Two Head Wu

Two-Headed-Wu Interface

Treat Two-Headed-Wu as deterministic infrastructure, not as another reasoning Agent. Keep reasoning and implementation in the current Codex or Claude Code session.

Resolve Before Use

1. Run wu resolve --agent two-head-wu --runtime <runtime> --project <id> when the project is registered. Omit --project only for global maintenance.
2. Add --intent <short-query> to obtain deterministic catalog matches. No match grants no additional permission.
3. Use only the returned Skill sets, packages, MCP servers, resources, and approval profile.
4. Keep formal deliverables in the current project. Two-Headed-Wu stores only its own state and audit records.

Use --json when another tool or Agent needs machine-readable output.

Owner Small Notebook

The notebook is a central Two-Headed-Wu resource even when the current Agent is working in another project. Do not copy it or create a per-project binding.

• Treat “记下来…”, “写到小本本上…”, and “记到小本本上…” as writes only when they are an instruction to this Agent and contain a fact to store. Quoted examples, negations, and discussion of these trigger phrases do not write.
• Strip the leading trigger and punctuation. Resolve personal-memory for project two-head-wu, then invoke personal-memory.remember.v1 through wu. Pass the memory body as data, preferably on stdin; never evaluate it as shell.
• Use one category: identity, preference, project, decision, or fact. For project memory, include only a registered project ID and short routing context—never an absolute path or copied implementation details.
• Confirm only after the invocation returns stored: true; include its pm-... ID. On failure, say that the notebook was not changed. Ordinary conversation and OpenClaw automatic capture never justify an “已写到小本本” confirmation.
• For a notebook recall request, invoke personal-memory.recall.v1 with recall --scope notebook --max-results 3. Do not load the whole notebook.
• Correct and forget by record ID. Without an ID, search the notebook first and proceed only when one result is unambiguous; otherwise ask which record.

Canonical invocation shape:

wu invoke personal-memory --project two-head-wu --runtime <runtime> \
  --interface personal-memory.remember.v1 -- remember [options]

Query And Run

• When the user asks what Two-Headed-Wu can do, run wu 能做什么 --json immediately. On Mini this reads the generated local catalog; on an enrolled Air it delegates to the signed dynamic directory. Present the named execution_entries/remote capabilities separately from the complete inventory: discovery never means that an adapter exists or permission was granted. Use wu 功能清单 --json when the local tracked source is explicitly required.
• Use wu search <query> to find registered Skills, capability packages, projects, and MCP servers.
• Use wu skills search <query> --project <id> --runtime <runtime> to search the approved cold Skill Catalog for the current Binding. Then use wu skills load <skill-id> ... to read its entrypoint and wu skills read <skill-id> <resource> ... for a referenced supporting file. These commands are read-only: loading instructions never authorizes a bundled script, network call, secret access, or other mutation.
• Use wu run <workflow-id> only for a workflow already present in catalog/workflows_registry.yaml.
• Supply --approve only after the user has authorized a workflow whose registry entry requires explicit approval.
• Use wu doctor when paths, runtime surfaces, or package manifests appear stale.
• After changing Two-Headed-Wu itself, run wu 更新说明 before staging files and wu 验收修改 before declaring completion. Repository pre-commit and pre-push hooks repeat the read-only verification automatically when installed; they never refresh or stage documentation.

Boundaries

• Never read or export raw passwords, tokens, private keys, browser stores, or unrelated personal data. Catalog entries may contain secret references only.
• Do not infer access from an installed Skill. Runtime exposure, Agent allowlists, project policy, and action approval all remain required.
• Do not bypass wu by treating documentation or a Skill prompt as an authorization system.
• If wu reports an unknown project or capability, stop that integration step and register or review it first.