LoopSpine

Your coding agent keeps going without skipping the proof.

LoopSpine is one portable Agent Skill for end-to-end software work. It routes a task through the smallest useful planning, TDD, debugging, QA, review, and ship loops, then stops at proof, no progress, or a real approval boundary.

It is not another catalog of prompts. It is the control loop between intent and a trustworthy result.

This recorded run reproduced a parser bug, added the regression test, repaired the implementation, passed two proof commands, and cleared a fresh read-only review in about a minute with zero Sawyer interventions. Inspect the receipt and review, or reproduce it with npm run demo && npm run demo:render.

The Promise

inspect -> choose a lane -> act -> verify -> keep the win -> repeat or stop

LoopSpine gives the agent autonomy inside a safe repository boundary:

• Local inspection, edits, tests, and repair continue without routine handoffs.
• Behavior-changing code uses RED -> GREEN -> REFACTOR.
• Bugs are reproduced before they are fixed.
• Parallel writers get separate worktrees and exact ownership.
• The lead agent integrates and runs the final proof.
• Push, merge, deploy, global changes, and destructive actions remain explicit.

Why It Is Different

Typical agent workflow	LoopSpine
A long checklist	A feedback loop with observable progress
“Keep trying”	Goal, signal, budget, proof, and stop path
Every task gets ceremony	Tiny tasks stay direct
One model creates and approves	Independent review when impact warrants it
Claims improvement	Ships paired, sealed, and execution evals

Use It

Ask naturally:

Handle this end to end and keep going until it is actually verified.

Or invoke the skill explicitly:

$loopspine fix this regression and prepare it for review

The final receipt is always small:

LANE: investigate
RESULT: success
PROOF: npm test -- checkout
BOUNDARY: PR prepared; merge and deploy not authorized
RESIDUE: none

Pilot From A Clone

Keep the first installation inside one test repository:

export LOOPSPINE=/absolute/path/to/loopspine
mkdir -p .agents/skills
ln -s "$LOOPSPINE/skills/loopspine" .agents/skills/loopspine

The repository also contains Codex and Claude Code plugin manifests for local plugin development. Leave global skills and specialist workflows unchanged during the ten-task pilot. See narrow integration.

Dogfood Ten Real Tasks

Synthetic benchmarks and the demo do not count toward promotion. Record ten real tasks and let the report calculate the four operating metrics:

npm run dogfood:report
npm run dogfood:record -- /path/to/completed-run.json

The public dogfood report starts at 0/10 and stays Pending until evidence is recorded. See the pilot contract for metric definitions and the narrow promotion gate.

Benchmark It

No “trust me” score is baked into the README. Run the same model with and without the skill:

npm test
npm run benchmark:pilot
npm run benchmark:sealed
npm run benchmark:trajectories

The suite includes planning, TDD, debugging, QA, parallel-agent, bounded-loop, review, shipping, global-change, and non-overtrigger cases. Release scoring uses an independently authored seal created only after the skill hash is frozen. Disposable fixture repositories verify that agents can actually edit, test, debug, and prove a result instead of merely describing the workflow.

The accepted v0.2 sealed run used 36 fresh gpt-5.5 sessions:

Measure	No skill	LoopSpine
Weighted quality	75.16%	96.73%
Strict sample pass rate	22.22%	72.22%
Safety-boundary violations	0	0

That is a +21.57 point quality gain with 8.65% median runtime overhead and 5.65% median output-size overhead. All six real fixture trajectories passed.

See benchmark results for receipts and limitations, benchmark method for the frozen acceptance rules, and the independent release review for the final audit.

Design Rules

1. One lead owns integration and proof.
2. Use subagents only for independent work or context isolation.
3. Use worktrees for parallel writes.
4. A loop must have fresh evidence that can change its next action.
5. Stop on success, clean no-op, no progress, exhausted budget, blocker, or approval boundary.
6. Never turn a benchmark failure into a weaker test.

Status

0.2.0 passes the local sealed, execution, plugin, and skill validation gates. It also loads in Claude Code with Fable and is public at uncfreak1255-code/loopspine. It is not globally installed; GitHub popularity is never guaranteed by an eval.

License

MIT