Cloud ultrareview fails reproducibly: 'all review agents terminated before completing' — 2 runs, 0 findings, 2 free credits consumed

Summary

Two consecutive cloud ultrareview runs (/ultrareview from the CLI) failed with the identical orchestrator error and produced zero findings, consuming 2 of the 3 free ultrareview credits with no output.

Environment

• Claude Code CLI on macOS (darwin arm64), launched from a git repo (private), model session: Fable 5.
• Both runs launched 2026-08-16 (~12:30 and ~13:40 PT).

Repro / what happened

1. /ultrareview on a branch diff of 107 files, +10,212/−445 → launched to the cloud → failed.
2. After trimming the diff, /ultrareview on 118 files, +10,389/−469 (~10.9K lines, under the stated 12,000-line limit) → identical failure.

Both runs ended with:

Review failed: all review agents terminated before completing. See session logs for details.

Evidence from the streamed progress (both runs, identical signature)

• ~3,900 <remote-review-progress> heartbeats in stage: finding, with bugs_found: 0, bugs_verified: 0, bugs_refuted: 0 for the entire run — the counter never populated once.
• 4 heartbeats in stage: verifying, then the terminal <remote-review>{"error": ...} frame.
• ~680KB of streamed output per run, no findings content in either.

Cloud session ids for your logs:

• Run 1: session_01EUtoU2Rz5X7uyNASeoCzHU
• Run 2: session_01Prhn57AjjTKsvY7qeMs7ve

Notes

• The same account's smaller-scope dialogs rendered fine (the confirmation UI, scope counting, and the 12,000-line limit check all behaved correctly; a separate attempt correctly refused at 12,516 lines).
• The failure looks orchestrator-side (agents terminating during the finding stage), not diff-content-related — the second run's diff differed substantially from the first and the signature was byte-similar.

Asks

1. Investigation of the two sessions above.
2. Restoration of the 2 consumed free ultrareview credits, since neither run produced any output.