Rick Review

Rick Review

Multi-agent code review. Up to seven Rick & Morty council agents review in parallel — three always run, four are routed by what the diff actually touches. You orchestrate, run three skepsis passes (cite → relevance → intent), and write a versioned report.

Quick Start

/rick-review            # reviews current branch / staged changes
/rick-review 213        # reviews PR #213
/rick-review --committed-only  # force committed-only scope even if working tree is dirty

Branches

Four implicit branches. Each short-circuits later steps; read this map first so you don't ambush yourself with a step that no longer applies.

• Mode: PR ($ARGUMENTS=<number>) vs current-branch. PR mode triggers Step 1's branch auto-swap and Step 3.5(b)'s PR-body extract.
• Council size: rickest rick (small, non-sensitive diff) vs full council. Rickest skips Step 5.5 deny-list and collapses Step 6's contradiction check.
• Version: N=1 vs N≥2. N=1 skips Step 5.5 (no prior kills to echo).
• Intel: with vs without rick-intel dossier (Step 1.5). Affects the Step 8 header line and Step 7.7's intent cross-reference.

Labels

Canonical label vocabulary — Step 5, Step 8, Step 4.7, and OUTPUT.md all reference these. Don't redefine.

Step 1 — Detect Context

git branch --show-current
git status --short
gh pr view --json number,title,headRefName,baseRefName 2>/dev/null

If $ARGUMENTS starts with a number: gh pr view ${ARGUMENTS%% *} --json number,title,headRefName,baseRefName (the %% * drops trailing tokens like --committed-only that Step 4 reads separately; passing the whole $ARGUMENTS makes gh error on unknown flags). Capture baseRefName as $BASE. Stacked PRs diff against their direct parent, not main.

Branch auto-swap (PR mode only): if headRefName from the PR differs from the current local branch, run gh pr checkout {number}. This fetches the PR head (handles forks correctly) and switches to it; git refuses the checkout when uncommitted work would be clobbered, so surface that error and stop if it fails. On success, print Switched to {headRefName} for PR #{number} and continue. Never silently review the wrong branch.

If no PR: detect default branch with git symbolic-ref refs/remotes/origin/HEAD | sed 's@^refs/remotes/origin/@@'. Fallback to main if that fails. Use as $BASE.

Step 1.5 — Intel Lookup

Check for an intel dossier per ../rick-intel/LOOKUP.md. If one exists, read its Story Context + Scope. Hold the AB-id, parent epic title, and one-line scope summary for inclusion in the report header (Step 8). The council agents review the diff, not the story — don't inject intel into agent prompts. No intel? Skip silently; the report header omits the story line.

Step 2 — Resolve REVIEW_NAME

See RESOLVE-FOLDER.md. Same folder identity threads through /rick-plan, /rick-save, /rick-fix, /rick-review-comments, /rick-recap.

Step 3 — Output Path and Version

Canonical report: docs/rick/<REVIEW_NAME>/review/current.md (overwritten each run).

Versioning: list docs/rick/<REVIEW_NAME>/review/v*.md, parse the integer from each v<N>.md (ignore -pre suffixes), N = max integer + 1. Sort by parsed integer, not lexically — v10.md sorts before v2.md and tail -1 gives v9.md at v10+, silently overwriting an existing snapshot. If no prior versions, this is v1.

Output files (full list and shape in Step 8 + OUTPUT.md): current.md, VERSIONS.md, v<N>.md, agents/v<N>/<agent-slug>.md. Create docs/rick/<REVIEW_NAME>/review/agents/v<N>/ before launching agents.

Step 3.5 — Gather System Context

Before launching agents, collect context they can't safely infer. Each subsection is gated — skip when its trigger doesn't fire. Section (a) contributes to $SYSTEM_CONTEXT (prepended to every agent's prompt in Step 5). Section (b) is held as $PR_INTENT for Step 7.7's cross-reference only.

a) Settled FIXED findings — only when prior versions exist

SETTLED_FIXED=$(awk '/✓ FIXED/' docs/rick/<REVIEW_NAME>/review/v*.md 2>/dev/null | sort -u)

If non-empty, wrap in this block and hold as $SYSTEM_CONTEXT:

Settled decisions from prior reviews of this branch (do NOT recommend reverting without new evidence):
<contents of $SETTLED_FIXED>

b) PR description intent — only when a PR was detected in Step 1

PR bodies routinely explain why something looks missing — scope splits ("except seed script"), deferred validation, intentional trade-offs. Read once upfront so Step 7.7 can cross-reference. Do not inject into agent prompts — PR narratives anchor agents on validating stated intent instead of scanning for anomalies the author didn't call out. Same reason Step 1.5 keeps intel out of agent prompts.

PR_BODY=$(gh pr view --json body --jq '.body' 2>/dev/null)

If non-empty, extract verbatim text (no rewriting) from these sections:

• ## Summary
• ## Security Notes / ## Security Considerations
• ## Performance Notes / ## Performance Impact
• ## Testing Instructions / ## Testing
• ## Deployment Notes / ## Rollback Plan
• ## Additional Context
• Any header line matching WORK ITEM:

Hold as $PR_INTENT for Step 7.7 only. If no PR was detected, $PR_INTENT stays empty and Step 7.7 skips the PR body block.

Step 4 — Build diff slices

Produce /tmp/rr-files.txt plus four per-agent slice files (broad, data, api, sec). Full procedure (scope detection, slices, slice cap) in SLICING.md.

Completion: /tmp/rr-files.txt and /tmp/rr-diff-{broad,data,api,sec}.txt exist; $BASE_REF and $SCOPE_LABEL set.

Step 4.5 — Shell Pre-Scan

Catch mechanically obvious nits before any agent fires. See PRE-SCAN.md for the pattern catalog (console.*, // TODO/FIXME, @ts-ignore, as any) and the awk script that produces $PRESCAN. Hits go two places: appended directly to the P3 table in Step 7, and wrapped in this block held as $PRESCAN_FINDINGS (prepended to every agent's prompt in Step 5, so agents don't burn tokens re-searching for what shell already caught):

Shell pre-scan already caught these — the orchestrator will file them at P3. Do NOT re-flag:

<contents of $PRESCAN, table rows verbatim>

Step 4.7 — Rickest Rick Mode

The full council is overkill on a 10-line bug fix. Rickest rick mode routes small, non-sensitive diffs to one Rick C-137. "Rickest" = the smartest Rick across the multiverse, not the fastest: one top-tier Rick covers every lens better than three generalist reviewers plus specialists.

Run the rickest gate per SLICING.md. True when: diff ≤ 50 lines AND ≤ 5 files AND specialized slices (data/api/sec) all empty.

When $RICKEST=true:

• Step 5 launches only Rick C-137.
• Step 5.5 skipped (single agent, nothing to echo).
• Step 6's cross-agent contradiction check collapses (the cite check still runs).
• Other six council slots render as Skipped (rickest rick) per Labels.

Completion: $RICKEST set; $SCOPE_LABEL suffixed if true.

Step 5 — Route the Council (parallel)

If $RICKEST=true from Step 4.7: launch only the Rick C-137 row below. Treat every other row as Skipped (rickest rick). Skip Step 5.5. Context assembly still runs — Rick C-137 gets the same substitutes and prepends, just solo.

Read AGENTS.md for each agent's full prompt template. Two mechanics: substitute {} placeholders into the template, prepend $ blocks in front of it. All triggered agents get the same context regardless of slice.

Substitute (placeholders declared in AGENTS.md):

• {CHANGED_FILES_LIST} — full /tmp/rr-files.txt
• {SCOPED_DIFF} — the agent's slice file from the table below
• {CLAUDE_MD_CORE} — CLAUDE.md excerpt (read once, inject all agents)
• {REVIEW_NAME} — from Step 2
• {SCOPE_LABEL} — from Labels

Prepend (blocks bolted onto the front of the prompt, each when non-empty):

• $SYSTEM_CONTEXT — Step 3.5(a) settled FIXED (revert-prevention)
• $PRESCAN_FINDINGS — Step 4.5 shell-verified nits (broad-slice agents must NOT re-scan for these)
• $DENY_LIST block — Step 5.5, N≥2 full council only (kill-prevention)

Model. Omit the model field on every Agent call — all agents inherit the parent session's model.

Launch all triggered agents in a single batched Agent tool call with subagent_type: general-purpose. Untriggered agents → Skipped (out of scope) (per Labels), not Clean.

Step 5.5 — Cross-version Deny-list (N >= 2, full council only)

Skipped when N=1 or $RICKEST=true. Otherwise, gather every prior "Killed in Step 7.5" bullet so agents stop re-raising findings the relevance check already discarded. Agents are stateless across versions: on deren-starter's 63-sentry-wiring branch, v11 through v14 each re-killed the same hedge-gated finding, burning council budget every round.

Kill-prevention lens (don't re-raise parasites). Step 3.5(a) is the revert-prevention lens (don't undo FIXED decisions). Both ship, they catch different misbehaviors.

DENY_LIST=$(awk '/^### Killed in Step 7\.5/{flag=1; next} /^#+ /{flag=0} flag && /^- /' docs/rick/<REVIEW_NAME>/review/v*.md | sort -u)

Empty? Skip the preamble. Otherwise prepend exactly this block to every agent's prompt, before the Step 5 placeholders:

Prior reviews of this branch killed the following findings via the Step 7.5 relevance check. Do NOT re-raise them unless the cited code has materially changed since the kill. Hedge-gated kills are particularly likely to look novel on re-read; they are not.

<contents of $DENY_LIST, bullets verbatim>

sort -u dedupes kills that recurred across multiple prior versions. The awk stop condition /^#+ / catches any subsequent heading at any level.

Step 6 — Skepsis: cite check

Every finding is a parasite until proven real. First skepsis pass: does the cite match the code as it stands today?

Short-circuit: every agent returned Clean → skip to Step 7.

Pre-existing code filter (mandatory). For each finding, locate the cited line in the diff slice the agent reviewed:

• Appears as a + (added) or -/+ pair (modified) → in scope.
• Appears as context (no prefix) or absent from the slice → existed identically on $BASE → kill.
• Whitespace-only changes don't count as "changed."

Verification budget. Apply the per-severity budget and grep recipe in PARASITE-CHECK.md. When $RICKEST=true, the budget still applies — only the cross-agent contradiction step (P0 check 4) is vacuous.

Completion: every finding marked REAL / PARASITE / DOWNGRADE; parasites removed silently from the report (don't list them).

Step 7 — Aggregate

1. Dedupe findings only when (a) file matches AND (b) line ranges overlap or are equal (treat 47 and 47-50 as overlapping) AND (c) the Issue text names the same defect. Different defects at the same line stay separate. Mark anything flagged by 2+ agents with ** — cross-cutting confirmation.
2. Group P0 / P1 / P2 / P3 (CRITICAL / HIGH / MEDIUM / LOW).
3. Within each tier, sort cross-cutting (**) rows to the top.
4. Append $PRESCAN rows (Step 4.5) directly to the P3 table — shell-verified, no parasite check needed. If a row's file:line exactly matches an agent-reported P3 with the same defect, drop the duplicate (keep the shell-verified one — its citation is mechanically guaranteed).

Step 7.5 — Skepsis: relevance check

Step 6 answered "is the cite real." This step answers "does it matter in production today, before any future change?"

Auto-keep. Any row with ** (cross-cutting) survives. Skip checks.

For each non-** row, run the primary check first; auto-kills are mechanical fallbacks.

Check 1 — Impact restatement (primary). Restate the finding in one sentence answering "what fails in production today, before any future change?" If the honest restatement needs "would," "could," "if," "after" → demote one tier. P3 demoted here is killed.

Auto-kill 1 — Self-defeat scan. Issue text contains "already covered," "tested elsewhere," "not a real defect," "documentation-level," "structurally cannot" → KILL. The agent told you it's not real.

Auto-kill 2 — Hedge scan. Issue text contains "if a future edit," "in case someone changes," "for symmetry," "for parity," "for completeness," "hypothetical" → demote one tier. P3 demoted is killed. Don't add present-tense consequence phrases ("could silently," "after a refactor") to this list — those describe real-today bugs, Check 1 handles them.

Re-number rows after kills and demotions. Tier counts Step 8 prints to chat and writes to VERSIONS.md are these post-7.5 numbers, not Step 7's raw aggregate. Append in the report:

### Killed in Step 7.5 (real, but didn't matter)
- [one-line summary]: [which check killed it]

(The exact ### Killed in Step 7.5 heading is the anchor Step 5.5's deny-list awk script greps for. Don't rename it.)

Completion: every non-** row evaluated; kills and demotions applied; sub-heading appended (or omitted when no kills).

Step 7.7 — Skepsis: intent check

Final skepsis pass: does the surviving finding contradict documented intent — the PR description body ($PR_INTENT from Step 3.5(b)) or the rick-intel dossier (Step 1.5)?

Skip if neither source exists.

See INTENT-CROSS-REF.md for the full demote / strengthen / mitigation / scope-flag rules across both sources.

Completion: every P0/P1 survivor compared against $PR_INTENT and the intel dossier; row either annotated ((addressed in PR body), (out of scope per intel), (strengthened — PR body promised X)) or left intact. Tier counts Step 8 prints are post-7.7.

Step 8 — Write Report and Print

Verdict: NEEDS FIXES if any P0 or P1 survives Step 7.5. Otherwise APPROVED.

Files to write (single batch)

All independent — write in one assistant response with parallel tool calls:

• docs/rick/<REVIEW_NAME>/review/current.md (canonical)
• docs/rick/<REVIEW_NAME>/review/v<N>.md (snapshot of the canonical)
• docs/rick/<REVIEW_NAME>/review/agents/v<N>/<agent-slug>.md (one per ran agent)
• Append to docs/rick/<REVIEW_NAME>/review/VERSIONS.md via the Edit tool — never Write (Write would clobber v1…v(N-1) history if the payload contains only the new line):

- v<N> | <YYYY-MM-DD> | <SCOPE_LABEL> | P0:<n> P1:<n> P2:<n> P3:<n> | <APPROVED|NEEDS FIXES>

Canonical report shape

See OUTPUT.md for the full structure and VERSIONS.md entry format. The Agents column uses the Labels vocabulary from the top of this skill.

Print to chat

Reviewed <REVIEW_NAME> (v<N>): P0:<n> P1:<n> P2:<n> P3:<n>. Verdict: <APPROVED|NEEDS FIXES>.
Report: docs/rick/<REVIEW_NAME>/review/current.md

Then stop. No "want me to start fixing the P0s." That's rick-fix's job.