autonomy-loop

autonomy-loop

Two Claude Code terminals on one repo — a builder and an adversarial reviewer — passing a git baton through a frozen-invariant safety gate. A Claude Code plugin you drop into any project and tune from one config file.

Honest framing up front: the idea of a self-driving builder/reviewer loop is not new (see Prior art). What this gives you is a specific, opinionated discipline — an adversarial multi-lens reviewer, a frozen-invariant gate, and a no-fabrication rule — wired together and tunable per project. The value is in the constraints, not the novelty.

---

What it actually is

Two terminals run claude in a /loop against the same repo on different git worktrees:

• Terminal 1 — Builder. Picks up the next task, writes code + a RED-before/GREEN-after test, runs the gate, commits to the work branch, hands off.
• Terminal 2 — Reviewer. Re-runs the gate itself, spawns a 5-lens critic panel (correctness · honesty · regression · security · UX), red-teams the opposite, fixes what's safe, flags the rest, hands back.

They never talk in chat. The only shared state is a committed markdown file — LOOP-STATE.md — that holds the baton (turn:, last-builder-sha, last-reviewed-sha, the next instruction each way). State lives in git, so a crash just resumes from the last commit.

flowchart LR
    subgraph T1["Terminal 1 - Builder"]
      B1[read baton] --> B2[write code + RED-GREEN test] --> B3[run gate] --> B4[commit + push work branch]
    end
    subgraph T2["Terminal 2 - Reviewer worktree"]
      R1[read baton] --> R2[re-run gate] --> R3[5-lens critic panel<br/>+ red-team-the-opposite] --> R4[fix safe / flag rest]
    end
    B4 -->|turn: reviewer| R1
    R4 -->|turn: builder| B1
    R3 -.->|frozen-drift / unsafe / deadlock| H[(turn: human<br/>FOR-REVIEW.md)]
    GATE{{gate-guard hook:<br/>blocks prod push, force-push,<br/>history rewrite, protected writes}}
    B4 -. every tool call .-> GATE
    R4 -. every tool call .-> GATE

Why you might want it

• The reviewer is adversarial by construction. Its only win condition is finding fault. It re-runs the gate from scratch (never trusts the builder's "tests pass"), spawns role-specialized critics, and must argue why the change is wrong before it's allowed to approve.
• A frozen invariant the loop can't quietly re-baseline. You declare what must stay byte-identical (golden/snapshot tests, a recorded API contract). Drift requires a human GO — the loop escalates, it doesn't overwrite.
• A no-fabrication rule. Every number a build emits must carry its sample size + interval, or say "building — N/30." A capability with no real data abstains visibly instead of inventing a plausible value.
• Cost-aware. Critics run on a cheap model in parallel; the expensive judge is only invoked when a wave escalates. See Cost.
• Portable. One autonomy.config.json drives the branch names, gate commands, models, protected paths, and honesty rule. The plugin is project-agnostic.

Install

# 1. add this repo as a plugin marketplace, then install
claude plugin marketplace add https://github.com/inferencegod/autonomy-loop
claude plugin install autonomy-loop

# 2. in your project, scaffold the config + baton
/autonomy-loop:autonomy-init

# 3. edit autonomy.config.json (branches, gate commands, protected paths), then:
#    Terminal 1:  claude  ->  /autonomy-loop:builder   ->  /loop 600
#    Terminal 2 (in the review worktree):  claude  ->  /autonomy-loop:reviewer  ->  /loop 600

/loop 600 = self-schedule every 600s (10 min). Commands are namespaced /autonomy-loop:<command>.

Configure

Copy autonomy.config.example.json -> autonomy.config.json at your repo root (it's gitignored — per-checkout). Key knobs:

knob	what it does
workBranch / prodBranch	the loop only ever pushes workBranch; prodBranch is gated
worktreePath	where Terminal 2's worktree lives
gate.test / gate.build / gate.lint	your real commands — the reviewer re-runs them
gate.coverage	optional third gate: a command that emits a coverage summary; the reviewer ratchets coverage so it can never silently drop (see Coverage ratchet)
gate.frozenInvariant	what must stay byte-identical without a human re-baseline
protectedPaths	edits and shell writes (rm/mv/cp/redirect/sed -i) here are blocked by the hook
models.builder / reviewerCritics / reviewerJudge	builder + cheap critics + escalation judge
honestyRule	the anti-fabrication contract injected into both prompts

Safety model & limitations