verify

Turn "it looks done" into a verdict backed by executed evidence. All paths below are relative to this skill's directory.

Three laws

1. No verdict without an executed command. Cite the command, exit code, and a line of output. A gate that could not run is not run, never passing.
2. No finding without a refutation attempt. Every model-authored candidate faces an independent skeptic. Survivors are reported; the rest are counted.
3. No repair that only silences the checker. The richer tiers refuse skipped or weakened tests, suppressions, widened types, swallowed errors, edited CI, and rewritten plans. scripts/forbidden-repairs.mjs enforces this.

Modes and tiers

Syntax: Codex uses $verify; the Claude plugin uses /maxgfr:verify; a standalone Claude skill uses /verify.

The default answers only whether the repository's commands pass now. It does not read the diff, check a plan, hunt defects, prove behaviour, or repair a failure. Use verify light when the change itself needs analysis and repair. Presets and full configuration: references/config.md.

Phase 0 — Resolve, then route

Do this in the main context before spending an agent.

1. Config first. Run node scripts/resolve-config.mjs --cwd <repo> --host <host> -- <arguments>. Pass its tier, mode, ref, and config unchanged. A tier-like branch needs --ref light; crosscheck is a modifier.
2. Gates. Run node scripts/detect-gates.mjs --cwd <repo> --pretty.
3. Run directory. Create <report.dir>/<YYYYMMDD-HHMMSS>/ and prune oldest runs beyond keep_runs.

The short route applies when the resolved config has only gates enabled and the loop disabled: spec: false, defects: false, behavior: "off", peer: false, loop.enabled: false.

• Run the gates lane once, at low effort. With no detected gate, spend no agent and return UNPROVEN; never escalate tiers.
• Skip delta, promise, ref, baseline, matrix, judging, reporter, and fix-loop work. A blocking gate that fails, times out, or cannot run returns FAIL after that pass. Report a non-blocking gate failure without sinking the verdict.
• Write the compact report from the returned structured gate evidence in the main context. This keeps the report path real without spending a reporter agent.

Any explicit richer tier, enabled analysis lane, peer crosscheck, or enabled loop takes the full route below.

Full-route pinning

1. Delta. User ref wins. Otherwise: no commits → whole working tree; dirty tree → git diff HEAD, cached diff, and commits since merge-base; clean tree → diff from git merge-base --fork-point origin/<default> HEAD; non-git → target directory. Always include ?? paths from git status --porcelain as whole-file additions; never mutate the index with git add -N.
2. Promise. User path → active host plan artifact → docs/plans/, docs/superpowers/plans/, specs/, .scratch/ → referenced issue → inferred intent, named as such. An approved docs/plans/ file is announced before agent work.
3. Baseline, only when repair is enabled. git stash create; empty output means use HEAD. It touches no ref or file.
4. Host, only for peer crosscheck. Pass the executing host, claude or codex; do not infer it from installed commands.

Stop if an explicit ref does not resolve or the full-route diff is empty.

Full pipeline

Read references/lanes.md for lane briefs.

Lane E is opt-in through crosscheck; its findings face the same skeptics. Read references/crosscheck.md for its brief and schema. An unavailable requested peer is named in RESIDUAL RISK.

Every stage inherits the session model unless config pins it. Prefer the highest host capability available:

1. Workflow. Call Workflow with workflows/verify.mjs and resolved Phase 0 data.
2. Parallel subagents. Follow references/fallbacks.md and its deterministic schedule.
3. Inline. Run the same enabled phases sequentially and report execution: inline.

Output contract

Return only: verdict and tier; for every tier below deep, a second line naming what was not checked; EVIDENCE rows with command, exit code, and first failing line; surviving findings with failure scenarios; REQUIREMENTS counts when the spec lane ran; RESIDUAL RISK; and a real report path. Examples and red flags: references/report.md.

Anything unverified or errored is named in RESIDUAL RISK. Findings killed by skeptics are a count, not a list. When report_path is null, write the short report yourself in the run directory before printing its path.