claude-code - 💡(How to fix) Fix ultrareview crashes "before producing findings" yet still consumes a free run

Summary

/ultrareview <PR> ran in the cloud and returned, twice in a row, only:

{"error":"Review crashed before producing findings. See session logs for details."}

Both attempts produced zero findings and no partial output, but each one still decremented the free ultrareview quota (the launch banner reported "Free ultrareview 1 of 3", then "Free ultrareview 2 of 3"). Net effect: 2 of 3 free runs spent for nothing.

Impact

• 2 free ultrareview runs consumed, 0 review output delivered in either.
• The failure is server-side and opaque to the client — the only client-visible signal is the generic crash string above; the referenced "session logs" are not accessible from the CLI.

Reproduction

• Client: Claude Code CLI (native install).
• Command: /ultrareview <PR-number> targeting a pull request in a private GitHub repository (org/repo intentionally omitted from this report — see Privacy note).
• Date of both occurrences: 2026-05-18, within roughly one hour of each other.
• The first run was against the PR's original state; the second was against the same PR after new commits were pushed. The error was byte-for-byte identical both times, which points to a deterministic, input-dependent crash rather than a transient outage.

Possible trigger (hypothesis)

The target PR has an unusually large diff: on the order of ~700 changed files and ~32,000 deletions, the large majority of which come from de-committing a previously-tracked generated build-output directory (i.e. the diff is dominated by bulk deletions of generated artifacts, not source changes).

A plausible root cause is the review harness attempting to fetch/process the full PR diff and crashing on diff size/volume before it reaches the actual source surface. Filing this as a hypothesis to aid triage, not a confirmed cause.

Expected behavior

Either:

1. ultrareview completes (or degrades gracefully — e.g. reviews the source-bearing subset and reports the rest skipped), or
2. if the harness crashes and produces no findings, the run is not counted against the free quota, and/or the consumed runs are automatically credited back.

Actual behavior

Crash with no findings and no partial output, free run consumed, no automatic credit, no actionable client-side detail.

Request

1. Investigate the server-side crash (logs referenced in the error message should exist for the two sessions below).
2. Credit back the 2 consumed free ultrareview runs given neither produced any output - for all ultrareview failures of this type for all accounts and customers.
3. If large-diff / bulk-deletion PRs are a known failure mode, consider a graceful-degradation path and a clearer client-side error.

Session references

These are the two tracked review sessions

• Run 1 (free run 1 of 3): https://claude.ai/code/session_01E1uqCVYDPKwmDQtG9MhCzL
• Run 2 (free run 2 of 3): https://claude.ai/code/session_01GGYPtCZ6QMzsbXHvM1Nxqh