scan（scan 效率）

Codebase Scan

Audit the codebase against another skill's criteria using a parallel agent team.

Workflow

1. Parse args & load skill

Extract the skill name from the args passed to this skill.

• If no skill name provided, list available skills in .claude/skills/ and ask the user which to scan against.
• If skill doesn't exist, list available skills and tell the user.
• Read .claude/skills/<name>/SKILL.md plus any files in references/ and rules/ subdirectories.
• Distill the skill's content into a numbered criteria checklist: a flat list of concrete, testable rules labeled C1, C2, C3, etc. Each criterion should be a single sentence describing what to check for.
• If the skill has no evaluable code criteria (e.g., workflow-only skills like why that don't define code patterns or rules), tell the user it's not scannable and stop.

2. Discover relevant files

Use the skill's criteria to infer file scope:

• React/frontend criteria → app/**/*.tsx, app/**/*.ts
• Backend criteria → services/**/*.ts
• General/mixed → both of the above
• CSS/styling → app/**/*.css, app/**/*.tsx

Always exclude: node_modules/, dist/, *.test.*, *.spec.*, **/migrations/**, **/*.d.ts, generated files.

Count candidate files. If zero, tell the user and stop.

3. Plan team composition

Split files into chunks by directory subtree so no file is assigned to two agents:

For very small scans (< 20 files), skip the team — scan inline and jump to step 6.

4. Spawn scan team

1. spawnTeam with name scan-<skill-name>
2. Create one TaskCreate per chunk with:

- Subject: Scan <directory-area> against <skill-name> criteria - Description: Include the full criteria checklist, the file scope (glob patterns), and the teammate instructions from references/teammate-instructions.md

1. Spawn teammates as general-purpose subagent_type (they need Read, Glob, Grep + team communication)
2. Name teammates scanner-1, scanner-2, etc.
3. Assign tasks via TaskUpdate with owner

5. Collect & synthesize

Wait for all teammates to complete their tasks. Each teammate reports findings in structured format:

FINDING: C<n> | SEVERITY | file/path.ts:LINE | Description

Collect all findings from teammate messages.

6. Generate report

Create .claude-scan/<skill-name>.md using the format in references/report-format.md.

Key sections:

• Executive Summary (2-3 sentences)
• Criteria Evaluated (table with violation counts)
• Findings by Severity (Critical → Warning → Info tables)
• Patterns Observed
• Statistics
• Recommended Fix Order (batched by non-overlapping file groups)

7. Cleanup & present

• If a team was spawned: send shutdown_request to all teammates, then cleanup
• Display inline summary: total findings by severity, top violated criteria, scan scope
• Tell the user the full report is at .claude-scan/<skill-name>.md
• Ask: "Want me to spawn a fix team to address these findings?"

8. Fix team (if user says yes)

1. Read the report's "Recommended Fix Order" section
2. Group fixes into non-overlapping file batches
3. spawnTeam with name fix-<skill-name>
4. Create tasks per batch — each task description includes:

- The specific violations to fix (from the report) - The relevant criteria definitions - Instructions from references/fix-team-instructions.md

1. Spawn general-purpose teammates named fixer-1, fixer-2, etc. (need Edit/Write)
2. Assign tasks and wait for completion
3. Shutdown and cleanup fix team
4. Ask user if they want to re-scan to verify fixes