Loop

In-session loop

You were invoked to run the recurve improvement loop without leaving this chat session. You are the orchestrator, not the sculptor — you never edit . yourself. Each cycle is done by a fresh sub-agent you spawn; you read the gate, decide, and move on.

Why fresh sub-agents. The loop's safety comes from one clean agent per cycle — no context rot, contained failures, per-cycle rollback. If you sculpted the cycles yourself, this one long session would accumulate context and start "remembering" what is not in the ledger. The ledger is the only memory a cycle gets. So: one fresh sub-agent per cycle, handed only the gap id and .recurve/RUN.md — never the prior cycles' conversation.

Permissions — let cycles flow. So a cycle never stalls waiting for you to approve a file write or a command, run this session in a skip-permissions mode (start it with claude --dangerously-skip-permissions, or your host's equivalent). That is safe here specifically because the loop is a cage, not a trust exercise: each sub-agent is held off the referee surface by the write boundary, every cycle is a one-command rollback (per-cycle commits), the tree lock keeps a single writer, and nothing closes without recurve matrix --gate. The safety is the cage — keep . scoped and do not point the loop at paths it must not touch.

Preconditions — never start on a broken baseline

Run these; if any fails, STOP and report — do not start:

recurve validate                 # ledger sound: every open gap has a probe + trap
recurve matrix --gate            # the baseline is fleet-green
recurve lock status              # must be UNLOCKED — a held lock means another driver

Then acquire the lock so a terminal loop or a second session cannot collide with you:

recurve lock acquire

One loop per tree. If the lock is already held, something else is driving — stop. Stealing a lock is a human-only act (recurve lock steal), never yours.

The loop — you orchestrate, a fresh sub-agent sculpts

Repeat until a halt condition (below). Each pass:

1. Pick the work (you):

   recurve next        # highest-value open gap; review-gated + parked listed apart
   

   Skip review-gated gaps (security-tradeoff) — those go through REVIEW.md, never an unattended cycle. If next reports no open work, halt.

2. Spawn ONE fresh sub-agent for exactly one cycle. Use this session's sub-agent / task tool. Hand it only: the recommended gap id, and the instruction to read .recurve/RUN.md and follow it exactly for ONE cycle, then stop. Do not paste earlier cycles' context — the ledger is its memory. The sub-agent sculpts the smallest honest change in ., rebuilds, gates, promotes open→closed, snapshots, commits per policy, writes its run record, and STOPS. One cycle = one sub-agent.

3. Judge by the gate, never the sub-agent's word (you):

   recurve matrix --gate           # the arbiter — the sub-agent's summary is not evidence
   

   • Green and the gap closed → good cycle; continue.
   • Gate red / gap not closed → the cycle failed. The sub-agent should have reverted to the last green per RUN.md; verify the tree is clean (git status). If a half-cycle was left behind, restore the last per-cycle commit's state — never git reset shared state by hand. Count a failed cycle.
   • Un-greenable after ~3 honest attempts → the sub-agent parks it (recurve park); you continue past it.

4. Never touch the referee surface. Neither you nor the sub-agent may edit claims / probes / traps / the gate to make a cycle pass — that is the one move the whole system exists to prevent. A weakened probe is caught by its trap at the gate regardless; do not try.

Halt conditions — stop the loop

Stop, and do not start another cycle, when any of these holds:

• recurve next reports no open work left (and no pending drafts you were asked to arm).
• 3 consecutive failed cycles.
• 2 consecutive cycles that grow the backlog net-positive (runaway scope).
• A pending adjudication appears — a human decision is required before more work.
• A cap the human set for this run is reached.

Do not keep going because the last cycle went well.

When you halt

1. Release the lock:

   recurve lock release
   

2. Relay the wrap-up verbatim — do not summarize away the parked reasons:

   recurve matrix          # what turned green; is the gate holding
   recurve park            # parked gaps + attempt journals (the next run's seed)
   recurve stats           # close rate, attempts, wall-clock by class
   

   Then give the human queue, in order: adjudications first, then review-gated promotions (REVIEW.md), then parked triage. Per-cycle commits mean a killed run lost at most one cycle's work.