Apd audit

APD Project Audit

Qualitative review of how APD is configured in the project — content quality, not just file existence. Pairs with verify-apd (mechanical checks).

When to use / When to skip

Use when:

• First session after /apd-setup — confirm everything is correct
• After manually editing agents, CLAUDE.md, or settings.json
• When the pipeline behaves unexpectedly
• When verify-apd passes but something "feels off"
• Before handing the project to another developer

Skip when:

• verify-apd itself is failing — fix those mechanical issues first
• You only need a yes/no health check — verify-apd is faster
• Mid-pipeline — audit is for between cycles, not during

What This Checks (verify-apd Does NOT)

verify-apd	/apd-audit
Files exist?	Content correct and complete?
JSON valid?	Hook if patterns correct?
Agents have model?	Agents have correct model/effort/color?
Pipeline runs?	Pipeline output matches expected format?
Mechanical ✓/✗	Qualitative review

Process

1. Run verify-apd first

bash ${CLAUDE_PLUGIN_ROOT}/plugins/apd/bin/core/verify-apd

If FAIL → fix those first. This skill builds on top of verify-apd, not replaces it.

2. Agent Quality

For each agent in .claude/agents/*.md:

Roles that must EXIST — check presence before quality:

• code-reviewer — missing → the reviewer advance BLOCKs
• adversarial-reviewer — missing → the reviewer advance BLOCKs (adversarial-agent-missing). Until v7.0 its absence silently disabled the whole adversarial layer, so a project that has been running "clean" without this file was running without the layer.
• supervisor — missing → supervision-missing at the verifier (every profile since v7.0)

Frontmatter check:

• model: — full ids only, never a bare alias. opus/sonnet resolve to whatever the runtime maps them to today, which is how a corpus moved between model generations unnoticed. Builders claude-sonnet-5, code-reviewer claude-opus-5, adversarial-reviewer claude-sonnet-5, supervisor claude-opus-5. Do not "fix" a model a profile owns: apd profile status shows which roles the declared MODEL_PROFILE manages. A conf with no row for a role means the template pin stands — flag a mismatch, do not rewrite it.
• effort: — builders xhigh; code-reviewer, adversarial-reviewer and supervisor max
• color: — should be set (purple/blue/green/cyan for builders, orange for reviewer)
• permissionMode: — builders bypassPermissions, reviewers plan
• memory: — project for builders, but none for adversarial-reviewer and supervisor. Those two carry the decontextualization contract; flagging them for a missing memory: project inverts the thing that makes them worth dispatching.

Hook check:

• if: field must be inside hook objects, NOT at matcher level
• No env var prefixes in if patterns (e.g., Bash(git *) not Bash(APD_ORCHESTRATOR_COMMIT=1 git *))
• Guard paths use ${CLAUDE_PLUGIN_ROOT}/plugins/apd/bin/adapter/cc/ (the CC shims; bin/core/ holds the runtime-neutral implementations they call)
• Builders declare: guard-scope, guard-bash-scope, guard-secrets, guard-git
• Reviewers declare: guard-secrets, guard-git (NO guard-scope — read-only)
• adversarial-reviewer must carry the guard-spec-blind marker (v7.0)

The per-agent hooks: block is DATA, not execution. It never fires (measured on CC 2.1.220) — enforcement runs session-level from hooks/hooks.json. But bin/lib/agent-scope.sh resolves an agent's writable scope by reading the guard-scope command out of that block when no YAML scope: key exists. So a wrong SCOPE_PATHS list there is a live enforcement defect, and a missing block on a writable agent fails closed.

Body check:

• Has FORBIDDEN section with commit prohibition
• Has workflow section
• Scope paths match guard-scope arguments

3. CLAUDE.md Quality

Check that CLAUDE.md has all required sections:

• ## Stack — technology table
• ## APD — orchestrator role description
• ### Pipeline — enforced pipeline reference
• ### Guardrails — guard script list
• ### Mandatory skills — the table must name apd-pipeline-guide (mandatory before every task since v6.15, hard-gated by .guide-marker); brainstorm is advisory, not the gate
• ### Human gate — approval requirements
• ### Session memory — session-log reference
• ## Anti-patterns — common mistakes
• ## Memory — @.claude/memory/ references

Check that CLAUDE.md does NOT contain:

• {{PLACEHOLDER}} unreplaced values
• References to old skill names (/apd-init, /github-projects, /miro-dashboard)
• superpowers:subagent-driven-development references (should use APD pipeline)

4. Settings Quality

Read .claude/settings.json and verify:

• enabledPlugins.superpowers@claude-plugins-official: false
• attribution.commit: "" (empty — no AI signatures)
• attribution.pr: "" (empty)
• permissions.allow includes Edit(.claude/memory/**) and every pipeline file the framework mandates writing: spec-card.md, implementation-plan.md, .adversarial-summary, .adversarial-rationale.md, .supervision-summary, .supervision-rationale.md, .guide-marker
• Edit(...) only — never ask for a Write(...) twin. Since CC 2.1.208 Edit(path) covers every file-editing tool and a Write(path) rule is inert for file-permission checks; apd-init strips legacy APD Write twins on each run, so demanding them here would make the audit and init undo each other on every session-start
• Notification hook configured

5. Workflow Rules

Read .claude/rules/workflow.md and verify:

• Uses apd pipeline commands (not apd-pipeline)
• Has step 9 (finish)
• Has the mandatory skills section (apd-pipeline-guide first — it carries the gate contract)
• Model discipline table present, written with full model ids (never bare aliases). The orchestrator's own model is not APD-managed; MODEL_PROFILE governs agents only

6. Pipeline Health

bash .claude/bin/apd pipeline status
bash ${CLAUDE_PLUGIN_ROOT}/plugins/apd/bin/core/apd-init --version

• Pipeline responds without errors
• Version matches expected

7. Memory Files

Check .claude/memory/:

• MEMORY.md — not empty, has project context
• status.md — has current phase
• session-log.md — exists (may be empty for new projects)
• No [fill in] placeholders in the last session-log entry (blocks new tasks)

8. Drift Detection (v6.10+)

Run the dedicated drift script — it scans three dimensions where projects typically lag behind the framework baseline:

bash ${CLAUDE_PLUGIN_ROOT}/plugins/apd/bin/core/pipeline-audit-drift

Dimensions checked:

1. .claude/settings.json deny patterns — compares against current framework baseline (8 mkdir patterns: 4 slash-prefixed + 4 bare-dir). Pre-v6.10 re-inits left projects with only 4 patterns; v6.10 closes the bypass vector by writing all 8 on re-init.

2. .claude/.apd-config APD_VERSION — compares against the currently loaded plugin version. Stale APD_VERSION means the project was configured under an older minor and may carry stale workflow/agent templates. Patch-only drift is INFO; minor-or-major drift is IMPORTANT.

3. .claude/rules/workflow.md content markers — checks presence of six guidance markers (Implements:, rationale gate, DEPRECATED, unconditional, apd-pipeline-guide, SUPERVISION). Missing markers mean workflow.md was last refreshed under an older framework, so the orchestrator never sees plan-spec consistency, the rationale gate, the v6.15 guide gate or the v6.30 supervision layer. Read the marker list from the script rather than this page if they disagree — the script is the authority.

4. Feature claim drift (v6.12.3+) — scans workflow.md and CLAUDE.md for orchestrator confabulation patterns claiming features the framework does not ship. Specifically: any line that mentions BOTH a contracts command (verify-contracts, apd contracts) AND an unsupported language (PHP/Python/Java/Go/Ruby/Kotlin/Rust). First documented instance: Festico apd-setup (2026-05-28) — orchestrator wrote "apd verify-contracts automatically checks PHP DTO ↔ TS types" which is false; framework supports TS ↔ C# only. Detection prevents silent gaps in cross-layer review coverage where humans rely on a feature that errors at runtime.

Output buckets: CRITICAL (drift blocks pipeline structurally, rare) / IMPORTANT (drift compromises guard coverage or orchestrator guidance, most common) / INFO (patch-level, non-blocking) / CLEAN (project tracks current baseline).

Recovery: all drift findings point to /apd-setup (v6.10+ auto-fixes settings.json missing patterns + refreshes workflow.md + bumps APD_VERSION in .apd-config). Manual fixes are documented per item in the drift script output.

Exit code: drift script exits 1 if any CRITICAL or IMPORTANT finding, 0 if only INFO or CLEAN. Use in CI / pre-commit hooks if desired.

Output Format

APD Project Audit — {project name}

CRITICAL:
  1. [file:line] Description

IMPORTANT:
  1. [file:line] Description

CLEAN:
  ✓ Agents (X builder + 1 reviewer)
  ✓ CLAUDE.md sections complete
  ✓ Settings configured
  ✓ Workflow rules current
  ✓ Pipeline healthy
  ✓ Memory files present

Result: X findings (Y critical, Z important)

Common rationalizations

Excuse	Reality
"verify-apd passes so it's fine"	verify-apd checks structure, not content quality
"Agents work, no need to audit"	Wrong model or unbounded scope wastes time and money
"CLAUDE.md looks ok"	Missing sections mean orchestrator skips important rules
"I'll fix it when it breaks"	Broken pipeline produces broken code silently

Examples

Example 1 — Builder agent missing scope.

Input: .claude/agents/backend-api.md frontmatter has model, effort, permissionMode: bypassPermissions, but no scope: line. The builder can write anywhere in the repo.

Output:

IMPORTANT:
  1. [.claude/agents/backend-api.md] scope missing — builder writes are unbounded
     Effect: out-of-scope edits slip past the guard that keys on declared scope
     Fix: declare the builder's scope paths (the `{{SCOPE_PATHS}}` the guard-scope
     and guard-bash-scope hooks enforce) so writes outside them are blocked

Example 2 — Stale skill references in CLAUDE.md.

Input: CLAUDE.md ### Mandatory skills table lists /apd-init and /github-projects as required. Both were renamed (/apd-setup, /apd-github).

Output:

CRITICAL:
  1. [CLAUDE.md:142] References renamed skill /apd-init
     Effect: orchestrator looks for a non-existent skill, falls back to ad-hoc setup
     Fix: replace `/apd-init` → `/apd-setup`, `/github-projects` → `/apd-github`

Example 3 — Orphaned scope path on a builder.

Input: .claude/agents/frontend-web.md has scope: src/components/** but the project moved to app/components/. verify-apd passed because the agent file exists; guard-scope blocks every builder write.

Output:

CRITICAL:
  1. [.claude/agents/frontend-web.md:8] Scope path src/components/** does not exist
     Effect: every builder write blocked by guard-scope — pipeline cannot ship
     Fix: update to `scope: app/components/**` (or run /apd-setup gap analysis)

Exit criteria

You're done when:

• Every agent has been opened and its frontmatter checked against the matrix in §2
• Every required section in CLAUDE.md is present and free of unreplaced {{PLACEHOLDER}} values
• .claude/settings.json has all four required keys (env, attribution, enabledPlugins, hooks)
• apd pipeline status runs without error
• Findings are sorted into CRITICAL / IMPORTANT / CLEAN buckets in the output format
• If any CRITICAL is reported, the user has been told what to fix and in what order

Hand-off

• After audit completes with CRITICAL findings → invoke /apd-setup to regenerate missing pieces
• After audit completes clean → continue with normal development
• If audit reveals a structural finding not covered by /apd-setup → escalate to user with concrete file:line references