source-command-setup-autonomy

name	source-command-setup-autonomy
description	Show or set the autonomy dial — how much the AI may do without asking (consent grade 1–4, orthogonal to the capability level).

source-command-setup-autonomy

Use this skill when the user asks to run the migrated source command autonomy.

Command Template

/autonomy — the consent dial (ADR-0041/0042)

Show or change how much the AI may do without asking in this project. This is NOT the capability level (/context-level, L1–L7 — what the kit can do); the dial is consent — what it may do on its own.

Run the script and relay its output verbatim (it echoes the first-person consequence text the user must see):

node contextkit/tools/scripts/autonomy.mjs            # show current grade + what each grade means
node contextkit/tools/scripts/autonomy.mjs 3          # persist grade 3
node contextkit/tools/scripts/autonomy.mjs 3 --session # this session only (auto-expires in 8h)
node contextkit/tools/scripts/autonomy.mjs --clear    # drop the session override
# grade 4 (EXPERIMENTAL, ADR-0045) — gated by the eligibility bar, SESSION-scoped by default:
node contextkit/tools/scripts/autonomy.mjs 4                     # session-only grade 4 (refuses if the bar fails)
node contextkit/tools/scripts/autonomy.mjs 4 --persist --confirm # persist grade 4 (only after seeing the consequence)
node contextkit/tools/scripts/autonomy-readiness.mjs            # measure self-coverage + attribution for the bar (EXPENSIVE)

Rules for you (the agent):

1. Never run the setter on your own initiative. Raising the grade is a floor area (grade-change → always human, ADR-0042). You run this command only when the user asks; you may suggest a grade change, never apply one.
2. When the user asks "what would grade N mean?", answer with the script's consequence text — do not paraphrase it weaker.
3. Grade semantics you honor everywhere (via resolveAutonomy, the single read path): 1 manual · 2 suggest+supervise · 3 auto-except-decisions (default, ADR-0058 — ADRs, pushes, high-risk and secret paths still come to the user) · 4 full-auto, experimental, telemetry/budget-gated (ADR-0045).
4. The non-negotiable floor holds at every grade: secret-bearing paths, gate/hook self-edits, force-push, ADR-class decisions and grade escalation are always the user's. No flag or config removes them.
5. Grade 4 is special (ADR-0045, amended ADR-0058). It is EXPERIMENTAL. The setter runs the deterministic eligibility bar first (≥30 transitions · ≥20 sessions · zero wiring-drift · self-coverage green · attribution present) and refuses naming the failing criterion if any miss — relay that verbatim. When the bar holds, the setter shows an informed-consent disclaimer and the human signs (--confirm is the [y] signature; omitting it is [n]/cancel). /autonomy 4 with no flags is session-scoped (auto-reverts); persisting needs the --persist --confirm signature after the disclaimer is shown. The self-coverage + attribution criteria come from autonomy-readiness.mjs — tell the user to run it if those are the blockers. Even at grade 4 the floor holds and /ship runs the hardened quorum (see /ship): a Critical from the security voice is a veto, and the human can yield/kill mid-run at any step.