Skip to content
Back to skills

Keel Autopilot

ASecurity

Gated autonomy for one session — advance phases back-to-back with the full /keel-phase-review run at every gate, auto-commit at phase ends, batch pushes for ONE approval. Stops on gate FAIL, UNCERTAIN decisions, architecture surprises, or security-adjacent changes. Push is never automatic. Solo sessions or an agent team's orchestrator ONLY — never a worker/co-agent (it runs phase-review and commits, both forbidden to them by rules §10.42).

  • 9 stars
  • 0 votes
  • 0 copies
  • 3 views
  • Added September 5, 2026
ai-agentsgogitsecurity

Security analysis

A100/100

Scanned September 5, 2026

npx -y skills add muratsilahtaroglu/claude-code-starter-kit --skill keel-autopilot --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Keel Autopilot?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Keel Autopilot
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/muratsilahtaroglu-keel-autopilot/badge)](https://www.skillsdirectory.com/skills/muratsilahtaroglu-keel-autopilot)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: keel-autopilot
description: Gated autonomy for one session — advance phases back-to-back with the full /keel-phase-review run at every gate, auto-commit at phase ends, batch pushes for ONE approval. Stops on gate FAIL, UNCERTAIN decisions, architecture surprises, or security-adjacent changes. Push is never automatic. Solo sessions or an agent team's orchestrator ONLY — never a worker/co-agent (it runs phase-review and commits, both forbidden to them by rules §10.42).
---

# /keel-autopilot — phases advance themselves; evidence gates don't move

When: the user explicitly turns it on ("/keel-autopilot", "autopilot", "work through the phases on your
own"). Autonomy is the USER's grant, never self-granted — and it changes WHO presses the button, not
WHAT the gates require.

**Who may run it: a SOLO session, or an agent team's ORCHESTRATOR — never a worker/co-agent.** This
mode runs `/keel-phase-review` (rule 2) and commits (rule 3), and §10.42 forbids a worker BOTH. So
check identity before the first step: the reground hook prints this session's agent from
`.claude/agent-team-sessions`; if it names a worker charter, say so and stop. Do NOT infer the answer
from `git config user.name` — same-machine agent sessions all share ONE git identity, which is exactly
why rule 4's ownership check cannot see this by itself.

Mode contract (4 rules):
1. **In-phase autonomy.** Work the wip phase's `TASKS ## Now` items back-to-back without pausing for
   confirmation; make routine judgment calls yourself (§10.36). Memory discipline stays hot-path:
   LESSONS the moment something is learned, TASKS updated as you go — autopilot without the rituals
   is just drift at higher speed. On a multi-user project, an `@`-assigned item still STARTS with the
   §10.41 spec briefing before the work (task-specific, self-explanatory questions; check the spec's
   Comprehension log first — an already-answered question is never re-asked).
2. **Gates stay evidence-based.** When `## Now` empties, run the FULL `/keel-phase-review` checklist
   yourself. PASS needs real evidence (tests green via `make test`, the gate's done-when observed) —
   never narrate a pass. PASS → flip the phase, seed the next `## Now` from PLAN.md, continue.
   FAIL → stop and report what's missing.
3. **Commit locally, batch the push.** Commit at each phase end (owner identity, §6.16); do NOT push.
   Accumulate the queue and present it at session end (or on request) for ONE approval.
   `settings.json`'s `ask` on push is the enforced backstop — never weaken it for this mode.
4. **Stop-and-ask triggers.** Halt and return to the user when: a gate FAILS · a decision is
   UNCERTAIN (verifier verdict or §10.37 grounding gap) · a scope/architecture surprise would need a
   new ADR · anything security-adjacent (secrets, auth, non-routine dependency changes) · a bulk
   output hits the `/keel-pilot` threshold (its human-routing rules override autopilot) · the same
   test is red twice after fixes · **the next `## Now` item or wip phase belongs to someone else** —
   autopilot NEVER does another's assigned work; stop and surface it (the parallel-work collision the
   ownership tag exists to prevent). Match ownership on the RIGHT key: on a human multi-user project
   that is `@owner` ≠ `git config user.name`; on a same-machine **agent team** every session shares one
   git identity, so compare the item's lane/`@tag` against this session's agent name from
   `.claude/agent-team-sessions` instead — the git check alone is inert there and silently passes.

Boundaries:
- **One session.** Cross-session automation (cron, /loop) is out of scope — re-request the mode each
  session, or make it a standing project agreement via a `LESSONS.md [rule]` line.
- **After any compaction, re-confirm.** The activation may live only in summarized-away conversation;
  ask one line ("autopilot still on?") before continuing autonomously. Note the activation in the next
  HANDOVER block (d) so the grant survives on disk.
- Never bypassed: push approval, `/keel-pilot` human routing, hook blocks, §0 bootstrap approvals,
  `/keel-distill`'s never-lossy-delete. The Stop hooks (`plan-phase-nudge`, handover reminder) keep
  firing — they are the net under the tightrope, not an annoyance to suppress.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.