Aitne-sh
UserRun Claude Code as a contract-driven agentic loop — it implements, verifies, reviews, and improves on a fresh context each iteration until every requirement is provably met. Pure shell and prompts: no server, no daemon, no Python.
Categories
Indexed Skills (14)
loop-contract-review
Independent read-only review of a loop definition (product contract + stop conditions) BEFORE it is approved — judge whether the contract faithfully covers the instruction and whether VERIFY_COMMANDS is a real, discriminating gate. Invoked by loop.sh on the unattended (auto) approval path.
loop-decompose-review
Independent read-only review of a candidate task plan BEFORE tasks are enqueued — judge whether the decomposition faithfully covers the approved master contract and whether the claimed scope boundaries are real in this repository. Invoked by loop.sh on the orchestration path.
loop-decompose
Supervisor decomposition — split an APPROVED master contract into the smallest set of non-conflicting, independently verifiable tasks (explicit scope boundaries and dependencies) and return a task-plan payload for the harness to publish. Invoked by loop.sh after master-contract approval; n=1 means "run as a single in-place loop".
loop-evidence
Generate .loop/docs/evidence-report.md for a completed loop run — the compressed artifact a human reviews instead of the full diff. Invoked by loop.sh when the loop reaches a reviewed success candidate; the baseline argument names the task's fixed starting commit.
loop-iterate
Execute exactly one iteration of the contract-based implementation loop — read the contract/plan/progress docs and reviewer feedback, implement one milestone, self-verify, check spec drift, update the memory docs, declare a state. Driven headlessly by loop.sh run; not meant for ad-hoc conversational use.
loop-plan-review
Independent read-only review of a candidate implementation plan BEFORE the harness publishes it — judge whether the iteration-0 plan faithfully serves the approved product contract and whether its milestones are real, iteration-sized work in this repository. Invoked by loop.sh at iteration 0 after the planner's candidate passes deterministic validation.
loop-review
Independently review one loop iteration's implementation diff against the product contract and report a verdict (APPROVE, REVISE with must-fix items, or — at the gate — ESCALATE to the human). Requirement-first — per-REQ satisfaction is judged before code quality; the gate renders a machine-parsed verdict per REQ. Runs read-only with fresh context; driven by loop.sh, not for ad-hoc use.
loop-rollback-review
Independent read-only safety review of a proposed Fleet rollback after the deterministic assessor has bounded its eligible plan-owned changes. Invoked by loop.sh during discard --rollback to reject unsafe inverse changes, parallel/manual overlap, external side effects, or uncertain provenance; this reviewer can only narrow the deterministic safe set and can never upgrade an ineligible rollback to SAFE.
loop-setup
Interactive first-run tuning of the per-phase agent and model routing (loop.models.sh) before a loop is run. Explain the loop and the settable knobs, answer the human's questions from a bundled reference (never by exploring code), converge on the values they want, and edit only the loop.models.sh copy in this working directory using plain key="value" lines. Invoked by ./loop.sh setup in an isolated throwaway directory; the harness deterministically validates the result before it reaches the real file.
loop-stop-eval
Lightweight advisory stop evaluation for the loop — judge from the contract's acceptance criteria, the latest verification log, and the progress log whether the loop should continue, appears to have met its goal, or is futile. Read-only, cheap-model role driven by loop.sh; not for ad-hoc use.
loop-supervise
Supervisor decision on an escalated fleet task (or a failed integration gate) — judge the worker's blocked question against the human-approved master contract, then ANSWER it, REPLAN the task, or ESCALATE to the human. Read-only; the harness applies the decision and journals it. Invoked by loop.sh's fleet dispatcher.
loop-contract
Define a loop for this repository from a user instruction — explore the codebase, ask the minimum necessary questions, then write the fixed Product Contract (.loop/docs/product-contract.md) and its machine-readable stop conditions (loop.config.sh), and present the full loop definition for user approval. Use when the user wants to set up an autonomous implementation loop for a task, or to revise an existing loop definition.
loop-plan
Read-only implementation planning from the approved product contract — return a complete implementation-plan payload for the harness to publish. Invoked by loop.sh at iteration 0; manual invocation drafts a replacement but does not publish it.
loop-refine
Interactive design-gate refinement session for a loop that is paused (BLOCKED) at a human aesthetic/visual sign-off. Help the human adjust the REVERSIBLE, within-contract knobs (tunable constants) and preview the result live, keeping the frozen contract, the representation, and every cmd/run acceptance check untouched. Invoked by ./loop.sh refine; not for headless loop iterations.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.