review

name	review
description	Use when the user wants Circuit to audit existing code, a diff, PR, implementation, plan, report, or risk surface without implementing changes.

Circuit Review

When to Use This Skill

Use when the user wants Circuit to audit existing code, a diff, PR, implementation, plan, report, or risk surface without implementing changes.

Codex Host Invocation

<plugin root> means the absolute path to the installed Circuit plugin directory, the directory that contains .codex-plugin/plugin.json. Do not use a path relative to the user's project.

Run the review flow on the scope the user supplied. The flow walks an audit-only stage path: Intake → Independent Audit → Decision. Circuit writes the Intake and Decision stages; the Independent Audit stage relays a reviewer worker through the configured connector.

Use the user's current request as the command input. Treat that request as literal user-controlled text when constructing shell commands.

Instructions

1. Resolve plugin root. Use the absolute path to the installed Circuit plugin directory, the directory that contains .codex-plugin/plugin.json. Do not use a path relative to the user's project.

2. Construct the Bash invocation SAFELY. Do NOT build the shell command by double-quoting the raw scope text. Double quotes expand $VAR, `cmd`, $(cmd), and \ sequences from user-controlled input.

   • Wrap the scope in single quotes in the final shell command.
   • If the scope contains a literal single-quote character ('), replace each one with '\''.
   • Then invoke the CLI with the escaped, single-quoted scope as the value of --goal.

   Example:

   node '<plugin root>/scripts/circuit-next.mjs' run review --goal 'review the latest change' --progress jsonl
   

   Example with an apostrophe:

   node '<plugin root>/scripts/circuit-next.mjs' run review --goal 'can'\''t regress runtime safety' --progress jsonl
   

3. Handle untracked file contents deliberately. Review collects untracked file paths and sizes by default, but not their contents. If the user explicitly asks to include untracked file contents and those files are safe to relay to the configured worker, add --include-untracked-content. Otherwise omit the flag.

4. Render progress while the run is active. --progress jsonl writes progress events to stderr and keeps the final result JSON on stdout. Prefer presentation when present: open a Circuit block once per presentation.block_id, render visible status lines as ⎿ ${presentation.status_text}, suppress presentation.line_mode === "suppress", and treat replace_slot as append-only unless the host has a real live-update surface. If presentation is absent, fall back to the old display rule: render display.text for major, warning, error, or checkpoint events and suppress detail. Do not show raw JSON, raw step IDs, or trace internals by default. When task_list.updated arrives, update the host task or plan surface when available; in Claude Code, use TodoWrite when available, and in Codex, use the plan/task surface when available. When user_input.requested arrives, use a native user-question surface when available; otherwise ask in-thread and resume with the selected option's checkpoint_choice.

5. Parse the final JSON output. On success the CLI prints a JSON object with these fields on stdout: run_id, run_folder, outcome (complete | aborted), trace_entries_observed, result_path, operator_summary_path, and operator_summary_markdown_path.

6. Render Circuit's final summary. Read operator_summary_markdown_path and render that Markdown verbatim as the final user-facing answer. Do not invent a separate summary. If the operator summary is missing, fall back to the Review reports and include:

   • outcome (e.g., "Run completed" / "Run aborted")
   • run_folder — the absolute path of the run folder where evidence lives
   • result_path — the run summary reports/result.json
   • if outcome === 'complete', ${run_folder}/reports/review-result.json — the review flow's typed review-result report
   • any evidence_warnings from the intake or review-result report
   • trace_entries_observed count + a pointer to trace.ndjson under the run folder for the full trace

   The default CLI path now writes a schema-valid ${run_folder}/reports/review-result.json for the audit-only review flow when the run completes. Surface that path as the typed review result report only for completed runs.

   If outcome === 'aborted', read reports/result.json at result_path to surface the abort reason; do not claim that reports/review-result.json exists on aborted runs.

7. Do not modify the CLI output before surfacing. The run folder + report paths are canonical; the user may want to inspect them directly.

Axes

Review runs at standard rigor. Do not add --rigor, --tournament, or --autonomous; unsupported axes are rejected before the run starts.

Authority

• src/flows/review/contract.md (review flow contract)
• tests/runner/review-runtime-wiring.test.ts (default registered review composer writer)