claws-mcp-cognition-enginelisted

## What this is The meta-layer for Claws MCP tool selection. Full per-tool reference lives in `references/tool-families.md`; this file is the orientation map. ## When to invoke - About to spawn 2+ workers and unsure which class tool to use. - A previous MCP call failed and you suspect tool-selection error. - Designing a wave plan and need the class/cost trade-off table. - First-time orchestrator or sub-worker in an unfamiliar role. ## When NOT to invoke - Single shell command — `claws_exec` is sufficient. - File read or grep — use Read/Bash tools, not MCP. - Conversational question with no terminal-side state needed. ## Steps Decision flow — run in order: 1. **Horizon?** <5 min → tool call. 5–60 min → `claws_worker`/`claws_lead`. 1–6 h → `claws_titan`. 4–8 h → `claws_legion`. 2. **Intent?** Spawn → orchestration family. Observe → Monitor + read-only. Modify → `claws_send`/`claws_publish`. 3. **TUI session?** Yes → `claws_worker(mission=...)`. Shell-only → `claws_exec`. 4. **Wave?** Single goal → `claws_worker`. Multi-worker → `claws_lead` + `claws_dispatch_subworker`. 5. **Observation:** Arm `Monitor(monitor_arm_command)` in the SAME turn as spawn. `claws_workers_wait` is fallback only. Class ladder: `LEGION claws_legion opus 4h` → `TITAN claws_titan opus 2.5h` → `LEAD claws_lead sonnet 1h` → `WORKER claws_worker sonnet` ## Examples ```bash claws_exec(command="npm test") # shell-only, no Claude claws_worker(mission="Fix null-check in foo.ts")