v0.4.30 to v0.4.30

6 added, 1 removed. Audit A to A.

---

name: emperor-dispatch

description: >-

Emperor Time — DISPATCH / Steal Chain. Consent-first enlistment of local

CLIs (Claude Code, Kimi, Codex, Copilot, opencode, Ollama). Use when the user

asks to use another agent, parallelize, or review via a different model.

license: MIT

metadata:

version: 0.4.30

chain: steal-chain

part-of: emperor-time

---

# Emperor Dispatch (Steal Chain wrapper)

1. Read `chains/steal-chain/SKILL.md` → one aspect

(`consent-protocol.md` first unless consent is already on the ledger).

2. Dowse the machine (`scripts/dowse.sh` / `dowse.ps1`) if the roster is stale.

3. Client consents per agent per task (or a standing policy they stated).

4. Workers receive the **work order + named files**, never the author's diary.

5. Capture to `.emperor/runs/<task>/<agent>/` (`prompt.md`, `out.txt`, `meta.md`).

6. Output is CONJECTURE until Judgment + `scripts/gate.sh g4`.

- 7. Sign-in is the client's terminal. You never run interactive logins.

+ 7. **MUST — quarantine HARD-GATE:** before merging worker output, run

+ `scripts/emperor quarantine <task-dir>` (or `steal-quarantine`).

+ Doctrine: `chains/steal-chain/quarantine.md`. Missing runs layout /

+ CONJECTURE start / ADMITTED|REJECTED → exit 1. G4 calls

+ `scripts/lib/quarantine.py` when steal activity is present.

+ 8. Sign-in is the client's terminal. You never run interactive logins.

## MUST — parallel-dispatch checklist for independent domains

When facing 2+ *independent* tasks / failures / subsystems (disjoint writable

scopes, no shared root cause) — before launching concurrent agents — open

`skills/emperor-dispatch/parallel-dispatch-checklist.md`

(Chain Jail leaf from Superpowers `dispatching-parallel-agents` → Identify

Independent Domains / Focused Agent Tasks / Parallel Dispatch / Review and

Integrate only)

and/or run `scripts/emperor parallel` (prints the mechanical PARALLEL / STEP /

MUST card).

One agent per independent domain. Focused self-contained briefs. All dispatches

in the same response for true parallelism. Review and integrate before done.

Do not load whole `dispatching-parallel-agents`; ET + emperor-dispatch

orchestrate. Sequential plan tasks stay `emperor subagent` (no parallel

implementers on the same plan). Inline stays `emperor execute`.