Dispatch code review to a subagent. Use when the user wants code reviewed — e.g. "review this", "review my changes", "check the code", "look over this", "code review", "any issues with this?", "is this code ok?", "review before merge", "review the diff", "what do you think of these changes?". Also triggers on requests to review specific files, commits, or branches.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Dispatch code review to a subagent. Use when the user wants code reviewed — e.g. "review this", "review my changes", "check the code", "look over this", "code review", "any issues with this?", "is this code ok?", "review before merge", "review the diff", "what do you think of these changes?". Also triggers on requests to review specific files, commits, or branches.
user-invocable
true
created
2026-02-17T00:00:00.000Z
updated
2026-07-06T00:00:00.000Z
model
opus
/cf-review
CLI Requirement: OPTIONAL — Uses the memory MCP from coding-friend-cli for fast indexed search and storage. Without the CLI: falls back to grep over docs/memory/ and direct file writes. Full functionality preserved, slower memory recall. See CLI requirements.
✨ CODING FRIEND → /cf-review activated
Review the code changes for: $ARGUMENTS
Auto-Triggered
This skill is automatically invoked by other skills — you don't always need to run it manually:
/cf-plan — runs /cf-review after all implementation tasks complete
/cf-fix — runs /cf-review after the fix is verified
/cf-optimize — runs /cf-review after the optimization is measured and verified
Workflow
Step 0: Custom Guide
Custom guide — auto-loaded below (if the raw command shows instead of its output, run it yourself):
If output is not empty, integrate returned sections: ## Before → before first step, ## Rules → apply throughout, ## After → after final step.
Step 1: Identify the target
If $ARGUMENTS is empty, review all uncommitted changes (git diff + git diff --staged)
If $ARGUMENTS is a file path, review that file
If $ARGUMENTS is a commit range (e.g., HEAD~3..HEAD), review those commits
If $ARGUMENTS is a natural language description (e.g., "the auth logic changes"), review all uncommitted changes but focus the review on the described area — filter findings to only report issues relevant to that description
If $ARGUMENTS contains --deep or --quick, use that mode (override auto-detection)
Codex dual-review flag:
If $ARGUMENTS contains --with-codex (or its alias --codex), set codex=true and strip the flag from $ARGUMENTS before any other parsing.
Otherwise, read review.withCodex from the config file (CF_CONFIG_FILE, default .coding-friend/config.json). If it is true, set codex=true (config-gated default; this is how auto-invokers like /cf-plan, /cf-fix, /cf-optimize opt in without passing the flag). If absent or false, codex=false.
When codex=true, the workflow runs Claude's own review (Steps 2–6) and a Codex review in parallel, then merges both (Steps 6.5–7). The Codex script (run-codex-review.sh) auto-selects its scope from git state so it covers committed work, not just the working tree: feature branch → codex review --base <base> (committed branch changes); on the base branch with unpushed commits → --base <upstream> (local commits not yet pushed); only uncommitted changes → --uncommitted; local-only repo → --commit HEAD. This lets Codex see a phase's changes even after they are committed — including on the base branch, where gather-diff.sh only reports uncommitted work (its committed-vs-base section is gated on being on a feature branch), so Codex covers committed-on-base work that Claude's own review would otherwise miss. Trade-off: a --base/--commit scope omits uncommitted/untracked files, and when Codex is unavailable that committed-on-base work degrades to a Claude-only review that may see nothing.
Target compatibility: the auto-scope above only matches the default target (empty $ARGUMENTS, or a natural-language description that still reviews the default change set). If $ARGUMENTS (after stripping flags) is a file path or a commit range (e.g. HEAD~3..HEAD), Claude reviews that specific target but Codex would review the unrelated default change set. In that case do NOT run Codex — print:
⚠ --with-codex only applies to the default uncommitted-changes review; Codex does not support the target <target>. Running Claude-only review.
Set codex=false and skip Steps 2.5/6.5.
External headless-reviewer flags:
After the codex block, parse --claude, --gemini, --cursor, --grok (and --with-<agent> aliases). Collect matched ones into an agents=[…] list; strip the flags from $ARGUMENTS.
Host-match no-op: read HOST from the session bootstrap context. Only skip a matching flag when there is positive evidence of the in-session host: a HOST: line exists and its value matches the flag. Concretely, if a HOST: line is present and equals claude and --claude was passed, drop claude from the list and print:
⚠ --claude skipped: Claude is already the in-session reviewer.
If noHOST: line is present — e.g. running under a host that does not emit the coding-friend session bootstrap, such as Grok, Cursor, or Gemini — or its value is anything other than the requested flag, do NOT skip: run the requested reviewer. Never assume the in-session assistant is Claude just because a flag says --claude. When in doubt, run it: skipping wrongly loses a review the user explicitly asked for, whereas running an extra review is harmless.
When agents is non-empty, the workflow runs Claude's own review (Steps 2–6) and each requested agent review in parallel (Steps 2.5/6.5), then merges all surviving sources (Step 7). Each agent is invoked via run-agent-review.sh, which feeds gather-diff.sh output through build-review-prompt.sh and runs the headless CLI in read-only mode. The four agents review the exact diff gather-diff.sh produces — the same change set Claude's own review sees (deliberately different from Codex's auto-scope).
Target compatibility: same restriction as codex — only the default target (empty $ARGUMENTS, or a natural-language description). If $ARGUMENTS (after stripping flags) is a file path or commit range, print:
⚠ External reviewer flags only apply to the default uncommitted-changes review; they do not support the target <target>. Running Claude-only review.
Clear agents=[] and skip agent Steps 2.5/6.5.
--out flag (manual external-review round-trip):
If $ARGUMENTS contains --out, set out=true and strip the flag.
Mutual exclusivity: if --out is combined with any headless-agent flag (--claude/--gemini/--cursor/--grok/--codex), print:
⚠ --out (manual external review) can't combine with auto reviewer flags — ignoring the agent flags.
Clear agents=[], set codex=false.
--out implies the default target only (same restriction as codex); on a file-path/commit-range target, print the analogous warning and ignore --out (set out=false).
When out=true: run Claude's own review (Steps 2–6), then emit a /cf-review-out-style prompt file with Claude's findings embedded (see Step 6.7). Skip codex/agent Steps 2.5/6.5 and Step 7 merge. Display Claude's report to the user, then show the "📝 Review Prompt Ready" panel.
Step 2.5: Spawn Codex review in the background (only when codex=true)
Skip this step entirely when codex=false.
Resolve the docs root and a label YYYY-MM-DD-review. Use CF_DOCS_ROOT (the absolute docs base dir from the session bootstrap context) so the result file lands in the project docs folder even when Claude is launched from a subdirectory — do NOT use the bare docsDir name with a cwd-relative path. Fallback: if CF_DOCS_ROOT is unset, use $MAIN_REPO_ROOT/<docsDir>. Spawn the Codex review as a background Bash process (run_in_background: true) so Claude's own review (Steps 3–6) runs concurrently:
Do NOT wait for or inspect the Codex result here. Immediately proceed to Step 3 and run Claude's own review (Steps 3–6) while Codex runs in the background — that concurrency is the whole point. The harness automatically notifies Claude when the background process exits (do NOT poll, sleep, or check /tasks — per the Bash tool contract: "you'll be notified when it finishes"). The Codex exit status and result are checked later, in Step 6.5.
Spawn external agent reviews in the background (only when agents is non-empty):
Skip when agents=[] or out=true.
For each agent in agents, spawn a background Bash process (run_in_background: true):
The script prints KEY=value lines: FILES_CHANGED, LINES_CHANGED, SENSITIVE, CHANGED_FILES, and MODE.
Use the MODE value directly — no further calculation needed.
Full 5-layer review. All security phases, concise.
DEEP
>10 files OR >300 lines OR sensitive paths touched
Full 5-layer + extended security. Data flow tracing. Exploit scenarios.
If SENSITIVE > 0, always escalate to DEEP regardless of size.
Step 4: Gather context (conditional — based on review mode)
QUICK mode: Skip this step entirely.
STANDARD mode: Search memory only (if memory_search tool is available). Call memory_search with: { "query": "<area being reviewed — e.g. auth, API, database>", "limit": 5 }. Use results as context hints for the review.
DEEP mode: Launch the cf-explorer agent to understand callers, dependencies, and data flows around the changed files. Use the Agent tool with subagent_type: "coding-friend:cf-explorer". Pass:
Explore the codebase context around these changed files: [list changed files]
Questions to answer:
What calls these files/functions? (callers, entry points)
What do these files depend on? (downstream effects)
What conventions and patterns exist in the surrounding code?
Are there related tests that should be checked?
Note: cf-explorer already checks memory internally — do NOT call memory_search separately when using cf-explorer.
Memory and explorer results are hints — always verify against actual code.
Step 5: Read changed files
Read changed files in full — do not review only the diff, understand the context.
Step 6: Dispatch the cf-reviewer agent
Use the Agent tool with subagent_type: "coding-friend:cf-reviewer". Pass the full context:
Review mode: [QUICK | STANDARD | DEEP]
Diff:
[full diff from Step 2]
Changed files (full content):
[full content of each changed file from Step 5]
Context (if gathered in Step 4):
[memory search results or cf-explorer findings, if any]
Run the review now. Return the unified report in the 🚨/⚠️/💡/📋 format.
Wait for the agent to return its report.
Step 6.5: Collect & normalize the Codex review (only when codex=true)
Skip this step entirely when codex=false.
Wait for the Codex background process to finish. By the time the cf-reviewer agent (Step 6) returns, the harness has likely already delivered the Codex completion notification. If it has not yet arrived, wait for it — do NOT poll or sleep. Only read the result file after the process has exited.
Check the Codex exit status (from the background process — its CF_CODEX=... stderr line and exit code):
CF_CODEX=unavailable (exit 127) → Codex not installed. Print:
⚠ Codex unavailable (not on PATH) — proceeding without it.
Set codex=false and skip the rest of this step (Step 7 will use the cf-reviewer report as-is).
⚠ Codex review failed () — proceeding with Claude-only review.
Set codex=false and skip the rest of this step.
CF_CODEX=empty (exit 0, no result file) → nothing committed or uncommitted to review. Print:
⚠ Codex found no changes to review — proceeding with Claude-only review.
Set codex=false and skip the rest of this step.
CF_CODEX=ok <file> (exit 0) → the review was written to the result file; continue. (An optional CF_CODEX_SCOPE=... line on stderr records which scope was used.)
Normalize the raw Codex result into the standard 4-section format:
(Use the same ${CF_DOCS_ROOT}-based path the result was written to in Step 2.5.) This emits a ## 🔍 Codex Review block with findings tagged **[Codex]** and severities mapped ([P2]→⚠️, [P3]→💡, anything else incl. [P1]/[P0]→🚨 so a top severity never fails silent). It never drops content — unparseable output is folded into the Summary section.
Never block the workflow on Codex — any failure degrades gracefully to a Claude-only review.
Collect external agent reviews (only when agents was non-empty at spawn time):
Skip when agents=[] or out=true.
Wait for each agent's background process to finish (same wait contract as Codex — harness notification, no polling).
Check each agent's exit status (CF_AGENT=… stderr line and exit code):
CF_AGENT=unavailable (exit 127) → print:
⚠ <Agent> unavailable (not on PATH) — proceeding without it.
Drop that agent.
CF_AGENT=error (non-zero exit) → print:
⚠ <Agent> review failed (<reason from stderr>) — proceeding without it.
Drop that agent.
CF_AGENT=empty (exit 0) → print:
⚠ <Agent> found no changes to review — proceeding without it.
Drop that agent.
CF_AGENT=timeout (exit 124) → print:
⚠ <Agent> review timed out (>Ns) — proceeding without it.
Drop that agent. (N = review.agentTimeout from config, default 300.)
CF_AGENT=ok <file> (exit 0) → keep its result file. No normalize step — stdout is already CF-format.
Never block the workflow on any external agent — failures degrade gracefully.
Step 6.7: Emit --out prompt file (only when out=true)
Skip when out=false.
After Step 6 returns Claude's report, emit a /cf-review-out-style prompt with Claude's findings embedded:
Write Claude's Step 6 report to a temp file.
Build the prompt file:
# CF_EMBED_CONTEXT_FILE must be set on the build-review-prompt.sh stage (it reads# the var) — NOT as a pipeline-leading prefix, which would only reach gather-diff.sh.mkdir -p "${CF_DOCS_ROOT}/reviews" && \
bash "${CLAUDE_PLUGIN_ROOT}/skills/cf-review/scripts/gather-diff.sh" | \
CF_EMBED_CONTEXT_FILE="$tmp_report" \
bash "${CLAUDE_PLUGIN_ROOT}/skills/cf-review-out/scripts/build-review-prompt.sh" \
"<label>""<docsDir>" \
> "${CF_DOCS_ROOT}/reviews/<label>-prompt.md"
Show the same "📝 Review Prompt Ready" panel, copy-paste external-agent instruction, and the > When all external agents finish, run /cf-review-in <label> to collect all results. reminder that cf-review-out Step 4 shows.
Display Claude's own report to the user before the panel so they see the findings that were embedded.
Skip Steps 7–10's normal completion banner — the review was exported, not completed in-session. Proceed to Step 8 (mark reviewed) if appropriate, then stop.
Step 7: Collect the report
Skip when out=true (Step 6.7 handles output).
When no external sources survived (codex=false and all agents dropped): the result of Step 6 is the final formatted report (🚨 Critical / ⚠️ Important / 💡 Suggestions / 📋 Summary). Do NOT reformat or restructure it — use it as-is in Step 10.
When any external source survived (codex and/or one or more agents): merge all reviews through the reducer. Dispatch the cf-reviewer-reducer agent (Agent tool, subagent_type: "coding-friend:cf-reviewer-reducer") with Claude's report as Source 1 and each surviving external review as an additional numbered source:
Merge these review reports into one unified, deduplicated, severity-ranked report.
Source 1 — Claude multi-agent review:
[the full report from Step 6]
Source K — <Agent> review: (for each surviving external source)
[the review content — normalized Codex block for Codex; raw result file for claude/gemini/cursor/grok agents]
Tag each external source's findings with [\<Agent\>] provenance (Codex → [Codex], gemini → [Gemini], etc.). Where multiple sources flag the same file:line for the same issue, merge into one finding (keep highest severity) and note agreement (raises confidence). Output the standard 🚨/⚠️/💡/📋 format.
Use the reducer's merged output as the report in Step 10.
Step 9: Smart capture (conditional — only if memory_store MCP tool is available)
If the review found architectural insights or recurring patterns worth preserving, call memory_store with:
type: "fact"
importance: 3
source: "auto-capture"
title/description/tags/content summarizing the insight
Skip if the review was routine with no notable findings.
Step 10: Final output
Display the full report followed by the status banner in a single message.
IMPORTANT: The structured report from step 8 and the banner below MUST appear together in the same final response. Do NOT split them across separate messages. This ensures the complete review is visible in the last message.
Display the cf-reviewer's report first, then append the appropriate banner. When any external source contributed, add a · Reviewed by: <in-session> + … suffix listing the in-session reviewer plus each external agent/codex that actually contributed (e.g. · Reviewed by: Claude + Codex + Gemini). Label the in-session reviewer by the HOST value from the session bootstrap context, capitalized (claude → Claude, codex → Codex); if no HOST: line is present, use the neutral label In-session AI — do NOT hardcode Claude, since the assistant running this skill may be another host (Grok, Cursor, …). Omit the suffix when only the in-session reviewer ran.
Skip this step's banner when out=true — Step 6.7 already showed the export panel.
