Audit Feature

Audit Feature

Orchestrates per-layer audits of a Flutter feature folder. Spawns one Explore subagent per detected layer in parallel, aggregates all violations, and offers a combined fix prompt.

This skill has no rules of its own — it reads each layer's skills/audit-<layer>-layer/rules/CATALOG.md and passes the rules inline to the corresponding Explore subagent.

Phase 0 — Resolve feature path

Read the user's request and extract a feature folder path. Accepted forms:

• Top-level feature: lib/src/features/<feature>/
• Sub-feature: lib/src/features/<parent>/<sub>/
• Bare feature name (e.g. auth) — expand to lib/src/features/auth/

If the path is ambiguous, ask exactly one question:

"Provide the feature folder path (e.g. lib/src/features/auth/) or the feature name."

Do not proceed until a path is confirmed.

Phase 1 — Detect layers

Check which layer directories exist under the resolved feature path:

<feature>/
  domain/          ← layer key: domain
  data/            ← layer key: data
  application/     ← layer key: application
  presentation/    ← layer key: presentation

Sub-feature / presentation-only shortcut: if only presentation/ is present (and no domain/, data/, or application/), delegate the entire audit to audit-presentation-layer by invoking the Skill tool with that skill name, passing the presentation/ path. Then stop — do not continue to Phase 2.

Otherwise, for each layer directory that exists, prepare a parallel Explore invocation (see Phase 2). Log a warning for any expected layer that is absent:

⚠️  No application/ directory found — skipping audit-application-layer.

Graceful degradation: if a layer's CATALOG file (skills/audit-<layer>-layer/rules/CATALOG.md) is missing, emit a warning and continue auditing the other layers:

⚠️  skills/audit-domain-layer/rules/CATALOG.md not found — skipping domain audit.
    Install the audit-domain-layer skill to enable this check.

Phase 2 — Parallel layer audits

For each present layer, spawn one Explore subagent simultaneously (all in a single response with multiple Agent tool calls). Pass to each Explore:

1. The full contents of the layer's rules/CATALOG.md (read it before spawning).
2. The layer directory path to scan.
3. A self-contained scan prompt (see template below).

Explore prompt template

You are performing a static architecture audit of Flutter <LAYER>-layer files.

## CATALOG (rules to enforce)
<paste full contents of skills/audit-<layer>-layer/rules/CATALOG.md here>

## Target
Scan all .dart files (excluding .g.dart, .freezed.dart) under:
  <feature_path>/<layer>/

## Instructions
1. List every .dart file found (relative path + approximate line count).
2. For each file, read the full content and apply every rule in the CATALOG above.
3. For each violation found, record:
   - file (relative path)
   - line number
   - rule_id
   - severity (error / warning / info)
   - brief message (one line)
4. Return results as a markdown table grouped by file, followed by a violation count
   summary: "N violations (E errors, W warnings, I info)".
5. If no violations found in a file, omit that file from the table.
6. If no violations found at all, write: "No violations — <layer> layer passes all rules."

Phase 3 — Aggregate and report

After all Explore subagents complete, merge their outputs into a single report:

## Full Feature Audit — <feature>

### Domain Layer
<paste domain Explore output — table + count>

### Data Layer
<paste data Explore output — table + count>

### Application Layer
<paste application Explore output — table + count>

### Presentation Layer
<paste presentation Explore output — table + count>
(platform-aware rules run by audit-presentation-layer; platform: <resolved>)

---
**Grand total**: N violations across K files (E errors, W warnings, I info)

If a layer was skipped (missing directory or missing CATALOG), include the warning in its section header instead of a table.

Phase 4 — Combined fix prompt

After the report, ask:

Apply fixes for which rule IDs? (comma-separated, "all", or "none")
List the rule IDs you want fixed:

On response:

• "none" or no response: done.
• "all" or specific IDs:
  1. Group selected IDs by layer.
  2. For each layer with selected fixes, apply heuristics from that layer's CATALOG (read from skills/audit-<layer>-layer/rules/CATALOG.md).
  3. For autofix_safe: true rules: apply edits directly, show diff.
  4. For autofix_safe: false rules: show the required transformation and ask for confirmation before editing.
  5. After edits, state which violations were resolved.

Never edit files that were not explicitly approved by the user.

Usage examples

• audit all layers of features/auth
• full feature audit for features/booking
• check all layers in lib/src/features/flight_plan/
• /audit-feature auth
• /audit-feature lib/src/features/booking/
• /audit-feature lib/src/features/home/home_screen/ ← sub-feature → presentation-only

Notes

• Paths are relative to the project root — always resolve from there.
• This skill spawns Explore subagents, not Skill invocations — Explore does not have access to the Skill tool, so rules are passed inline in the prompt.
• audit-presentation-layer is the exception: for the presentation shortcut path it is invoked via the Skill tool (which supports platform detection from pubspec.yaml). For full multi-layer audits, the presentation layer is scanned via an Explore subagent using audit-presentation-layer's CATALOG directly.
• To modify per-layer rules, edit the respective skills/audit-<layer>-layer/rules/CATALOG.md — this orchestrator reads them at runtime.