audit-plan（审计计划）

Plan Flow Audit

YOUR ARGUMENTS: $ARGUMENTS

Arguments are provided above via $ARGUMENTS. Parse them to determine the plan folder path (e.g., plans/detector-refactor). Do NOT ask the user to re-provide the path.

This skill performs a structural flow audit of a multi-phase implementation plan. It assesses whether phases connect coherently as a pipeline — NOT whether individual phases comply with templates (that's /review-plan's job).

Positioning: This audit runs BEFORE per-phase reviews. Phases have NOT been reviewed or polished yet. Focus on intended design vs. planned structure, not on template completeness or placeholder content.

Critical

• Do NOT modify plan files, phase files, or codebase files — this skill only writes the audit report
• Read ALL phase files, not just a sample — the value of this audit is completeness
• Ground findings in actual codebase files, not just plan documents
• When you claim a phase targets a file that doesn't exist, verify with Glob first
• Distinguish between intentional design choices and actual discrepancies
• Tolerate rough edges — phases may have incomplete sections, placeholder content, or imperfect formatting. That's /review-plan's job to fix. You're checking whether the PLAN DESIGN is sound.

Task Tracking

Tasks survive context compacts — skipping this check causes lost progress and repeated work.

Before starting work, run TaskList to check if tasks already exist from a previous session or before a compact. If tasks exist:

1. Read existing tasks with TaskGet for each task ID
2. Find the first task with status pending or in_progress
3. Resume from that task — do NOT recreate the task list

If no tasks exist, create them after reading the master plan (Step 1):

Example task list:

Task 1: Read the master plan
Task 2: Read all phase files and extract metadata
Task 3: Build dependency graph + bail-out check
Task 4: Assess data flow consistency
Task 5: Evaluate phase ordering
Task 6: Identify stale artifacts
Task 7: Assess risk for pending phases
Task 8: Write audit report
Task 9: Return summary

Mark each task in_progress when starting and completed when done.

Why This Audit Exists

/review-plan checks individual files for template compliance and codebase patterns. This audit checks the connections between them — structural design issues that only appear at the whole-plan level.

This runs FIRST (before /review-plan) because structural issues invalidate all per-phase review work. No point polishing Phase 5's code blocks if Phase 5 depends on a table that Phase 3 doesn't actually create.

The audit report feeds into /implement — it checks the flow audit verdict before building. A structural audit catches design problems that per-phase reviews cannot see.

Output Location

The flow audit report goes in the reviews/planning/ subfolder alongside per-phase reviews:

Write to: {plan-folder}/reviews/planning/flow-audit.md

Examples:

• plans/250202-voice-assistant/reviews/planning/flow-audit.md
• plans/250202-api-refactor/reviews/planning/flow-audit.md

Create the reviews/planning/ directory if it doesn't exist.

Note: Per-phase template/codebase reviews go in the same folder — see /review-plan. Code/implementation reviews go in reviews/code/ — see /code-review.

Step 1: Read the Master Plan

Read {plan-folder}/plan.md completely. Extract:

1. Phase table — all phases with titles, status, and any noted issues
2. ADRs / Decision Log — decisions that constrain phase design
3. Critical Issues section — known problems already documented
4. Architectural North Star — patterns all phases must follow

Note any phases marked as deprecated, renumbered, or amended.

Step 2: Read All Phase Files

Use Glob to find all phase files: {plan-folder}/phase-*.md

For EACH phase file, extract from the frontmatter and overview:

Build a mental model of the full dependency graph as you read.

Placeholder Check (Informational Only)

Run the placeholder validator to note which phases still have skeleton content:

echo '{"cwd":"."}' | uv run $CLAUDE_PROJECT_DIR/.claude/hooks/validators/validate_no_placeholders.py \
  --directory {plan-folder} --extension .md

This is NOT a blocker. Since this audit runs before /review-plan, phases may still contain [To be detailed], TBD, or skeleton markers. Note these in the report for awareness, but do NOT fail the audit because of them. Reviewers will clean them up. Only flag placeholders in critical structural sections (dependencies, data flow descriptions) as a concern.

Step 3: Build Dependency Graph + Bail-Out Check

From the data collected in Step 2, construct the full dependency graph.

For each phase, document:

• Declared dependencies (from frontmatter dependencies: field)
• Implicit dependencies (phase references data/files/services created by another phase but doesn't declare it)
• What downstream phases depend on it

Check for:

Bail-Out Assessment

After building the dependency graph, evaluate whether the plan is fundamentally broken. If ANY of these conditions are true, STOP the audit immediately — write a short report with assessment "Unusable" and return:

If the plan is bail-out level broken: Skip Steps 4-7. Write a minimal report (Step 8) with:

• Overall Assessment: Unusable — Needs Restructuring
• The specific bail-out condition(s) detected
• A brief recommendation for what needs to change
• Return immediately (Step 9)

This saves the user from waiting through a full audit of a plan that needs to be thrown out and rewritten.

Step 4: Assess Data Flow Consistency

Trace how data flows through the plan pipeline. For plans involving data transformations (API -> storage -> consumption), check:

1. Source consistency — Do all phases agree on where data comes from?
2. Schema consistency — Do phases that share data structures use the same types/interfaces?
3. Pattern consistency — Is there one architectural pattern or do different phases use contradictory approaches?

- Example: Some phases have components query the database directly while others expect data passed via context/props

1. Table/entity naming — Do all phases reference the same table names consistently?
2. File path consistency — Do phases reference files that actually exist at those paths?

Verify against codebase: For key files referenced by "Done" phases, use Glob to check they exist. For critical claims (e.g., "Phase 11 updated the orchestrator to load summary data"), read the actual file to confirm.

Step 5: Evaluate Phase Ordering

Assess whether the current phase sequence makes logical sense:

1. Infrastructure before consumers — Are foundational phases (schemas, services, APIs) ordered before phases that use them?
2. Producers before consumers — Does the phase that creates data run before phases that read it?
3. Refactoring order — For refactoring plans, is the new system built before the old system is deprecated/removed?
4. Test timing — Are testing phases positioned to test code that already exists?
5. Cleanup last — Are deprecation, migration, and drop phases at the end?

Flag ordering issues like:

• A consumer phase runs before its data producer
• A centralized service is created AFTER the things that should use it
• Tests are written before the code they test exists
• Cleanup runs before all references are migrated

Step 6: Identify Stale Artifacts

Check for:

Step 7: Assess Risk for Pending Phases

For each pending phase, evaluate risk based on:

Step 8: Write Audit Report

Write the report to: {plan-folder}/reviews/planning/flow-audit.md

Create the reviews/planning/ directory if it doesn't exist.

If the report already exists (re-auditing after fixes), Read the existing file first — the Write tool silently fails on existing files without a prior Read. A re-audit should fully overwrite the previous report since the value is a fresh holistic view, not incremental updates.

Report Structure

# Flow Audit: {Plan Title}

**Audited:** {date}
**Phases reviewed:** {count} ({done count} Done, {pending count} Pending, {deprecated count} Deprecated)
**Overall Assessment:** {Coherent | Minor Issues | Significant Issues | Major Restructuring Needed | Unusable}

---

## 1. Dependency Graph

{Visual or textual representation. Use indentation to show dependency chains.}

### Dependency Issues

| #   | Issue | Phases Affected | Severity | Suggested Fix |
| --- | ----- | --------------- | -------- | ------------- |
| 1   | ...   | P05, P12        | High     | ...           |

---

## 2. Data Flow Analysis

### Architecture Pattern(s)

{Describe the data flow pattern(s) used across the plan}

### Inconsistencies

| #   | Issue | Phases Affected | Details |
| --- | ----- | --------------- | ------- |
| 1   | ...   | ...             | ...     |

---

## 3. Phase Ordering Assessment

### Current Order

{List phases in order with one-line description}

### Ordering Issues

| #   | Issue | Current Order | Suggested Order | Rationale                                       |
| --- | ----- | ------------- | --------------- | ----------------------------------------------- |
| 1   | ...   | P12 after P11 | P12 before P11  | Loader should exist before orchestrator uses it |

---

## 4. Stale Artifacts

| #   | Artifact | Type            | Location                      | Action Needed     |
| --- | -------- | --------------- | ----------------------------- | ----------------- |
| 1   | ...      | Deprecated file | phase-17-unit-tests-update.md | Delete or archive |

---

## 5. Risk Assessment (Pending Phases)

| Phase | Risk   | Key Risk Factors                                   | Recommendation                    |
| ----- | ------ | -------------------------------------------------- | --------------------------------- |
| P12   | High   | Depends on P11 which may have ordering issues      | Review dependency direction       |
| P14   | Medium | Modifies Edge Function with complex existing logic | Read function before implementing |

---

## 6. Recommendations (Priority Order)

1. **[Critical]** {Description} — Fix before continuing implementation
2. **[High]** {Description} — Fix before implementing affected phases
3. **[Medium]** {Description} — Fix for plan hygiene
4. **[Low]** {Description} — Nice to have

Step 9: Return Summary

After writing the report, return to the main agent:

1. Report location — file path
2. Overall assessment — one-line verdict
3. Critical/High issues count — how many need attention before implementation continues
4. Top 3 findings — the most impactful discoveries
5. Recommendation — whether to proceed, fix first, or restructure
6. Context hygiene: "Consider running /compact before starting the next workflow step."

Resuming After Context Compact

If you notice context was compacted or you're unsure of current progress:

1. Run TaskList to see all tasks and their status
2. Find the in_progress task — that's where you were
3. Run TaskGet {id} on that task to read full details
4. Continue from that task — don't restart from the beginning

Tasks persist across compacts. The task list is your source of truth for progress, not your memory.

Pattern for every work session:

TaskList → find in_progress or first pending → TaskGet → continue work → TaskUpdate (completed) → next task

Constraints

• Source files are read-only — do NOT modify plan files, phase files, or codebase files. The only file you write is the audit report at {plan-folder}/reviews/planning/flow-audit.md
• All phases — read every phase file, not just a sample
• Codebase grounding — verify that key files referenced in phases actually exist at those paths. Use Glob to spot-check
• No template checking — that's /review-plan's job. Focus on structural design, not format or polish
• Tolerate rough edges — phases have NOT been reviewed yet. Placeholder content, incomplete sections, and formatting issues are expected. Only flag these if they obscure structural intent (e.g., dependencies section is entirely TBD)
• Severity calibration — only mark issues as Critical if they would cause implementation failure (circular deps, fundamentally wrong ordering). Use High for likely problems, Medium for structural concerns, Low for suggestions
• Bail-out early — if the plan is fundamentally broken (Step 3), stop and say so. Don't spend tokens auditing an unusable plan
• Pipeline awareness — this audit runs BEFORE /review-plan. Your verdict determines whether per-phase reviews proceed or the plan gets sent back to the planner

Troubleshooting

Too Many Phase Files

Symptom: Plan has 30+ phases and reading all of them is slow.

Fix: Read frontmatter + overview section only (first 60 lines) for initial pass. Deep-read only phases flagged as problematic during the dependency/flow analysis.

Plan Has No Dependency Metadata

Symptom: Phase files don't have dependencies: in frontmatter.

Fix: Infer dependencies from content — look for "requires Phase X", "uses table from Phase Y", "after Phase Z is complete" in the overview and prerequisites sections. Flag the missing metadata as a stale artifact.

Phases Have Lots of Placeholders

Symptom: Many phases have [To be detailed] or TBD in their content.

Fix: This is expected — the audit runs before reviews clean these up. Only flag placeholders if they appear in structural sections (dependencies, data flow, key file paths) where you can't assess the design without them. Note the count in the report for the reviewers' benefit.