Menu
General-purpose harness for scheduled repo work โ stage in an isolated worktree, verify with a genuinely separate second session, ship only after a pass. Read-only โ reports, never delivers.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
| name | loop-harness-verification |
| description | General-purpose harness for scheduled repo work โ stage in an isolated worktree, verify with a genuinely separate second session, ship only after a pass. Read-only โ reports, never delivers. |
The prompt carries intent; this skill carries the procedure โ keep it a checklist, not an essay. Adapted from Forward Future Loop Library #17 (Istasha) to this library's L1-first, report-only shape.
loop-harness-verification-state.md's last run.loop-verifier agent, or a second claude -p / opencode run call โ
never the same pass that staged the output) to check the candidate
against explicit criteria.harness-report.md recommending delivery. Do NOT
deliver it yourself.loop-harness-verification-state.md
and retry, but only within the configured retry limit.harness-report.md: source revision, staged output, verifier
result, delivery status, next scheduled run โ the source's own required
fields.loop-harness-verification-state.md, append one
line to loop-run-log.md, commit both.
Based on SOC occupation classification
Reviews pending PRs, excludes stale/unfinished work, combines the rest into a release batch built from main. Read-only โ reports the batch, never actually releases.
Writes a scoped spec packet for one ticket, reviewed by a genuinely independent adversarial pass that also regression-checks prior findings. Read-only outside the draft โ never writes the canonical spec.
Defines every required outcome and its evidence before acting on long-running work, then audits each requirement as proved/weak/missing/ contradicted. Only run when explicitly asked for a completion audit.
Runs an adversarial review cycle against a PR (Claude implements/fixes, Codex or a subagent stand-in reviews), up to 5 iterations. Read-only on the real PR โ drafts fixes, never pushes them.
Turns a ticket or bug report into a review-ready patch. Read-only โ reproduces, roots out the cause, drafts a fix, reports it. Never opens a PR, never folds in unrelated refactors.
Inspects branches, PRs, commits, and worktrees for what's stale vs. valuable. Read-only. Never deletes uncertain work or someone else's PR without confirmation.