feat(loop): add autonomous run-loop with planner agent and frozen gate

[senblet]

Adds an end-to-end autonomous path to the template: a human writes TASK.md, a planner subagent turns it into an executable contract, then the loop iterates to green without further input.

The core idea

The planner's deliverable is the gate, not a checklist. Each plan item must name the check that proves it, so "the plan is done" becomes verify.sh exiting 0 rather than the agent's own judgment. An agent that decides its own plan is complete is just self-assessment with extra steps — this keeps the template's central invariant intact under full autonomy.

Three rails

• Red before green. The freshly-written gate must exit non-zero before any work starts. A gate that's already green proves nothing, and this is the cheap defense against the planner writing weak checks to make its own life easier.
• The gate is frozen. scripts/hashes.sh fingerprints verify.sh and TASK.md into STATE.md; every pass re-checks them. Editing the thing that judges you is how a loop manufactures success, so a mismatch aborts the run.
• The planner can refuse. It runs the four-box test itself and returns REFUSED for unloopable tasks — replacing the human who used to say "this shouldn't be a loop."

With nobody watching, escalation means a clean abort with a written reason, never working around the contract.

/run-loop phases

0  preflight: reject placeholder TASK.md, note dirty tree, reset STATE.md
1  spawn planner (synchronous) -> REFUSED aborts, no retry
2  arm: gate must exit non-zero, then fingerprint -> frozen
3  iterate until green / cap / stuck / tampered
4  report the literal exit code

Files

New: TASK.md, .claude/agents/planner.md, .claude/commands/run-loop.md, scripts/hashes.sh

Modified: LOOP.md (two new stop conditions: stuck, tampering), STATE.md (phase + fingerprint fields), scripts/verify.sh (freeze note; fail-closed default unchanged), .claude/commands/iterate.md (hash check, never-edit-the-gate rule), .claude/commands/setup-loop.md (scoped to project setup only), CLAUDE.md, README.md

Notes

• /run-loop deliberately does not commit. An unattended run producing unreviewed commits is a worse default than one leaving a diff.
• The planner's four-box test swaps "repeats weekly" for "bounded" — the README version decides whether to build a loop at all, the planner's decides whether a given task belongs in one. Documented in the README.

Testing

No test framework in this repo (markdown + two shell scripts). Both scripts pass bash -n, hashes.sh runs clean, and verify.sh still fails closed at exit=1 as intended.

Run /iterate by hand on a real task before trusting /run-loop — what needs proving is whether the gate is honest, not whether the agent can code.

🤖 Generated with Claude Code