git:20260905.49c56b2 to git:20260905.6c19234

2 added, 2 removed. Audit A to A.

---

name: repo-harness-check

description: Verification entrypoint for repo-harness workflow readiness. Runs workflow gates, task sync, contract checks, inspector, and migration dry-run before merge or release.

when_to_use: "repo-harness-check, check agentic workflow, verify harness, pre-merge workflow check, release readiness, validate tasks-first contract"

---

# repo-harness-check

Use this command when the user asks whether the harness, migration, or release surface is ready.

## Protocol

1. Confirm the repo path and report dirty-worktree boundaries.

- 2. Read the canonical verification policy from the target repo's root agent context (`CLAUDE.md` for Claude, `AGENTS.md` for Codex) `## Required Checks` section. Select its applicable tier from the complete diff and risk, state the reason, and run that tier's commands; a command listed under a conditional full-suite tier is not mandatory for every task. Preserve stronger active-contract and CI requirements. Docs-only/ledger closeout does not alone trigger full tests; unknown impact, high-risk, cross-module, and release changes require the full tier. Use the global/package helper runtime where applicable; this self-host repo may use root `scripts/`. If the section is missing or empty, report the missing policy rather than inventing commands. Report exact results and unrun checks, and never claim full-suite validation from focused tests.

+ 2. Read the canonical required-check list from the target repo's root agent context (`CLAUDE.md` for Claude, `AGENTS.md` for Codex) `## Required Checks` section, then apply its scope conditions and run the applicable commands through the global/package helper runtime. This self-host source repo may also use root `scripts/` for source-only maintenance commands. Use focused checks by default; a conditional full-suite command is not an unconditional requirement. Consume current canonical acceptance evidence for already-satisfied expensive criteria. After a bounded follow-up edit, have the parent retain the baseline full-run identity and revise final criteria to focused delta checks when no explicit full-suite requirement or uncovered integration risk remains; never relabel a baseline pass as current-subject full evidence. If `## Required Checks` is missing or empty, report that as the first blocking finding instead of substituting a default list.

3. Run advisory readiness when available:

- `repo-harness run check-agent-tooling --host both --json`

4. Treat missing CodeGraph or missing Codex `health`/`check`/`mermaid` as hard failures.

5. Treat Waza staging drift as a yellow readiness flag; report the fix or acceptance reason without failing the repo gate.

6. Report skill eval authority when release/readiness evidence depends on skill

effectiveness:

- authoritative: non-dry-run `bun run benchmark:skills --eval <slug>` with

`full_test_count > 0`, `dry_run_ratio <= 30%`, and graders reported

- non-authoritative: dry-run-heavy or all-dry-run evidence

- unavailable: no current eval evidence; report the benchmark command needed

7. Summarize pass/fail evidence, yellow flags, eval authority metrics, and the next blocking command if any.

## Delegation Brief Evidence

- A file-coupled `contract-run` worker is instructed to self-run every `exit_criteria` command and paste the exact command line plus output before reporting done. When checking such a run, confirm that evidence was actually pasted, not merely asserted, and re-run any criterion you cannot confirm from the transcript.

+ A file-coupled `contract-run` worker prepares executable evidence through `verify-sprint --prepare-acceptance`. Inspect the immutable run artifact and its current subject, contract, target revision, toolchain context, and per-criterion results. Valid executed and exact-context reused passes count. Return missing, stale, or failed criteria to the canonical runner; do not independently launch the same suite or accept transcript assertions as evidence.

## Root Cause Evidence Review

For a `bugfix` contract, confirm `## Root Cause Evidence` states a testable `root_cause`, a working `repro`, a `regression_guard` that also appears under `exit_criteria.tests_pass`, and a `pre_fix_failure_artifact` showing a non-zero `PRE_FIX_EXIT=` line for that guard, not a passing run. Also check whether `task_profile` was mislabeled or left out entirely: an omitted `task_profile` defaults to legacy pass-through (non-bugfix) by design, so confirm that default is actually correct here rather than an evasion of the gate.

## Failure Modes

- If any required workflow gate fails, report the first blocking command and stop the readiness claim.

- If advisory tooling times out, report advisory evidence as unavailable instead of passing it.

- If eval evidence is missing or all dry-run, mark it non-authoritative or unavailable for skill effectiveness and name the repair command.

## Boundaries

- Does not mutate repo files by default.

- Does not silently ignore CodeGraph readiness failures, advisory tooling hangs, or skipped checks.

- Does not claim skill-effectiveness authority from dry-run benchmark output.

- Does not claim release readiness if source repo and installed runtime copy are out of sync.

- Does not maintain its own copy of the required-check list; the target repo's root `## Required Checks` section is the single source of truth.