quality-harness

Execute an active Accepted ADR task-by-task, and record the tool-written evidence that a task is finished. Use when the user invokes `/quality-harness:adr-execute` with an active ADR path, asks to implement an accepted current decision, or asks to mark a task done, tick off a task, update a task's status, or record that work passed — marking done without `adr-verify` is the one thing this skill exists to prevent. Do not execute archived, superseded, withdrawn, Proposed, or Draft records; retire records with `adr-retire` instead.

Retire or archive ADR records without erasing decision authority, rewriting historical evidence, or losing open obligations. Use when the user asks to archive, retire, freeze, supersede, or reduce noise from an ADR corpus. Do not use to hide a failing current ADR, delete decision history, or execute archived work.

Create a proposed, executable decision record for a durable technical choice. Use when the user asks for an ADR or when work changes a public contract, persistent state, trust boundary, cross-component ownership, or costly-to-reverse architecture. Do not use for bounded implementation, open-ended requirements discovery, or ADR retirement.

Create or audit a current-state architecture document. Use when the user invokes `/quality-harness:arch-write`, asks for an architecture map/audit, or an accepted structural decision changes module boundaries or ownership. Use grill-only mode for a read-only architecture audit. Do not use for a single decision record, bounded implementation, or speculative future design without code evidence.

Ask a fresh-context, read-only Codex GPT-6 Astra session for grounded advice in any project. Use for an independent second opinion on architecture, ADR shaping, design tradeoffs, debugging hypotheses, incident reasoning, or a difficult technical decision. Routes high, xhigh, or ultra effort from task risk and breadth. Do not use for a verdict-bearing change review (use codex-review), routine implementation, or an unscoped “look at everything” request.

Run a fresh-context, read-only Codex GPT-6 Astra review in any project. Use for an independent review of uncommitted changes, a commit, a branch/base diff, an ADR, a plan, or explicitly named artifacts; use when the user asks for codex-review, an external Codex pass, or a verdict-bearing review. Routes high, xhigh, or ultra effort from task risk and breadth. Do not use for open-ended design or debugging counsel (use codex-advise) or for implementation.

Run every reader this plugin ships over a corpus you do not own — on a platform the maintainers cannot run — and report everything it printed, verbatim. Use when asked by another session to probe a repository, when preparing a release that changes what a reader says, or when a reader's answer about your own corpus looks wrong. Do not use to fix anything in the probed repository; this skill reads and reports only.

Implement a decided code change with the smallest coherent diff and fresh executed evidence. Use for bug fixes, bounded features, and small refactors when the request authorizes implementation. If a durable design decision is discovered, return it to the main coordinator rather than silently expanding scope. Do not use for review-only work.

Break a mechanism on purpose and check whether anything notices, to find tests and contracts that assert nothing. Use when the user asks whether the tests actually test anything, wants mutation testing, suspects a suite is green for the wrong reason, asks "are these tests real", or wants confidence before trusting a passing run. Do not use to judge a diff or to review code someone wrote — that is `review`; this skill never reads for quality, it only measures what a suite detects.

Operate the harness itself — what is actually installed, whether a home gate is a forwarder or a stale copy, how to verify an upgrade, and how to read a lint finding's severity. Use when a gate behaves unexpectedly, after `claude plugin update`, when a standalone copy may be shadowing the plugin, when deciding whether a finding blocks, or when a session is about to write operating instructions into a project's CLAUDE.md. Do not use for authoring or executing a record — that is `adr-write` and `adr-execute`.

Document a material, recurrent, or production-relevant failure—or a genuinely reusable lesson—as a structured postmortem. Use after resolving a qualifying failure or review finding, or whenever the user explicitly asks. Do not auto-trigger for every routine bug, expected test failure, or low-impact one-off fix.

Apply universal scope, simplicity, verification, and risk-routing discipline to substantive code changes. Use in the main coordinator for implementation planning and completion; do not use as a second lifecycle router or preload its full body into child agents.

Default code review skill, reading a diff and judging it. Use when the user asks for a review, PR review, audit of changes, a second opinion on implementation quality, or a test-quality review of tests as written. Loads a performance or security mode from references/ when the review centers on those. Do not use for implementation work, and do not use to measure what a suite actually detects — breaking a mechanism to see whether anything notices is `mutation-audit`.

Discover and write verifiable requirements before design or implementation. Use when the user invokes `/quality-harness:spec-write`, asks for a spec/use cases/BDD scenarios, or brings a feature idea too ambiguous to plan; use grill-only mode to audit an existing plan, ADR, or PR description without editing it. Do not use when behavior is already decided and the task is bounded implementation or ADR retirement.

Main-session coordinator for substantive development goals. Use when the user invokes /quality-harness:work, sets a development goal and wants the lifecycle handled end to end, or requests non-trivial implementation without naming a narrower skill. Classifies once, freezes scope, routes only the necessary stages, and owns final evidence. Never invoke inside a spawned agent or workflow role; do not use for conversational questions.