loop-engineeringlisted
Install: claude install-skill xsefirosus/sefi-agents
# Loop Engineering
A loop discovers work, hands it off, verifies it, persists state, and reschedules itself.
Every loop implements all five moves. This skill maps them to repo mechanics.
User instructions always override this skill.
All factual output follows the anti-hallucination skill: cite or mark UNKNOWN, never guess.
agentic-signals: goal_intake, refusal_gate, verification, loop_discipline, close_out
(goal_intake's behavior: `skills/sefi-orchestration/references/goal-intake.md`;
close_out's: `skills/sefi-orchestration/references/close-out.md`)
## The five moves (mapped to repo mechanics)
- Discovery: a skill invoked by the automation reads CI / issues / commits / state and
judges actionability. Preflight first: `scripts/probe-tools.sh --loop <spec>` probes the
spec's `requires-tools:` line before the move runs.
- Handoff: one git worktree per finding, per the worktree procedure in `docs/LOOPS.md`.
- Verification: the qa-engineer plus an executed stop condition, judged separately from
the generator.
- Persistence: `state/*.md` committed, carrying a cross-iteration notes bridge (what the
next iteration must know).
- Scheduling: cloud cron or a local interval.
## Stop condition = a grep-countable artifact, not a self-declaration
The product-manager's numbered-checkbox plan is "done" only when every box is checked -- counted
by grep, never the generator declaring "I'm done." (AutoGPT's `finish`-tool self-completion
is the anti-pattern; a predecessor's 184-green