| name | issue:review |
| description | Run QA criteria against the current issue's codebase and produce structured findings. Use when asked to review, audit, or QA the current issue. Triggers on /issue:review. |
Issue Review
Run the QA criteria for the current issue against the codebase. Produces a findings file that mirrors the criteria structure with results nested inline.
Tiers
Groups are tagged by tier. Select one or more tiers to run — each runs independently (not cumulative).
| Tier | When to use |
|---|
| gate | Every pass. Fast, automated, binary. |
| quality | Structural + product quality. Adds judgment. |
| polish | Late-stage. Edge cases, manual QA. |
To run: /issue:review gate, /issue:review quality, /issue:review all. No argument prompts for selection.
Requirements
bd CLI (beads plugin active)- Git repo on a feature branch with issue key (e.g.
feature/hea-3849-*)
- Epic with
--external-ref pointing to the issue key
- Issue directory with numbered phase subdirectories (the README chain)
- A QA phase directory (CRITERIA.md will be scaffolded if missing)
Steps
1. Resolve Issue Context via Epic
Follow the flo chain to discover the issue directory and QA phase:
- Resolve the epic from the current branch (same as
flo:next — parse branch for issue key, find epic with matching --external-ref)
- Read the epic's
design field — it points to the issue directory
- Read the issue directory's
README.md — it lists the numbered phase directories (1-spec, 2-design, 3-plan, 4-qa, etc.)
- Follow the chain to the QA phase directory
- Read the QA
README.md to understand the process and file layout
- Read the spec (
1-spec/SPEC.md) and terminology (2-design/terminology.md) — these are referenced by the criteria and needed for judgment-based checks
2. Read or Bootstrap Criteria
Check if the QA directory has a CRITERIA.md file.
If CRITERIA.md exists: Read it. Parse the tier sections and group structure. Identify:
- Available tiers and which groups each includes
- The criteria items within each group (the nested list structure)
If CRITERIA.md does not exist: Bootstrap it from the template asset at .claude/skills/issue_review/CRITERIA.template.md. Copy the template into the QA directory as CRITERIA.md. Then tell the user:
Scaffolded CRITERIA.md from template. Fill in criteria under each tier section (Gate, Quality, Polish) before running a review. See HTML comments in each section for guidance.
Stop here — do not proceed to tier selection or execution until the user has populated the criteria.
3. Determine Tier
Check $ARGUMENTS for a tier name (e.g. /issue:review quality) or all. Valid values are the individual tier names from the criteria file (e.g. gate, quality, polish) plus all.
Each tier runs independently — picking quality runs only quality checks, not gate + quality. all runs every tier.
- If a valid tier or
all is provided as argument → use it, skip prompting
- If no argument or invalid argument → use AskUserQuestion to ask which tier(s) to run. Options: each tier from the criteria file +
all. Enable multiSelect so the user can pick multiple.
4. Determine Findings File Number
Look for existing FINDINGS.*.md files in the QA directory. The new file is FINDINGS.{N+1}.md where N is the highest existing number. If none exist, start at FINDINGS.1.md.
5. Execute Criteria Checks
For each group within the selected tier ceiling, execute the criteria checks against the codebase.
How to check: Read relevant source files, grep for patterns, run commands (lint, typecheck) as appropriate. Each criterion is a judgment call — evaluate it and record what you find.
Important:
- Record every criterion's outcome — both passes and failures. Passes go under
## PASSING, failures under ## FAILING.
- For gate-tier mechanical checks, actually run the commands (
npm run check:types, etc.) and report pass/fail.
- For judgment-based criteria, read the relevant code and apply the criterion.
- Exclude
.claude/, .beads/, .issues/ from source scans (per criteria instructions).
6. Write Findings File
Write 4-qa/FINDINGS.{N}.md using the inline findings format described below.
Findings File Format
The findings file splits results into two top-level sections: FAILING (findings that need attention) and PASSING (criteria that passed). Within each section, tiers provide the next level of grouping, preserving the full criteria tree structure underneath.
FAILING comes first — failures are always the first thing a reader sees.
# QA Findings — Pass {N}
Branch: `{branch}`. Date: {date}. Tier: `{tier}`.
---
## FAILING
### {Tier Name}
- **{GROUP NAME}**
- **{Sub-group}** — {original criterion text}
- {criterion item that had a finding}
- FINDING: {what was observed, with file:line references}
- FINDING: {another observation if multiple}
- {another criterion item with a finding}
- FINDING: {observation}
---
## PASSING
### {Tier Name}
- **{GROUP NAME}**
- {criterion that passed} — Pass.
- {another criterion that passed} — Pass.
Rules:
- Preserve the original criteria text exactly — the criterion line is verbatim from CRITERIA.md
- Findings are always prefixed with
FINDING: and indented one level below the criterion
- Include
file:line references wherever possible
- Within each status section, omit any criterion, sub-group, group, or tier that has zero entries
- Omit the
## FAILING section entirely if everything passed
- Omit the
## PASSING section entirely if everything failed
- Passing criteria get a brief
— Pass. suffix (or a short note if context is useful, e.g. — Pass. No errors.)
Example:
## FAILING
### Quality
- **CONSISTENCY**
- **DRY**
- No business logic repeated — same validation rule, transformation, or filter predicate in multiple places
- FINDING: `nonExcluded` filtering (`plan.records.filter(r => r.exclusionType === null)`) appears in wizardService.ts:142, :287, :341, :403, :512. Extract to helper.
- FINDING: `formatCurrency()` identical in Step3Validation.tsx:23, Step6Review.tsx:45, Step7Apply.tsx:67.
- **Pattern consistency**
- Error handling: every mutation surfaces failures to the user (onError callback, callout, or toast)
- FINDING: Steps 1, 3, 4, 6 have no `onError` callback on their mutations. Steps 2, 5, 7 do.
---
## PASSING
### Gate
- **MECHANICAL**
- Lints pass (project lint command) — Pass. CI lint check green.
- Types pass (`npm run check:types`) — Pass. No errors.
- Build succeeds (project build command) — Pass. Built in 19.41s.
- CI checks green (`gh pr checks`) — Pass. All 8 checks green.
### Quality
- **CONSISTENCY**
- **Terminology** — the terminology doc is the authority
- Every step component filename follows `Step{N}{RouterKey}` — Pass.
- Prisma model names match terminology doc — Pass.
7. Report Results to User
Summarize:
- How many criteria were checked
- How many findings were produced
- Which groups had the most findings
- Any critical/blocking findings (e.g. gate failures)
8. Propose Beads
Analyze the findings and propose beads to create. Group findings into logical work units:
- Batch related findings into single beads (e.g. all DRY violations into one "Extract shared utilities" bead)
- Identify which findings are blockers vs improvements
- State which epic they'd be assigned to (the current issue)
- Present the proposed beads to the user for approval before creating
Use bd create syntax in the proposal so the user can see exactly what would be created.
Notes
- Context discovery follows the same epic → issue directory → README chain as
flo:next. This skill does not hardcode paths — it discovers the QA phase by reading the issue structure.
- The findings file is append-only per pass — never overwrite a previous pass's file
- Gate failures should be called out prominently — if mechanical checks fail, note that nothing else matters until they're fixed
- HUMAN-IN-THE-LOOP items (in polish tier) cannot be executed by the agent — present them as a checklist for the user instead of attempting to evaluate them
