[BUG] ultrareview: crashed before producing findings on large PR"

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Title

ultrareview: "Review crashed before producing findings" on large folded PR (no usable output)

What happened

Ran /ultrareview 13777 against housing-cloud/web#13777. The job completed (status went to completed, billing hit), but the only output was:

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

Re-ran it; same result. No findings, no partial output, no stack trace surfaced to the client.

Repro

• Target: housing-cloud/web#13777 (private repo)
• Command: /ultrareview 13777
• PR shape: 69 commits, 202 files changed, +51,640 / -397 lines (a Graphite stack of ~42 PRs folded into one via gt fold --close)
• Sessions (for log access):
• https://claude.ai/code/session_014cGJwoxgsvbUfAj9M4s1or
• https://claude.ai/code/session_012rSGfQp4ZWit8TfPvy9WH7

Expected

Either findings, or a clear error explaining what went wrong (timeout, size cap, parser failure, etc.) so the user knows whether to retry or restructure the PR.

Actual

Generic "Review crashed before producing findings. See session logs for details." with no actionable detail. The user has no way to read the session logs themselves.

Suspected cause

Likely a size-related crash — smaller PRs in the same repo work normally; this one is unusually wide because it's a folded stack. Could be a memory cap, file-count cap, diff-size cap, or per-file timeout in the review pipeline.

Asks

1. Surface a real error (size limit hit / OOM / parse error / etc.) instead of the generic crash message
2. If there's a size cap, document it and bail early with a "PR too large for ultrareview" message before billing
3. Investigate the two linked sessions for the actual stack trace

Environment

• Claude Code CLI on macOS (darwin 24.1.0)
• Repo: housing-cloud/web (private), branch feat-reports-v2-reports-db, base main

What Should Happen?

/ultrareview should either return findings or surface a real, actionable error (e.g. "PR exceeds size cap of N files / N lines"). The user should know whether to retry, restructure the PR, or wait — and ideally not be billed for crashes that produce no output.

Error Messages/Logs

{"error":"Review crashed before producing findings. See session logs for details."}
(That's all the client got — the user has no way to read the session logs themselves.)

Steps to Reproduce

1. In a Claude Code session, run /ultrareview <PR#> against an unusually large PR (e.g. a Graphite stack folded via gt fold --close into a single branch).
2. PR shape in our case: 69 commits, 202 files changed, +51,640 / -397 lines.
3. Wait ~5–10 min for the cloud job to complete.
4. Observe the only output is the generic crash error above. No findings, no partial output, no traceback.

Reproduced twice in a row against the same PR (housing-cloud/web#13777). Sessions for log access:

• https://claude.ai/code/session_014cGJwoxgsvbUfAj9M4s1or
• https://claude.ai/code/session_012rSGfQp4ZWit8TfPvy9WH7

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.123

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

_No response_