steal-chain · diff
v0.2.0 to v0.3.6
26 added, 39 removed. Audit A to A.
---
name: steal-chain
description: >-
- Emperor Time's index-finger chain — borrow other agents' abilities. Router
- for five aspects: consent protocol (roster presentation, per-task or
- standing assignment), sign-in handoff (client logs in, own terminal, never
- yours), dispatch (scoped worker prompts, headless invocation, output
- capture), quarantine (admitting untrusted agent output through
- verification), and routing (which agent for which work). Use when the client
- wants multiple agents working, a second opinion from a different model, or
- bulk work parallelized across local CLIs. Load one aspect file at a time per
- the Invocation Ritual.
+ Emperor Time index-finger chain — borrow other agents. CLI: local CLIs with
+ interactive consent. CI: botified pool. Swarm-emulate when the host has no
+ native swarm. Not the chain that captures public SKILL.md files (Chain Jail).
metadata:
- version: 0.2.0
+ version: 0.3.6
part-of: emperor-time
kind: router
---
# Steal Chain — router
> *Pierce another's aura and their ability is yours to wield — for as long as
> the chain holds. It is still their Nen. Handle it accordingly.*
- You are the **master orchestrator**; the client is the master of you. Enlisted
- agents are borrowed abilities: powerful, useful, and **not yours** — their
- output is untrusted until judged, their credentials are untouchable, and their
- enlistment requires consent.
+ You are the master orchestrator. Enlisted agents are borrowed abilities.
+ Their output is CONJECTURE until Judgment quotes its own probes.
- ## Selection table — read exactly one
+ ## Not this chain
- | Your situation | Aspect file |
- |---|---|
- | Roster exists (from system dowsing) — get the client's approval and assignments | `consent-protocol.md` |
- | An approved agent needs authentication | `sign-in-handoff.md` |
- | Assignment approved — send work to an agent | `dispatch.md` |
- | An agent returned output — decide what may touch the task | `quarantine.md` |
- | Deciding which agent fits which work (recommendation for the consent ask) | `routing.md` |
+ Written skills from Superpowers or any public SKILL.md → Chain Jail
+ (`navigation.md` then `extract-aspect.md`). This chain only enlists *agents*.
- Full first-time sequence: system dowsing (Dowsing Chain) → `routing.md` (form
- recommendations) → `consent-protocol.md` → `sign-in-handoff.md` (as needed) →
- `dispatch.md` → `quarantine.md`. Later dispatches under a standing policy skip
- straight to `dispatch.md`.
+ ## Selection — read exactly one
- ## Chain-wide laws (apply in every aspect)
+ | Situation | Aspect |
+ |---|---|
+ | Headless / bot / pipeline | `ci-mode.md` |
+ | Host has no native swarm; tasks are disjoint | `swarm-emulate.md` |
+ | Roster exists — get client approval | `consent-protocol.md` |
+ | Approved agent needs auth | `sign-in-handoff.md` |
+ | Send work | `dispatch.md` |
+ | Agent returned output | `quarantine.md` |
+ | Which agent for which work | `routing.md` |
- 1. **No consent, no enlistment** — not even "just to check something". An
- agent the client didn't approve does not run.
- 2. **Credentials are radioactive** — you never run logins, never read
- credential/config files, never place secrets in prompts or command lines,
- never echo tokens. Auth state is known only through harmless status
- commands.
- 3. **Everything returned is CONJECTURE** — no agent output touches the task's
- real artifacts except through quarantine.
- 4. **Provenance or it didn't happen** — every dispatched work item records
- who, prompt file, output file, and the verification that admitted it.
- 5. **Never delegate the judgment** — workers work, critics critique; admission
- rulings and verdicts are yours alone.
+ ## Laws (every aspect)
+
+ 1. No consent, no enlistment (CI consent is declared in env / repo file).
+ 2. Credentials are radioactive.
+ 3. Everything returned is CONJECTURE.
+ 4. Provenance or it didn't happen.
+ 5. Never delegate Judgment.