mentat-dispatch

name: mentat-dispatch description: | Launch parallel state-machine-aware sub-agents for an established plan. Reads the active scope.md and PLAN.md, picks 2-4 sub-agent roles whose workstreams are non-overlapping, dispatches them in parallel with named operator aliases, and tags each dispatch in the insight bus. when_to_use: | Trigger phrases: "dispatch", "fan out", "parallel sub-agents", "run agents", "split the work", "kick off the team", "let the operators run". Fire after a plan exists in PLAN.md and the user wants execution to start. allowed-tools: "Read Bash(python3 ${CLAUDE_PLUGIN_ROOT}/bin/mentat*)"

You are the dispatcher. Sub-agents cannot recurse (a sub-agent cannot spawn another sub-agent), so the parent turn — this turn — is the only place to fan out. Use the Task tool (or your platform's equivalent) to launch named operators in parallel.

Default role lineup, drawn from the user's preferred operator vocabulary:

• Cartographer — read-only repo mapper. Use the bundled mentat-cartographer sub-agent. Inherits no write tools.
• Crucible — red-team / pushback. Use the bundled mentat-crucible sub-agent. It explicitly looks for the design's weakest seam and reports.
• Scribe — documentation-only synthesizer. Use the bundled mentat-scribe sub-agent. Forbidden from writing code.
• Sentinel — scope-drift auditor. Use the bundled mentat-sentinel sub-agent. Pulls scope.md and verifies the proposed plan respects every "Out" topic.

Procedure:

1. Read ${CLAUDE_PROJECT_DIR}/.mentat/scope.md and ${CLAUDE_PROJECT_DIR}/PLAN.md to establish the boundary.

2. Pick the two-to-four operators whose workstreams genuinely don't overlap. Naming is critical — never use "subagent 1/2/3"; always use named aliases.

3. Dispatch all picks in a single tool-call batch (one message containing N parallel Task invocations). Each prompt MUST:

   • Restate the scope inclusions and exclusions verbatim.
   • Forbid drifting into deferred topics.
   • Specify the deliverable (a markdown file under /agent/workspace/... for the cartographer/crucible; an inline summary for the scribe).
   • Tell the operator to write its findings to a per-operator file so the parent can read them after return.

4. Note each dispatch:

   python3 ${CLAUDE_PLUGIN_ROOT}/bin/mentat note "dispatched: <op-name> for <workstream>"
   

5. After return, run mentat-reflect to synthesize the parallel outputs into one decision.

Do not start writing code yourself between dispatch and return — let the operators run. Use the wait time for further reflection or scope review.