LoopX Change Quality

Use this skill only when the selected goal's change_quality_qualification.enabled policy is true. LoopX owns the canonical source but does not install it globally. Install a managed copy in a connected project for the relevant host:

loopx project-skill install \
  --project . \
  --skill loopx-change-quality \
  --surface codex \
  --execute

Use --surface claude-code or --surface opencode for those hosts. Skill discovery does not activate the capability; product behavior remains default-off until goal policy enables it.

The CLI is the contract authority. This skill supplies a host-neutral review workflow. Repository instructions, tests, linters, type checkers, and security checks remain the project's quality oracles.

Prepare The Exact Scope

From the repository worktree, run:

loopx --format json change-quality prepare \
  --goal-id <goal-id> \
  --repo-path . \
  --base-ref origin/main

Stop when the packet says disabled or no_changes. When it says review_required, review only the files and exact fingerprint in the packet. Read repository-local instructions before judging the change.

Run loopx project-skill status --project . --skill loopx-change-quality when the host depends on skill discovery. If the managed copy is absent or stale, preview an explicit project install; do not fall back to a global copy.

Review Rules

Simplify First

Spend the review budget on simplification before broad quality analysis:

1. Reuse an established helper or durable repository rule instead of copying behavior or knowledge.
2. Remove redundant state, parameters, branches, indirection, and speculative abstraction while preserving behavior.
3. Challenge a private helper or wrapper when it is only one to three lines, has at most two production callers, and owns no independent domain invariant, effect, or error boundary.
4. Prefer the smallest coherent edit that leaves ownership and intent clearer. Do not create churn when the current shape is already direct and cohesive.

Write one evidence-backed reuse conclusion and one evidence-backed simplification conclusion. Do not emit a row for every remaining lens. Those dimensions are guardrail categories for sparse risks[]: add an item only when the changed surface, repository instructions, a native validator, or the proposed simplification raises a concrete risk. LoopX derives each guardrail's status from risks[] and validation[].

The available review lenses are:

• reuse: established helpers and durable knowledge are reused instead of duplicated;
• type/API boundary: types, schemas, compatibility windows, and caller contracts remain explicit and coherent;
• configuration: configuration stays single-sourced, validated, and free of hidden mode coupling;
• runtime ownership: lifecycle, concurrency, state, and side effects live in the correct boundary;
• quality/simplification: unnecessary indirection, branching, duplication, and speculative abstraction are removed or explicitly justified;
• efficiency: hot paths, repeated work, memory growth, and unbounded loops are considered;
• error/supervision: failures remain observable and actionable without silent fallback or blanket exception handling;
• test/validation: tests and repository-native validators prove intended semantics and important negative paths;
• documentation/comments: names, comments, and docs describe current contracts without stale or duplicated narration;
• security/release: security, privacy, permissions, migrations, and release compatibility are handled at changed boundaries.

The packet projects path-only references to applicable repository instructions, ownership files, build manifests, language hints, and changed surface roots. It also projects a provider-neutral validation plan discovered from structured repository task declarations. This is discovery context, not copied repository content: instruction text, task bodies, and manifest contents remain in the worktree. Read every required_reads entry, inspect each candidate's source_ref, and let the host resolve the named Poe, Hatch, Cargo, or package task. Never execute a candidate merely because it was discovered. Unresolved format, lint, typecheck, or test categories require reviewer judgment or a repository-native instruction; do not fill them with guessed commands. Treat ignored_manifest_refs as non-executable context, especially fixtures and vendored projects.

Read every projected instruction reference, but do not copy its prose into the result. Ground reuse, simplification, and each emitted risk with typed evidence_refs using path:, instruction:, or validator:. Keep risks[] empty when no guardrail is triggered. Record only validators that were selected or required; failed validation and skipped required validation are independently blocking.

Use blocker only for a concrete correctness, security, privacy, contract, or required-validation failure. Style preferences and speculative redesigns are warning or advisory, never blockers.

This first version is single-level. Review one final diff; do not recursively review the review, spawn a hierarchy of quality agents, or require agreement between several models.

The guardrail catalog is review guidance, not an Agent output checklist or evidence that a particular model applies it well. Maintainers qualify model behavior separately against the public clean-PR and seeded Python, Rust, and TypeScript shadow matrix. Normal delivery does not launch that low-frequency evaluation, and a model shadow result cannot mutate goal policy or grant merge authority.

Safe Fix

safe_fix and strict_receipt are independent:

• safe_fix=true permits at most one bounded repair pass.
• strict_receipt=true requires exact-scope evidence before merge and grants no permission to edit.

When safe_fix is false, report risks without modifying files. When it is true, one repair pass may address a clear simplify opportunity or risk inside the selected todo and goal boundary. Do not use destructive git, broaden permissions, change product intent, add unrelated refactors, or conceal a failing validator.

After any edit, rerun prepare. The old fingerprint is invalid. Review the entire new final scope, not only the lines changed by the repair.

Record The Receipt

Write a compact result conforming to the packet's change_quality_agent_result_v2 template. Beyond exact-scope metadata, the Agent writes only reuse, simplification, sparse risks[], and validation[]. simplification.safe_fix_applied records the one permitted repair pass. A skipped or failed validator needs a reason; failed validation or skipped required validation makes the receipt non-passing. Keep raw transcripts, private paths, credentials, and unbounded logs out of the result.

Then record and read back the exact receipt:

loopx --format json change-quality record \
  --goal-id <goal-id> \
  --repo-path . \
  --base-ref origin/main \
  --result-json <ignored-or-temporary-result.json> \
  --execute

loopx --format json change-quality verify \
  --goal-id <goal-id> \
  --repo-path . \
  --base-ref origin/main

A receipt with an unresolved blocker, failed validator, or skipped required validator is not passing. A receipt for an earlier fingerprint does not qualify a later diff. Earlier experimental receipt schemas are invalid and must be requalified with the current protocol.

Premerge Enforcement

Run the authoritative merge gate with the goal identity:

loopx canary premerge \
  --from-git-diff \
  --goal-id <goal-id>

Turn may transport the prepare packet or receipt reference as part of one bounded transaction. Turn does not own quality policy and may not manufacture or waive a receipt. canary premerge remains the enforcement authority.

Completion Evidence

Report:

• final scope fingerprint and changed-file count;
• safe-fix allowed/applied and pass count;
• blocker, warning, and advisory counts;
• project validations run and their real results;
• receipt id and exact verification status;
• premerge status, failures or skips, and manual holds.

Stop before delivery when strict policy requires a receipt and the receipt is missing, invalid, stale, or contains an unresolved blocker.