auto-coder - SKILL.md Agent Skill

name: auto-coder description: Autonomous spec-driven development agent. Syncs DEV_SPEC.md into chapter-based reference files, identifies the next pending task from the schedule, implements code following spec architecture and patterns, runs tests with up to 3 auto-fix rounds, and persists progress with atomic commits. Use when user says "auto code", "自动开发", "自动写代码", "auto dev", "一键开发", "autopilot", or wants fully automated spec-to-code workflow.

Auto Coder

One trigger completes read spec → find task → code → test → persist progress.

Optional modifiers: append a task ID (e.g. auto code B2) to target a specific task, or --no-commit to skip git commit.

Pipeline

Sync Spec → Find Task → Implement → Test (≤3 fix rounds) → Persist

Pause only at the end for commit confirmation. Run everything else autonomously.

⚠️ CRITICAL: Activate .venv before ANY python/pytest command (idempotent, re-run if unsure).

Windows: .\.venv\Scripts\Activate.ps1

macOS/Linux: source .venv/bin/activate

Reference Map

All files under .claude/skills/auto-coder/references/:

File	Content	When to Read
`01-overview.md`	Project overview & goals	First task or when needing project context
`02-features.md`	Feature specifications	When implementing feature-related tasks
`03-tech-stack.md`	Tech stack & dependencies	When choosing libraries or patterns
`04-testing.md`	Testing conventions	When writing tests
`05-architecture.md`	Architecture & module design	When creating/modifying modules
`06-schedule.md`	Task schedule & status	Every cycle (Sync Spec step)
`07-future.md`	Future roadmap	When planning or assessing scope

1. Sync Spec

python .claude/skills/auto-coder/scripts/sync_spec.py

Then read the schedule file to get task statuses:

Read .claude/skills/auto-coder/references/06-schedule.md

Task markers:

Marker	Status
`[ ]` / `⬜`	Not started
`[~]` / `🔶` / `(进行中)`	In progress
`[x]` / `✅` / `(已完成)`	Completed

2. Find Task

Pick the first IN_PROGRESS task, then the first NOT_STARTED. If user specified a task ID, use that directly.

Quick-check predecessor artifacts exist (file-level only). On mismatch, log a warning and continue — only stop if the target task itself is blocked.

3. Implement

Read relevant spec from .claude/skills/auto-coder/references/:
- Architecture: 05-architecture.md
- Tech details: 03-tech-stack.md
- Testing conventions: 04-testing.md
Extract from spec: inputs/outputs, design principles (Pluggable? Config-driven? Factory?), file list, acceptance criteria.
Plan files to create/modify before writing any code.
Code — project-specific rules:
- Treat spec as single source of truth
- Use config/settings.yaml values, never hardcode
- Match existing codebase patterns and style
Write tests alongside code:
- Place in tests/unit/ or tests/integration/ per spec
- Mock external deps in unit tests
Self-review before running tests: verify all planned files exist and tests import correctly.

4. Test & Auto-Fix


Round 0..2:
  Run pytest on relevant test file
  If pass → go to step 5
  If fail → analyze error, apply fix, re-run

Round 3 still failing → STOP, show failure report to user

5. Persist

Update DEV_SPEC.md (global file): change task marker [ ] → [x]
Re-sync: python .claude/skills/auto-coder/scripts/sync_spec.py --force
Show summary & ask:

✅ [A3] 配置加载与校验 — done
   Files: src/core/settings.py, tests/unit/test_settings.py
   Tests: 8/8 passed
   Commit: feat(config): [A3] implement config loader

   "commit" → git add + commit
   "skip"   → end
   "next"   → commit + start next task

On "next", loop back to step 1 and start the next task.