Update Grok review mechanics

[kelchm]

What changed

• Replace the plan-mode agentic review recipe with a throwaway-worktree recipe using approvals.
• Add a separate plan-mode recipe for pure single-generation reviews with all inputs embedded.
• Document the observed prompt-size ceiling, cancellation/truncation signatures, and shared output cap.
• Record the verified -w / --worktree semantics from Grok 0.2.103.

Why

FanaBridge review sessions on 2026-08-02 reproduced plan-mode cancellation at turn 1–2 whenever Grok attempted a tool call, including read-only commands such as git diff. A 140 KB prompt was also silently truncated, while 119 KB worked. Long outputs continue to cancel mid-generation.

Impact

Future sessions have distinct safe recipes for tool-using reviews and tool-free generation, plus recognizable failure signatures and conservative input/output limits.

Validation

• Confirmed -w and --worktree-ref behavior with grok 0.2.103 --help.
• Ran git diff --check.
• Confirmed the tracked source matches the rendered ~/.claude/CLAUDE.md.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 2cac7fe2-07cc-4c12-9afc-d81587dc3266

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[kelchm]

there has to be a better pattern than this -- further testing needed.

[kelchm]

Superseded by #7.

The two central claims here don't survive testing, and both trace to experiments run inside /tmp:

• "--sandbox does not prevent edits to the working directory" — grok's read-only profile leaves /tmp, /var/tmp, /var/folders and ~/.grok writable by design. sandbox-events.jsonl shows the original runs used /private/tmp/.../scratchpad/. Outside those paths the sandbox blocks the edit tool, shell writes, and new-file creation at the OS level.
• "-w gives effective read-only protection for the real checkout" — -w is silently ignored in headless mode. No worktree is created, edits land in the real checkout, and nothing is printed to stderr. With --always-approve that recipe is more dangerous than the one it replaced.

Also: the 119–140 KB input ceiling doesn't reproduce (151 KB and 204 KB read in full), and grok does ship a headless /review skill.

#7 rebuilds the guidance on --deny rules (which fail closed) and plain git worktree, and adds a grok-implementation skill.