claude-code - 💡(How to fix) Fix [BUG] /ultrareview crashes in Dedupe stage after Verify completes — findings produced but never delivered to CLI

Preflight Checklist

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

What's Wrong?

/ultrareview 2 ran the review pipeline through Setup → Find → Verify successfully, generating 8 confirmed findings (visible in the cloud session console UI), then crashed in the Dedupe stage. The CLI received only a generic "Review crashed before producing findings" error and consumed one free-tier credit, despite the cloud agent having actually produced findings that were never delivered.

What Should Happen?

When the review pipeline produces confirmed findings in the Verify stage, those findings should be delivered to the CLI even if a downstream stage (Dedupe) fails. At minimum, a partial-findings result should be returned rather than a generic crash error. A credit should not be consumed for a review whose findings never reach the user.

Error Messages/Logs

CLI output (final task notification result):

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

  Cloud session UI state at time of crash (https://claude.ai/code/session_01N8DwK7zN8CG1Lsx8EYvqNv):

  Initialized session
    ✓ Set up a cloud container
    ✓ Cloned repository
      Run setup script (none configured)
    ✓ Started Claude Code

  Review failed
    ✓ Setup
    ✓ Find       — 25 candidates
    ✓ Verify     — 8 confirmed, 17 refuted
    ✗ Dedupe     — 8 issues, then:
        [heartbeat] 11:14:13
        [heartbeat] 11:14:23
        (no further events)

  Browser console errors observed on the session viewer page
  (`https://claude.ai/code/session_01N8DwK7zN8CG1Lsx8EYvqNv`):

  GET /api/claude_code/organizations/{org}/user_settings           → 404
  GET /api/organizations/{org}/model_configs/claude-opus-4-7[1m]   → 404
  POST /v1/toolbox/shttp/mcp/{uuid}                                → 405
  [REACT_QUERY_CLIENT] QueryClient error: Tu: Not found

  `/v1/sessions/{session_id}/events?limit=1000` (without `after_id`) returns
  404, so the full event log isn't recoverable post-hoc.

Steps to Reproduce

1. From the CLI: /login (succeeds)
2. /ultrareview 2 against a large PR
   • Repo: sjanoe123/medecon-stack, PR #2
   • PR diff: 71 files, +32,897 / -52, ~33k diff lines
   • Branch release/v1.2 → main, mergeable
3. CLI confirms launch: Free ultrareview 1 of 3. Ultrareview launched for sjanoe123/medecon-stack#2
4. Wait ~5–10 minutes
5. Task notification fires completed; CLI prints the JSON error above
6. Cloud session UI shows Dedupe-stage failure as described

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

No response

Claude Code Version

2.1.132 (claude code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Session identifiers:

• Session: session_01N8DwK7zN8CG1Lsx8EYvqNv
• Cloud session: cse_01N8DwK7zN8CG1Lsx8EYvqNv
• Org: e66b7663-4efe-407e-8c01-5599cc0a5993
• Run timestamp: 2026-05-07 ~11:14 UTC

Credit impact: One free-tier /ultrareview credit (1 of 3) was consumed without findings being delivered to the CLI. Refund requested.

Suspected separate bugs surfaced during diagnosis:

1. The model_configs API doesn't recognize claude-opus-4-7[1m] (the 1M variant) — returns 404 in the session viewer.
2. The Streamable-HTTP MCP toolbox endpoint (/v1/toolbox/shttp/mcp/{uuid}) returns 405 Method Not Allowed during session setup.
3. The /v1/sessions/{session_id}/events endpoint requires after_id; without it, returns 404. Makes post-hoc debugging of crashed sessions difficult unless DevTools was open at page load.

Suggested mitigations (in priority order):

1. Make Dedupe failures non-fatal — return Verify-stage findings to the CLI with a "dedupe failed, findings may contain duplicates" warning.
2. Stream findings to the CLI as Verify produces them, so a downstream stage failure doesn't lose the whole review.
3. Provide an out-of-band fetch (e.g., /ultrareview --fetch <session_id>) to retrieve findings from a completed-but-undelivered cloud session.

Reproducibility note: This was the first /ultrareview invocation on this account after a fresh /login. PR diff size (~33k lines) may be a factor in the Dedupe-stage hang, but Find/Verify completed successfully on that same diff, so size alone is unlikely to be the root cause.