AI 工程从零到一（中文）

Given a repo and the agent product running inside it, produce schema-first state files for the workbench.

Produce:

1. schemas/agent_state.schema.json covering required keys, allowed status values, array-vs-null discipline, and a schema_version integer.
2. schemas/task_board.schema.json covering task id pattern, allowed owners, allowed statuses, and acceptance arrays.
3. tools/state_manager.py exposing load, commit, and update with temp-and-rename atomic writes.
4. tools/migrate_state.py scaffold for the next schema bump, fail-loud if the file is from an unknown version.
5. agent_state.json and task_board.json seeded at schema_version: 1 and a fresh backlog.

Hard rejects:

• A schema without a schema_version field. Migrations are not optional.
• Allowing null where an array is expected. null is a write-time bug masquerading as data.
• A writer that uses plain open(path, "w"). Atomic writes only; partial files corrupt the source of truth.
• Storing tokens, raw chat transcripts, or PII inside state. State is for repo-relevant facts.

Refusal rules:

• If the repo has no version control, refuse to ship state files. Atomic writes plus git diff is the durability story.
• If the project does not have at least one acceptance command to validate the done transition, refuse the status: done enum value. Adding done without an acceptance check is theater.
• If the project intends to share state across processes without a lock strategy, surface that finding before shipping; atomic rename is necessary but not sufficient.

Output structure:

<repo>/
├── agent_state.json
├── task_board.json
├── schemas/
│   ├── agent_state.schema.json
│   └── task_board.schema.json
└── tools/
    ├── state_manager.py
    └── migrate_state.py

End with "what to read next" pointing to:

• Lesson 35 for the initialization script that calls the manager on startup.
• Lesson 38 for the verification gate that reads state to score completion.
• Lesson 40 for the handoff generator that consumes the same schema.