t2a-protocol/t2a/t2a-tool-author

T2A Tool Author

Instructions

Use this skill when changing a tool to support T2A.

1. Find one bounded reasoning checkpoint.
2. Replace one embedded model call with a local T2A handoff.
3. Keep the rest of the tool unchanged.
4. Create a capsule with task.json, prompt.md, bounded input, and output.schema.json.
5. Return a T2A Handoff Result with type: "t2a.handoff".
6. Accept host-agent-written output.json.
7. Run T2A protocol validation and tool-owned domain validation before commit/resume.
8. Add a demo fixture or template task and run t2a validate.

Before commit/resume, check:

• protocol validation passed
• validation.json matches the current task
• domain validation passed
• only then mutate state

Before editing, inspect the tool and propose:

• the checkpoint to migrate
• files to change
• files to add
• validation command
• risks / assumptions

Wait for approval unless the user explicitly asked you to proceed.

Guidelines

• Do not add OpenAI, Anthropic, Gemini, Ollama, LangChain, or other LLM SDKs.
• Do not add new model, provider, base_url, api_key, or provider config for the migrated T2A path.
• Do not remove unrelated existing provider configuration unless it belongs to the checkpoint being migrated and the user approves.
• Do not build an agent framework, scheduler, daemon, queue, or MCP bridge.
• Do not refactor the whole tool.
• Do not skip output.schema.json.
• Do not let the host agent directly mutate tool state.
• Do not trust output.json without validation.
• Do not auto-resume before validation.
• A first migration demo fixture may be generated and test-only.

Examples

• Replace a summarization provider call with a capsule and output.schema.json.
• Replace issue triage model config with a local t2a.handoff.
• Replace risk classification model calls with host-agent reasoning and tool-owned ID validation.