crashed mid-stream. No findings emitted either time.

Workaround

/review 103 against the same PR ran cleanly. Will append /review findings as a comment for comparison.

Session IDs

• First failed run: session_01MaTToiAnCNTuPvi7cR6jVr
• Second failed run: unknown (will append in a comment if surfaced)

Notes for whoever picks this up

• If this is reproducible against any moderately-sized PR (~1k additions, ~17 files, mix of TS/JSON/md/JS), there may be an upper-bound or timeout issue in the ultrareview pipeline.
• /review worked on identical input — the differentiator is whatever ultrareview does extra.
• Two crashes in a row from the same input strongly suggests it's not transient.