coder

name	coder
description	Write/modify code for autopilot tasks
agent	.claude/agents/coder.md

Coder Skill (Wrapper)

Invokes coder subagent for writing/modifying code within a single task.

Architecture: This skill is a WRAPPER over .claude/agents/coder.md. The agent file is the source of truth for the coder prompt.

When to Use

Internal: Called by autopilot for each task in implementation plan

Standalone: Rarely used directly — prefer using autopilot workflow

Invocation

Task tool:
  description: "Implement task N"
  subagent_type: "coder"
  prompt: |
    TASK: {task description from plan}
    ALLOWED FILES: {from spec}

    Execute and report files_changed: [list]

Output

files_changed:
  - path/to/file1.py
  - path/to/file2.py
status: completed | blocked

Module Headers Workflow (MANDATORY)

When working with a file:

1. OPENED → read module header
2. Header empty? → create before changes
3. MADE changes
4. RE-READ header → update if needed (Uses, Used by, Role)
5. SAVED

Module Header Format

"""
Module: {module_name}
Role: {one-line purpose}
Source of Truth: {what is authoritative — SQL RPC, this file, etc.}

Uses:
  - {module}: {what classes/functions}

Used by:
  - {module}: {for what purpose}

Why here: {context if non-obvious}

Glossary: ai/glossary/{domain}.md
"""

Post-Change Verification (MANDATORY)

After modifying a file:

1. If changed term/naming:

   grep -rn "{old_term}" . --include="*.py" --include="*.sql" --include="*.ts"
   

   Result must be 0.

2. If changed API/signature:

   grep -rn "{function_name}" . --include="*.py"
   

   All calls updated?

3. If added new term: → Add to corresponding ai/glossary/{domain}.md

Notes

• Coder MUST respect File Allowlist from spec
• Coder MUST use Edit/Write tools, not Bash
• Coder outputs only file paths, not implementation details