Codex rescue task gets stuck in infinite tool-call loop during code review

Description

When using the Codex plugin for Claude Code (codex:rescue) to review code changes, the Codex task occasionally gets stuck in an infinite tool-call loop. It reads all the necessary files and converges on findings, but then loops back to re-reading the same files instead of producing a final review output.

Steps to Reproduce

1. Use Claude Code with the Codex plugin (v1.0.2)
2. Invoke codex:rescue with a code review task, e.g.:

   Review the changes between commits <base> and HEAD. Focus on Spectre.Console integration, null safety, behavioral parity...
   

3. Codex reads all relevant diffs and source files successfully
4. Instead of producing a final summary, it loops back to re-reading the same files

Observed Behavior

• Codex's last progress message: "我已经收敛到几条具体问题了" (I've converged on a few specific issues)
• After this, it re-enters file reading commands (nl -ba, sed -n, rg) on files it has already read
• The loop continues until the task is terminated
• No final review output is produced

Expected Behavior

Codex should produce a final review summary after reading all necessary files, not loop back to re-reading them.

Environment

• Claude Code: v2.1.90
• Codex plugin: v1.0.2
• Codex CLI: v0.118.0
• OS: macOS (Darwin 25.4.0, Apple Silicon)
• Runtime mode: shared session (direct startup)

Additional Context

• This happened during a review of ~300 lines of changes across 6 files
• A separate Codex task reviewing the same codebase (full repo review, ~2000 lines) completed successfully in ~5 minutes
• The stuck task was in "verifying" phase when terminated
• The issue seems intermittent — not every review task triggers it

Workaround

Cancel the stuck task and retry. The retry (or a concurrent task) may complete successfully.

Start by reproducing the issue with the codex:rescue entry point during a shared-session code review, focusing on the transition from the verifying phase to final output. No source files or tests are named, so trace the repeated nl -ba, sed -n, and rg calls and compare with a review that completes successfully. Done means the task produces a final review summary instead of re-reading files indefinitely.