Menu
Adversarial code review on a single fix branch — codex (structural) and gemini (logic) volley until convergence or budget. The harness picks the entry, runs the test attestation, and records the verdict; the skill only does the LLM-shaped review.
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.
Based on SOC occupation classification
Filter recent Hacker News for live threads on June's topics and hand him the list. Separately, flag the rare subset where work he has already done settles a specific quoted claim, and draft that comment as a plain argument with no link to his own writing. Use when June asks to check HN or find threads worth reading or commenting on.
Scan blog posts for AI writing tics and missed opportunities for human voice. Reports subtractions (AI patterns to remove) and additions (wordplay, arc, and claims to strengthen).
Poll June's Hacker News feeds, report what is new since the last run, and idempotently arm a background watch for replies to his comments. State persists in ~/Documents/sweep/ so it survives across sessions. Use when June asks what is new on HN, whether anyone replied, or to start watching a thread.
Argument-level compression by a strengthened absence test (does the argument stand with the same force, clarity, evidence, scope, pacing, and misreading-resistance?). Removes passages that fail it (self-referential cruft, cross-section restatement, self-recap, re-derivation, scaffolding, expired reader-orientation, stakes inflation, inert hedges, redundant examples); grafts good rhetoric onto an adjacent essential point rather than deleting. Idempotent. The structural cousin of /tighten (word-level) and /not-but (negation).
Audit a technical candidate's resume/CV by verifying its claims against the public contribution graph instead of taking the text at face value. Sorts every claim onto the self/peer/world attestation lattice, applies a cost-and-stake filter so cheap proxies (stars, downloads, backlinks) don't count, runs LIVE checks (GitHub, arXiv, package registries, CVEs), and reports two separate axes — merit (staked acceptance of the candidate's specific work) and attention (timestamped output) — plus a verifiability score in [0,1]. For technical recruiters and hiring managers. Invoke when given a candidate — a GitHub handle, a resume/CV, an arXiv author, a personal site — and asked to verify, audit, vet, or check their claims. Runs at two depths — a quick triage scan or a deep per-claim audit — and elicits which. Checks the provenance of a personal site or self-hosted resume (Internet Archive + git history) as an un-backdatable anti-tailoring anchor. Needs tool access (Bash/curl or WebFetch).
Find and defuse "not X but Y" / "isn't X; it's Y" / "X — not Y" constructions. Deterministic grep for detection (no LLM blind spot), the not-but subagent for triage. Split-or-cut tail negations ("X, not Y") aggressively — promote the point to its own sentence or delete it; recast in place only mid-clause "not X but Y" that genuinely can't be split; keep just the rare iconic thesis line literal.
| name | qa |
| description | Adversarial code review on a single fix branch — codex (structural) and gemini (logic) volley until convergence or budget. The harness picks the entry, runs the test attestation, and records the verdict; the skill only does the LLM-shaped review. |
| argument-hint | <repo> [--entry BRANCH] [--budget SECONDS] [--dry-run] |
| allowed-tools | Read, Write, Edit, Bash, Glob, Skill |
You were invoked by the harness with a prompt describing what to review and a fix branch checked out in a worktree. Run codex (structural) and gemini (logic) as adversarial reviewers on the diff. Report verdicts. Stop.
The harness owns everything that isn't LLM-shaped: queue state, schema, entry selection, idempotency, budget timing, status appends, gate validation, resume after timeout. Don't reimplement it in this skill. The CLI is the contract.
| What you need | How to get it |
|---|---|
| Diff and branch context | The harness has already checked out the branch in $worktree. git diff against base. |
| Test attestation (fail-on-master / pass-on-fix) | sweep qa test --repo … --branch … --worktree … --test-cmd … |
| Run one codex round | sweep qa codex --repo … --branch … --worktree … |
| Run one gemini round | sweep qa gemini --repo … --branch … --worktree … --round N |
| End-to-end on one entry (no Temporal) | sweep qa full --repo … --branch … --worktree … --test-cmd … |
| Record the verdict | The composer (qa_one_entry / sweep qa full) writes the gate line. Don't append jsonl by hand. |
If a CLI call rejects your input, read the error and fix the input — don't paper over it in the skill prompt. That's the skills-lack-determinism loop: deterministic guardrails return useful errors, the agent closes the loop. If the error reveals a missing CLI affordance, add it to sweep qa rather than encoding the workaround here.
What the skill actually contributes — the part the CLI can't do — is framing the two reviewers and reading their verdicts.
Codex (structural):
Review this fix as a maintainer of
<repo>seeing it for the first time. Find structural problems: missing edge cases, wrong abstractions, scope creep, hidden coupling, missed invariants. Don't comment on style. Verdict:pass/reject/revise. Ifrevise, list specific changes.
Gemini (logic):
Trace the logic in this diff. Find: inverted boolean conditions, off-by-one errors, missed branches, race conditions, type confusions. The fix must be correct, not just plausible. Verdict:
pass/reject/revise.
Round loop, per reviewer:
pass → done.revise → apply codex's revise suggestions (auto), surface gemini's revise notes (don't auto-apply — gemini is better at logic traces than rewrites). Re-run test attestation. Re-send.reject → stop the loop for that reviewer, continue the other.Convergence: stop on pass, on round 3, or when the harness signals budget. The composer records the final verdict; the skill doesn't write jsonl.
Both reviewers see the same diff. Don't shape the prompt to one model's strengths — the point of two reviewers is independent blind spots.
Never edit the diff to satisfy a reviewer if it changes the issue's intent. A reviewer asking for a change the issue's commenter didn't ask for is scope creep; surface it to the human, don't fold it in.
If sweep qa codex returns a credit/quota error, fall back to an Opus subagent playing the structural-reviewer role with the same framing. The composer records codex_provenance: "opus-fallback" so the retro can separate real-codex outcomes. Codex credits restore 2026-05-17; remove this path after retro confirms parity.
If gemini is stubbed (no subscription), the composer records gemini_first/last = "stubbed". The downstream hook will block the PR — that's the contract, not a bug to work around.
sweep does — the skill is handed one.)/drip does, downstream.)/drip creates PRs.)The CLI is the interface. Any process that calls sweep qa full (or composes sweep qa test + the codex/gemini steps) substitutes for this skill. Manual reviewers, external CI signals, alternative reviewer pairs all plug into the same composer.
After all narration, the last printed line must be a single JSON object matching this schema:
{
"verdict": "pass | fail | needs_human",
"issues_found": [],
"reason": "short string, ≤200 chars",
"rejected": false,
"reject_reason": null
}
verdict=pass means both adversaries converged on "ship it" and the test attestation is intact. fail means at least one substantive bug found that blocks. needs_human means non-substantive concerns that need operator judgment (maintainer-style mismatch, ambiguous bug). issues_found is an array of short strings; empty when verdict=pass.
Set rejected=true with a reject_reason when QA cannot run (missing worktree, unrunnable tests, broken branch — distinct from "tests failed," which is verdict=fail). Rejected jobs route to ~/.sweep/inbox/rejected.jsonl for operator review.
Print the JSON as the literal last line, no trailing prose, no markdown fence.