| name | risk-gate |
| description | Triages a diff/PR/branch by change risk and escalates only the risky parts to deeper review. Runs a cheap deterministic scoring pass (danger paths like auth/payments/migrations, change size, blast radius, churn) to bucket changed files Low/Medium/High, then escalates only HIGH-risk files to a heavyweight separate-model review (review-board or codex-second-opinion). Use when user says "risk gate", "what's risky here", "triage this diff", "is this safe to merge", or invokes /risk-gate. This is the cheap entry point of the review ladder — start here when unsure how much review a change deserves. |
| user-invocable | true |
| argument-hint | [PR number | branch | (default: working diff)] |
| allowed-tools | Bash, Read, Write, Glob, Grep, Agent |
Risk Gate
A triage-then-escalate conductor. Most changes are low-risk and don't need a
heavyweight multi-model audit; a few are dangerous and deserve everything you've
got. Risk Gate spends the expensive review only where it pays by scoring the
diff cheaply first.
risk = likelihood-of-failure × impact-of-failure. The score approximates both:
impact from where the change lands (auth, payments, migrations, money/LLM
spend), likelihood from how much/complex the change is (churn, size, blast
radius).
Step 1: Resolve the target & score it (cheap, deterministic)
Resolve the diff base (PR / branch / working tree — see review-board for the
same resolution), then run the scorer:
python3 "${CLAUDE_PLUGIN_ROOT}/skills/risk-gate/scripts/risk_score.py" [BASE_REF]
It emits JSON: per changed file an added/removed count, a score, a bucket
(high/medium/low), and the reasons that fired. The signals and their
weights are documented in
${CLAUDE_PLUGIN_ROOT}/skills/risk-gate/references/risk-signals.md — read it if
you need to explain or tune a score. Weights are heuristic defaults, not law:
treat the buckets as a starting point and apply judgment in Step 2.
If there are no changes, stop and say so.
Step 2: Confirm the buckets (judgment pass)
The script is keyword/size based and will mis-rank some files. Skim the diff for
each medium/high file and adjust:
- Promote a
low/medium file to high if it touches an invariant, a security
boundary, or irreversible state the keywords missed.
- Demote a
high file that's a rename, comment, generated artifact, or pure test
churn.
State the final bucketing and why anything moved.
Step 3: Escalate ONLY the high-risk set
-
No high files → no escalation. Report the risk map, recommend a normal
review, stop. (This is a valid, common outcome — don't manufacture risk.)
-
Some high files → escalate just those files to deep review. Prefer,
in order of availability:
- The
review-board skill (multi-lens + Codex), scoped to the high-risk files.
- Else, the
codex-second-opinion skill (separate-model) if the Codex CLI is present.
- Else, spawn your own parallel Claude lens audit over the high-risk files.
Pass the deep reviewer only the high-risk file list so it doesn't waste the
expensive run on trivia.
TEMPLATE / TUNING: the escalation threshold and which deep reviewer to call
are the two knobs you'll adjust. Default threshold = any high file.
Step 4: Report
Write ./.risk-gate-out/RISK-REPORT.md:
- Risk map — table:
| File | Bucket | Score | Why |, sorted high→low.
- Escalated — which files went to deep review and the reviewer's verdict.
- Gate decision — one line:
proceed / review-then-proceed / block until fixed.
Print the map + decision to the user.
Error handling
| Scenario | Action |
|---|
| No changes | Stop; report empty diff. |
risk_score.py errors | Surface stderr; fall back to a manual skim and say the score was unavailable. |
| No deep reviewer available (no review-board / Codex) | Do the Claude lens audit yourself; note reduced model diversity. |
| Deep review finds nothing on a high-risk file | Keep it flagged review-then-proceed; a clean deep review lowers, not erases, risk. |
