[Bug] ultrareview: All reviewer agents terminate at Verify stage, Dedupe never executes
Bug Description
ultrareview fails at the Verify stage: all reviewer agents terminate, Dedupe never runs, free run consumed
Environment Info
- Platform: darwin
- Terminal: vscode
- Version: 2.1.235
- Feedback ID: 2d55dc30-5382-49bc-8d36-45c24fe0b617
Errors
[]
Summary
/code-review ultra on a GitHub pull request failed with "Review failed: all review agents terminated before completing." The cloud session shows Setup and Find completed successfully and Verify failed with no visible output. Dedupe never started. No findings were returned, nothing was posted to the PR, and the run consumed a free ultrareview.
Environment
- Claude Code 2.1.235
- macOS (Darwin 25.3.0), zsh, interactive session
- Auth: claude.ai account (Pro/Max free-run allotment)
- Target: a pull request on github.com in a private repository, open and not draft
- Diff size: 68 changed files, 4,323 changed lines (+3,483 / -840)
- Date: 2026-08-19
- Cloud session id and repository details available privately on request
Reproduction
/code-review ultra <PR url>- In the launch dialog, chose to post the findings to the PR.
- CLI confirmed: "Free ultrareview 1 of 3. Ultrareview launched ... runs in the cloud."
- Roughly 5 minutes later the CLI received a failed task notification.
Expected
Verify runs, Dedupe runs, findings return to the session and are posted to the PR.
Actual
Cloud session UI:
Resumed session
Resumed your cloud container
Refreshed your repository
Started Claude Code
Review failed
Setup (ok)
Find (ok)
Verify (failed)
Dedupe (never started)
The Verify log pane contains only:
[heartbeat] 14:53:57
[heartbeat] 14:54:07
Session error payload:
<remote-review>{"error": "Review failed: all review agents terminated before completing. See session logs for details."}</remote-review>
CLI task notification:
status: failed
summary: Cloud review failed: orchestrator reported an error:
Review failed: all review agents terminated before completing.
See session logs for details.
Nothing was posted to the PR.
What this is not
Ruled out before filing:
- Not a diff-size refusal. 68 changed files and 4,323 changed lines, against the documented 500-file / 8,000-line limits. The size gate also runs before any review work, and this run got through Setup and Find, so it never fired.
- Not container or clone setup. Setup completed, and Find agents ran successfully in the same container.
- Not repo-side config. The repository's project-level hooks are scoped to a path that has no uncommitted changes in a fresh clone, so they are inert there, and its committed symlinks are repo-relative and resolve after clone. Find succeeding in the same container corroborates this.
- Not a 30-minute timeout. The failure arrived within minutes, and the error text is the agent-termination message, not the timeout message. This appears distinct from #54229 and #53939.
The failure is specific to the Verify fan-out.
Diagnostics not available locally
TaskOutput on the local task id returns "No task found", and the local session transcript holds only the one-line failure summary, so there is no local log to attach. The cloud session log shows no more than what is quoted here.
Requests
- Root-cause why the Verify agent fleet terminated. Whatever killed those agents is not surfaced anywhere the user can see, in the CLI or in the cloud session view.
- Surface a real reason in the Verify pane. "All review agents terminated" plus two heartbeat lines gives a user nothing to act on, and no local artifact survives to diagnose from.
- Reconsider charging a free run for this. The docs state a run that fails to complete still consumes one, but Find had already completed and its results were discarded when Verify died. Same request as #49706.
Possibly related
- anthropics/claude-code#49706 (failed run consumes a free run)
- anthropics/claude-code#54229 (hangs after start, no findings, free runs consumed)
- anthropics/claude-code#53939 (cloud completes, CLI reports failed)