rb-triage

name	rb:triage
description	Triaging review findings: pick fix now, skip, or defer. Prioritizes Blockers (Iron Law violations, security). Triggers: "triage", "which to fix", "prioritize findings".
argument-hint	[path to review file]
effort	low

Triage Review Findings

Batch-select findings from the review before creating follow-up work.

Iron Laws

1. Auto-include all Iron Law violations - these are never optional
2. Separate safety fixes from style preferences - prioritize data integrity
3. Group related findings - fix together what belongs together
4. Defer performance optimizations without profiling data
5. Use structured multi-select, not freeform chat prompts — when asking the user which findings to keep, use a checkbox-style selection UI

Workflow

Review File
    ↓
Parse Findings (Blocker / Warning / Suggestion)
    ↓
┌─────────────────┐
│ Auto-Categorize │
├─────────────────┤
│ • Blocker       │ → always include (auto, never optional)
│ • Warning       │ → recommend (default selected; user may defer)
│ • Suggestion    │ → defer (default unselected; user may include)
└─────────────────┘
    ↓
Present Grouped Findings
    ↓
User Selects Items
    ↓
Write Triage Output
    ↓
Next Action

Severity Classification

Map review buckets (Blocker | Warning | Suggestion) to triage priorities. Within each bucket, order security findings only by evidence_mode (defined in ${CLAUDE_PLUGIN_ROOT}/skills/security/SKILL.md § "Evidence Mode" and ${CLAUDE_PLUGIN_ROOT}/agents/security-analyzer.md § "Evidence Mode (mandatory)"):

1. runtime-confirmed — act first
2. configuration-risk
3. static-signal
4. requires-human-validation — surface to user for decision

Non-security findings carry no evidence_mode. Retain agent-emitted order within the bucket.

Blocker → Always Include

• Iron Law violations: transaction safety, after_commit discipline, JSON-safe args
• Security issues: SQL injection, XSS, mass assignment, hardcoded secrets
• Data integrity: missing constraints, race conditions, incorrect null handling
• N+1 queries: database performance killers
• Transaction safety: operations outside transactions that should be inside

Warning → Recommend Include

• Error handling (non-Iron-Law): missing error handling on external API calls. rescue Exception / rescue ::Exception / rescue_from(Exception) / rescue_from ::Exception is Iron Law 18 → Blocker. Bare rescue defaults to StandardError, not a Law 18 violation.
• Test coverage: missing tests for critical paths
• API contract: breaking changes to public APIs
• Documentation: missing docs for public methods
• Refactoring opportunities: complex methods, Law of Demeter violations
• Naming: unclear variable/method names

Suggestion → Defer by Default

• Code style: inconsistent formatting (if not auto-fixable)
• Performance: without profiling evidence
• Micro-optimizations: minor efficiency gains
• Subjective preferences: personal style differences
• Future-proofing: speculative abstractions
• Comments: missing or misleading comments

Triage Process

Step 1: Load Review File

Argument resolution FIRST. Triage operates ONLY on a consolidated review artifact (no plain description / no feature input):

$ARGUMENTS shape	Action
empty / absent	Resolve newest consolidated review: glob .claude/reviews/*-*.md, EXCLUDE .provenance.md sidecars (suffix -{datesuffix}.provenance.md) — accept only files whose basename ends in -{datesuffix}.md with no .provenance segment. Direct children of reviews/ only (no per-reviewer subdirectory). Pick the file with the most recent mtime. If none found, STOP. Print: No consolidated review artifact found under .claude/reviews/. Run /rb:review first.
.claude/reviews/{review-slug}-{datesuffix}.md (direct child of reviews/)	proceed
.claude/reviews/{agent-slug}/{review-slug}-{datesuffix}.md (per-reviewer artifact under subdirectory)	STOP. Print: Path is a per-reviewer artifact, not the consolidated review. Run with the consolidated path: .claude/reviews/{review-slug}-{datesuffix}.md
.claude/reviews/{review-slug}/RUN-CURRENT.json (manifest)	STOP. Print: Path is the manifest, not the consolidated review.
anything else (plain text, non-review path)	STOP. Print: /rb:triage operates on a consolidated review artifact. Run /rb:plan {description} for feature planning, or /rb:review first to produce a review artifact.

After argument resolution passes, read the consolidated review markdown.

Source section	Extract
Preamble	**Date**:, **Complexity**:, **Files Changed**:, **Reviewers**: (metadata)
## Reviewer Coverage	Per-reviewer recovery state + finding counts
## Reviewer Verdicts	Per-reviewer raw + canonical verdicts
## Summary (then **Verdict**: line immediately AFTER the Summary table per ${CLAUDE_PLUGIN_ROOT}/skills/review/references/review-playbook.md § "Consolidated Review Format")	Severity counts + consolidated verdict
## At-a-Glance Finding Table	Per-finding row: # / Finding / Severity / Confidence / Reviewer / File / New?
## Blockers ({n}) / ## Warnings ({n}) / ## Suggestions ({n})	Detailed finding bodies: **File**:, **Reviewer**: ... | **Confidence**:, **Issue**:, **Why it matters**: (Blockers only), **Current**: + **Suggested**: (Blockers code blocks), **Recommendation**: (Warnings) or **Suggestion**: (Suggestions)
## Pre-existing Issues (unchanged code)	Findings to surface in plan's ## Pre-existing Issues (informational) only
## Test Coverage Gaps ({n}) (REQUIRES CHANGES verdict only)	One row per uncovered NEW public surface; columns # / Surface / File / Why uncovered / Suggested test. Step 2b + Step 5 auto-include each row as a Phase 1 [test] task.

For NEW At-a-Glance rows (New? = Yes), match each row to its detailed finding by the Finding column text — the At-a-Glance Finding cell MUST equal the ### N. {Title} heading under the appropriate ## Blockers / ## Warnings / ## Suggestions section verbatim. Use (File, Reviewer, Severity) as a corroboration tuple. Title text disambiguates the case where one reviewer reports two findings of the same bucket against the same path; the tuple alone is ambiguous in that case. The Iron Law label for THIS NEW row's plan task line comes from the issue text combined with the iron-laws.yml registry (e.g., "Multi-step operations without transaction wrap" → Iron Law 5).

Pre-existing At-a-Glance rows (New? = Pre-existing) have NO backing ### N. {Title} heading — ## Pre-existing Issues (unchanged code) lists them as bullets, not headings. The Finding cell carries a concise description of the unchanged-code issue. Surface them in the plan's ## Pre-existing Issues (informational) section unchanged; do NOT attempt the title-verbatim match. Pre-existing findings NEVER produce a plan task line, so NO Iron Law label / [annotation] / [Pn-Tm] is derived for them — Step 5 emits them as bullets, not as - [ ] checkboxes (per Step 2 + Step 5 contract).

The plan's [annotation] value MUST be one of the canonical Set A entries per ${CLAUDE_PLUGIN_ROOT}/skills/plan/references/planning-workflow.md § "Plan Generation":

Annotation	Use for
[direct]	most common — generic Ruby work, refactors, glue
[active record]	models, migrations, ORM-touching findings
[hotwire]	Turbo Streams, Frames, Stimulus, broadcasts
[sidekiq]	jobs, queues, Sidekiq-specific findings
[concurrency]	threads, fibers, async, race conditions
[security]	auth, SQL injection, XSS, secrets, authorization
[test]	spec / test additions or fixes

Do NOT emit [ruby], [testing], [grape], [ar], [sequel], [perf], or [general-purpose] — those are descriptive narrative labels (or subagent type names), not plan-task annotations, per planning-workflow.md § "Plan-Task Annotations Cross-Reference". Do NOT use the underscore form from iron-laws.yml category: field. /rb:work parser routes on the canonical Set A only.

Step 2: Auto-Categorize via Review Bucket

Read each finding's bucket (Blocker | Warning | Suggestion) AND its New? column. Apply this filter FIRST:

• New? = Pre-existing → informational only; never auto-include, never offer in selection UI. List under a separate "Pre-existing Issues" section in the fix plan for context.
• New? = Yes → proceed to bucket-based categorization below.

For NEW findings:

• Blocker → always include (auto-selected, never optional)
• Warning → recommend (default selected; user may defer)
• Suggestion → defer (default unselected; user may include)

Step 2b: Verdict gate + Phase 1 auto-include source

Consolidated **Verdict**:	Phase 1 source (auto-include)
BLOCKED	every NEW Blocker row from At-a-Glance + matched ## Blockers detail
REQUIRES CHANGES	every row from ## Test Coverage Gaps ({n})
PASS WITH WARNINGS / PASS	none — Phase 1 heading omitted entirely

BLOCKED + REQUIRES CHANGES are mutually exclusive per playbook STEP 4 (BLOCKED requires blockers > 0; REQUIRES CHANGES requires blockers == 0). One source populates Phase 1, never both.

Gap row → Phase 1 task line shape (REQUIRES CHANGES only):

- [ ] [P1-T{n}][test] Add spec for {Surface} — test-coverage gap
  (REQUIRES CHANGES); source {review-path}

/rb:plan {review-path} is the gaps-only entrypoint (no selection UI). /rb:triage is the default; covers all mixed cases.

Step 3: Group Findings

Group NEW findings by [file, bucket] so batch-fixable items appear together in the selection UI. The review artifact does NOT carry a stable rule_id; do not invent one.

Step 4: Present Multi-Select UI

Use AskUserQuestion with multiSelect: true. Auto-include rules per Step 2b verdict gate:

Verdict	Auto-include (no UI)	Present in selection UI
BLOCKED	every NEW Blocker (B<n>)	NEW Warnings (W<n>), NEW Suggestions (S<n>)
REQUIRES CHANGES	every Test Coverage Gap row (G<n>)	NEW Warnings (W<n>), NEW Suggestions (S<n>)
PASS WITH WARNINGS / PASS	none	NEW Warnings (W<n>), NEW Suggestions (S<n>)

Never surface pre-existing findings in the selection UI. Drive selection entirely through AskUserQuestion options; never accept freeform-typed commands. Present this option set in the multi-select prompt, bucket shortcuts first then individual items:

Option label	Effect
All Warnings	Select every NEW Warning finding. Leave Suggestion selections untouched.
All Suggestions	Select every NEW Suggestion finding. Leave Warning selections untouched.
Skip all Warnings	Defer every NEW Warning finding. Leave Suggestion selections untouched.
Skip all Suggestions	Defer every NEW Suggestion finding. Leave Warning selections untouched.
Group by file	Re-render the option list grouped by file. Do NOT change selection state.
W<n> / S<n> (one row per finding)	Select the individual NEW Warning / Suggestion. Description: file, line, one-line reason.

Accept combinations of shortcuts and individual picks in one response (e.g., All Warnings + S2 + S5). Resolve combinations in selection order: cancel an earlier All Warnings when a later Skip all Warnings follows; let individual W<n> / S<n> picks override their bucket-level skip. Split into multiple screens when

6 selectable items.

Summarize state back, e.g.: auto-included: B1 B2 B3 | selected: W1 | deferred: W2 S1 S2 or for REQUIRES CHANGES: auto-included: G1 G2 | selected: W1 | deferred: W2 S1.

Step 5: Write Triage Output

Save the plan to .claude/plans/{slug}/plan.md using the canonical template at ${CLAUDE_SKILL_DIR}/references/triage-plan-template.md. Emit the template's full surface in this order:

1. # Plan: {slug} heading.
2. **Status**: IN_PROGRESS / **Created** / **Last Updated** preamble (per ${CLAUDE_PLUGIN_ROOT}/skills/work/references/file-formats.md so /rb:work parses + resumes correctly).
3. ## Metadata (Source Review, Generated By, Triaged By).
4. ## Summary table (Bucket / Total / Selected / Deferred / Excluded).
5. Phase 1 (auto-include; heading per Step 2b verdict gate):
   • ## Phase 1: Blockers [PENDING] (BLOCKED verdict) — task lines - [ ] [P1-Tn][annotation] ... per NEW Blocker.
   • ## Phase 1: Test Coverage Gaps [PENDING] (REQUIRES CHANGES verdict) — task lines - [ ] [P1-Tn][test] ... per gap row.
   • Phase 1 heading OMITTED entirely for PASS / PASS WITH WARNINGS verdicts (no auto-include source).
6. ## Phase 2: Warnings (selected) [PENDING] — task lines - [ ] [P2-Tn][annotation] ... (only NEW Warnings the user selected at Step 4). Omit the heading when zero warnings selected.
7. ## Phase 3: Suggestions (selected) [PENDING] — task lines - [ ] [P3-Tn][annotation] ... (only NEW Suggestions the user selected via S<n> at Step 4). Omit the entire phase heading when zero suggestions selected.
8. ## Deferred Findings (NEW Warnings + Suggestions the user did NOT select; excluded entirely from any Phase).
9. ## Pre-existing Issues (informational) — every New? = Pre-existing finding from Step 1, file + line + reviewer + one-line note. NEVER emit as - [ ] task lines.
10. ## Next Steps — /rb:work invocation, time estimate, deferred-review note.

Decision Tree

Read the review's existing bucket per finding. Override only when the bucket is missing or ambiguous:

Review bucket = Blocker?
├── YES → always include
│         └── never optional
│
└── NO → bucket = Warning?
          ├── YES → recommend (default selected; user can defer)
          │
          └── NO → bucket = Suggestion?
                    ├── YES → defer (default unselected; user can include)
                    │
                    └── NO → bucket missing / ambiguous → ask user

Batch Operations

Select All in Bucket

Blockers are auto-included by Step 4; do NOT surface a "Select All Blocker" shortcut in the selection UI. Offer only Warning / Suggestion + cross-cutting filters:

[All Warnings]
[All Suggestions]
[All in File: app/models/user.rb]
[All of Type: N+1]

Smart Grouping

Bucketize findings via reasoning, not by running code. Step 3 already pins the primary key: group NEW findings by (file, bucket). Review artifacts carry no stable identifier beyond that tuple — do NOT invent one.

Grouping rule	Effect on selection UI
Same (file, bucket) with ≥ 2 occurrences	Surface together; recommend fix_together
Findings touching the same domain layer (e.g., 2+ in app/models/*)	Optional cross-cutting shortcut (e.g., "Select All in app/models/"); recommend fix_sequentially
Different buckets in same file	Keep separate — auto-include / recommend / defer follows the bucket, not the file

Integration with Workflow

From Review

# After review completes
/rb:review app/models/user.rb
# → Creates `.claude/reviews/{review-slug}-{datesuffix}.md`

# Then triage
/rb:triage .claude/reviews/fix-user-model-20260505-123000.md
# → Creates .claude/plans/fix-user-model/plan.md

# Then work on selected items
/rb:work .claude/plans/fix-user-model/plan.md

Direct to Work

If triage generated a plan with selected Blocker items:

/rb:work .claude/plans/fix-user-model/plan.md
# → Works on the selected findings as normal plan tasks

Edge Cases

Empty Selection

If user selects nothing:

No findings selected. Options:
[Review All Findings Again]
[Create Note: "Reviewed, no action needed"]
[Cancel]

All Blockers

If all findings are Blockers:

All 5 findings are Blockers and must be addressed.
Estimated time: 2 hours

[Start Work]

[Start Work] invokes /rb:work against the generated plan.

Contradictory Findings

If findings conflict:

⚠️ Warning: Contradictory findings detected

Finding A: "Extract method to reduce complexity"
Finding B: "Inline method for clarity"

These may conflict. Review together?
[Review Both] [Skip Both]

Output Format

Sole output: .claude/plans/{slug}/plan.md. Tasks MUST use the canonical - [ ] [Pn-Tm][annotation] ... format from ${CLAUDE_PLUGIN_ROOT}/skills/work/references/file-formats.md + canonical Set A annotations from ${CLAUDE_PLUGIN_ROOT}/skills/plan/references/planning-workflow.md § "Plan Generation". /rb:work parser routes on this format and resumes via --from Pn-Tm.

Example for BLOCKED verdict (Phase 1 = Blockers):

# Creates `.claude/plans/{slug}/plan.md` with selected findings as tasks

## Phase 1: Blockers [PENDING]

- [ ] [P1-T1][active record] Fix transaction wrap in `User.create_with_profile` — Iron Law 5
- [ ] [P1-T2][sidekiq] Fix JSON-safe args in `EmailJob#perform` — Iron Law 9
- [ ] [P1-T3][active record] Fix N+1 in `OrdersController#index` — Iron Law 3

## Phase 2: Warnings (selected) [PENDING]

- [ ] [P2-T1][test] Add edge-case spec for `PaymentService` refund branch

## Phase 3: Suggestions (selected) [PENDING]

- [ ] [P3-T1][direct] Extract `RETRY_LIMIT` constant in `SyncService`

Example for REQUIRES CHANGES verdict (Phase 1 = Test Coverage Gaps):

# Creates `.claude/plans/{slug}/plan.md` with selected findings as tasks

## Phase 1: Test Coverage Gaps [PENDING]

- [ ] [P1-T1][test] Add spec for `PasswordsController#create` — test-coverage gap (REQUIRES CHANGES); source .claude/reviews/...md
- [ ] [P1-T2][test] Add spec for `PasswordsController#create` 429 throttle branch — test-coverage gap (REQUIRES CHANGES); source .claude/reviews/...md

Phase 1 heading is OMITTED entirely for PASS / PASS WITH WARNINGS verdicts. Phase 2 + Phase 3 headings are OMITTED when zero items selected at Step 4.

Best Practices

1. Always explain why a finding is a Blocker
2. Show estimated effort for each item
3. Group by location when possible for efficient fixing
4. Preserve excluded items for future reference
5. Link to documentation for Iron Law violations
6. Suggest alternatives for deferred items (e.g., "Use RuboCop auto-fix")

Commands

Command	Description
/rb:triage	Triage most recent consolidated review
/rb:triage <file>	Triage specific review file

For in-UI bucket shortcuts (All Warnings, Skip all Warnings, Group by file, …), surface them as AskUserQuestion option labels per § "Step 4: Present Multi-Select UI"; never prompt the user to type them as commands.

References

Need	Reference
auto-approve / usually-fix / often-skip rules + severity reclassification	${CLAUDE_SKILL_DIR}/references/triage-patterns.md
canonical fix-plan output template (Step 5 destination format)	${CLAUDE_SKILL_DIR}/references/triage-plan-template.md

Trust States

When triaging a review with a provenance sidecar, read the sidecar's trust_state (see ${CLAUDE_PLUGIN_ROOT}/references/output-verification/trust-states.md):

• conflicted: surface findings in a dedicated section labeled "Conflicted evidence — resolve before triage".
• missing: tag each finding [unverified]; treat as hint, not decision.
• weak: include findings but sort after runtime-confirmed evidence.
• clean: proceed with normal triage.

Related — invoke manually if needed

• Adversarial review needed → /rb:challenge (adversarial-mode review)
• PR review comments to address → /rb:pr-review (PR review-comment handling)