workflow-quality-audit

name	workflow-quality-audit
description	Domain-agnostic quality audit pipeline. Enumerates items, loads standards, captures evidence, evaluates per-item, posts structured verdicts, verifies coverage, and summarizes findings. Works for UI components, code review, SDLC, icon design, animation, or any domain with items to audit against criteria.
disable-model-invocation	true
context	fork
agent	general-purpose
allowed-tools	Read, Write, Edit, Bash, Glob, Grep, Task

Workflow: Quality Audit

A 7-step pipeline for auditing any collection of items against a set of standards. Domain-agnostic by design — the caller supplies a config block that plugs in the item source, standards document, evidence capture method, evaluation template, and reporting target.

When to Use

• After completing a body of work (design wave, sprint, migration) that needs QA sign-off
• Before a release to verify quality across a surface area
• For structured reviews: code, visual, accessibility, compliance, performance
• Any time you have N items and need a per-item verdict against defined criteria

Config Schema

Supply a YAML config block when invoking. All fields are optional — the pipeline will prompt for missing required fields.

# --- Required ---
domain: <string>              # Human label (e.g., "kingly-design-system", "api-endpoints")

items:
  source: <enum>              # bd-children | file-glob | manual-list | api
  parent: <string>            # bd parent bead ID (if source=bd-children)
  glob: <string>              # File glob pattern (if source=file-glob)
  list: [<string>, ...]       # Explicit item list (if source=manual-list)
  api: <string>               # API endpoint/command (if source=api)

standards:
  doc: <path>                 # Path to standards/checklist document
  sections: [<string>, ...]   # Specific sections to evaluate (optional, default=all)
  known_gaps: <string>        # Section to skip or flag separately

# --- Optional ---
evidence:
  type: <enum>                # screenshot | test-run | recording | measurement | source-read
  capture: <string>           # Shell command template for capture
  naming: <string>            # Filename template: {id}, {slug}, {date} supported
  storage: <path>             # Output directory for evidence artifacts


               
            
                  
               
             


                 
   [, ]  
             


                 

Pipeline Steps

Step 1: INVENTORY

Enumerate all items to audit.

Source	Method
bd-children	bd list --parent {parent} — collects bead IDs + titles
file-glob	Glob({glob}) — each matching file is an item
manual-list	Literal list from config
api	Execute {api} command, parse JSON/lines output

Output: An ordered list of (id, title/name, metadata) tuples. Persist to .lev/reports/audit-{domain}-{date}/inventory.json.

Step 2: STANDARDS

Load and parse the evaluation criteria.

1. Read the standards.doc file
2. If sections is specified, extract only those sections
3. If known_gaps is specified, flag that section as informational (not scored)
4. Parse into a checklist of (category, criterion, weight) triples

Output: Structured checklist ready for per-item evaluation.

Step 3: EVIDENCE

Capture per-item evidence using the configured method.

Type	Method
screenshot	Execute {capture} command per item, save to {storage}/{naming}
test-run	Execute test suite, capture per-item results
recording	Capture screen/interaction recording per item
measurement	Run measurement command, capture metrics
source-read	Read source files associated with each item

Evidence is stored at {storage}/{naming} with template variables expanded. If capture fails for an item, mark it EVIDENCE_FAILED and continue.

Output: Evidence manifest mapping each item ID to its artifact paths.

Step 4: EVALUATE

Run the evaluation checklist against each item's evidence.

Team delegation pattern:

• If agents > 1 and parallel: true, partition items across N evaluator agents
• Each evaluator receives: item subset, standards checklist, evidence paths
• Use split_by to control partitioning (round-robin if not specified)

Per-item evaluation produces:

## {item.title}

### {category_1}
- [ ] {criterion}: {PASS|FAIL} — {note}
- [ ] {criterion}: {PASS|FAIL} — {note}

### {category_2}
...

**Verdict:** {verdict} ({pass_count}/{total_count})
**Issues:**
1. {issue description} — severity: {P0-P3}

Built-in templates (see templates/ directory):

• component-audit — Visual + Code Quality + Accessibility + Standards Compliance
• visual-audit — Visual Fidelity + Storybook Coverage + Code Structure
• code-review — Correctness + Style + Performance + Security + Tests
• generic — User-defined categories (bare scaffold)
• custom — Load from custom_path

Step 5: REPORT

Post the structured verdict for each item.

Target	Method
bd-comments	bd comment {item.id} --body "{verdict}" per item
markdown-files	Write {output_dir}/{item.id}-verdict.md per item
stdout	Print verdicts to console

Each verdict includes: item ID, title, verdict label, pass/fail counts, issue list.

Step 6: VERIFY

Confirm 100% coverage.

1. Load inventory from Step 1
2. Load all verdicts from Step 5
3. Check: every item has exactly one verdict
4. Check: no items are EVIDENCE_FAILED without a retry or exception note
5. Check: verdict taxonomy matches config (no unknown labels)

Output: Coverage report — {covered}/{total} with any gaps listed.

If gaps exist, report them and optionally re-run Steps 3-5 for missing items.

Step 7: SUMMARIZE

Aggregate findings into a top-level summary.

Scorecard format (default):

Domain: {domain}
Date: {date}
Items: {total} | PASS: {n} | NEEDS_WORK: {n} | BLOCKED: {n}
Coverage: {covered}/{total} ({pct}%)
Issues: {total_issues} (P0: {n}, P1: {n}, P2: {n}, P3: {n})
Top systemic patterns:
  1. {pattern} — {count} items affected
  2. {pattern} — {count} items affected

Write full report to {output_dir}/audit-{domain}-{date}.md.

Team Delegation

For audits with >5 items, delegate to parallel evaluators:

Role	Responsibility
Coordinator (you)	Steps 1, 2, 6, 7 — inventory, standards, verify, summarize
Evaluator 1..N	Steps 3-5 — capture evidence, evaluate, post verdicts

Evaluators receive a self-contained work packet:

• Their item subset (IDs + titles)
• The full standards checklist
• The evidence capture config
• The report template + target config

Evaluators return a manifest: {items_evaluated, verdicts, issues_found, evidence_paths}.

Domain Examples

UI Components (the original use case)

domain: kingly-design-system
items: { source: bd-children, parent: oclaw-top.2 }
standards: { doc: packages/kingly-ui/DESIGN_STANDARDS.md }
evidence: { type: screenshot, capture: "apptestr screenshot" }
evaluation: { template: component-audit, agents: 2, parallel: true }
reporting: { target: bd-comments, verdict_taxonomy: [PASS, NEEDS_WORK] }

Code Review

domain: api-endpoints
items: { source: file-glob, glob: "src/routes/**/*.ts" }
standards: { doc: docs/API_STANDARDS.md }
evidence: { type: source-read }
evaluation: { template: code-review, agents: 2 }
reporting: { target: markdown-files, output_dir: .lev/reports/api-review/ }

SDLC / Sprint Review

domain: sprint-42
items: { source: bd-children, parent: sprint-42 }
standards: { doc: docs/DEFINITION_OF_DONE.md }
evidence: { type: test-run, capture: "pnpm test --filter {id}" }
evaluation: { template: generic }
reporting: { target: bd-comments, verdict_taxonomy: [SHIP, FIX, HOLD] }

Icon Design

domain: app-icons
items: { source: file-glob, glob: "assets/icons/**/*.svg" }
standards: { doc: docs/BRAND_GUIDE.md, sections: [Icons, Color Palette] }
evidence: { type: measurement, capture: "convert {path} -resize 16x16 /tmp/{slug}-16.png" }
evaluation: { template: visual-audit }
reporting: { target: markdown-files, verdict_taxonomy: [APPROVED, REVISE] }

Validation

The audit is complete when:

• Inventory persisted with item count logged
• Standards checklist parsed and loaded
• Evidence captured for all items (or failures documented)
• Every item has exactly one verdict
• Coverage is 100% (or gaps explicitly noted)
• Summary report written to output directory
• All P0/P1 issues have corresponding tracking items (if using bd)

Notes

• For bd-children source, items inherit the parent's epic/project context
• Evidence capture failures are non-fatal — the item gets EVIDENCE_FAILED and continues
• The known_gaps standards section is included in reports as informational but not scored
• Parallel evaluation scales linearly: 2 agents = ~2x throughput for >10 items
• Custom verdict taxonomies must have at least 2 labels (one positive, one negative)
• Templates are composable — component-audit includes visual + code + a11y categories