菜单
Send independent adversaries on a bug hunt — a structural pass then a logic-tracing pass, from different model families, iterated to convergence (fix, re-hunt, repeat until zero new bugs). Falls back to a Sonnet adversary when external models are unavailable or contaminated.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
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 | bug-hunt |
| description | Send independent adversaries on a bug hunt — a structural pass then a logic-tracing pass, from different model families, iterated to convergence (fix, re-hunt, repeat until zero new bugs). Falls back to a Sonnet adversary when external models are unavailable or contaminated. |
| argument-hint | ["spec-file or scope description"] |
| allowed-tools | Read, Edit, Write, Bash, Grep, Glob, Agent, AskUserQuestion |
Adversarial review by two independent adversaries, iterated to convergence. Independent Verification & Validation — the adversary has no stake in the implementation.
The structural adversary catches structural issues (missing baselines, overclaims, architectural gaps). The logic adversary catches logic bugs (inverted boolean conditions, decision tree errors, spec-to-code mismatches). Run the structural adversary to convergence first, then the logic adversary on its approved result. If the logic adversary finds bugs, fix and re-run the structural adversary to verify the fix didn't introduce regressions, then the logic adversary again. Converged = both approve with zero new findings.
Adversary selection. The adversary's value is independence: prefer different model families for the two passes (different priors catch different bugs; same-family review shares blind spots). For contamination-controlled work, the adversary must also be cutoff-clean relative to the system under review — a reviewer that post-dates the code could "know" the intended fix and leak it into the verdict.
Fallback to Sonnet. When external adversary CLIs are unavailable, OR when contamination control rules them out (frontier external models often post-date the system under review), fall back to a Sonnet adversary (Claude alone, pinned to a clean-cutoff Sonnet). Same-family review is weaker on diversity — flag that in the verdict — but it preserves cleanliness and availability. Two clean adversaries from different families is ideal; one clean adversary beats two contaminated ones.
One of:
docs/agent-participation.md) — the adversary reviews implementation against itgit diff main...HEAD)Read the actual diff under review (git diff main...HEAD, or the specific fix being validated), not just the file names. The changed lines are the center; the blast radius is that code plus what it touches: callers of changed functions, code that reads state the diff now writes differently, and the control-flow paths the change newly reaches or newly skips. Scope the hunt to that radius.
Two in-scope finding types when validating a fix:
Out of scope: pre-existing bugs unrelated to the diff. Finding them neither validates nor invalidates the fix; note them separately if worth it, but don't chase them — that's scope creep that starves the verdict on the change.
(Two modes: fix-validation — a diff is under review, scope = blast radius, as above. full-surface — a spec or whole module is under review, scope = all of it. If a spec exists or no diff is given, you're in full-surface mode; otherwise fix-validation. The convergence note about "leaving scope" below applies only to full-surface.)
Send the adversary on a bug hunt via stdin pipe. External adversary CLI:
cat <<'PROMPT_EOF' | <adversary-cli> -
<prompt here>
PROMPT_EOF
Sonnet-adversary fallback (external unavailable or contaminated): dispatch a subagent pinned to a clean-cutoff Sonnet (its own agent type), given the same prompt and the system-under-test access (e.g. via the project's shell helper), web tools disabled.
The prompt must include:
docs/bug-hunt.md or append a new section)mix test / npm test / equivalent and report failuresRead the findings. For each bug:
Ask the user before deferring anything severity:high.
Fix all accepted bugs. Run tests. Commit.
Send the adversary again with an updated prompt that lists ALL previously found bugs (fixed and deferred) and instructs:
Loop steps 3–5 until the adversary reports zero new bugs. That's convergence.
Halt condition: if the hunt hasn't converged after the fix-count stops decreasing for 3 consecutive rounds at the same count, halt and report to the user. The implementation may have structural issues the spec didn't anticipate.
Round 1:
Bug hunt. Read [spec file] for the contract. Read [file list] for the implementation.
Find bugs: logic errors, spec violations, integration seams, security issues, edge cases, resource cleanup gaps, and common mode assumptions.
For each bug: describe it precisely, cite file and function, explain the impact, suggest a fix (don't fix it).
Write findings to [output file]. Run [test command] and report failures.
If zero bugs found, write: "# Bug Hunt Round 1\n\nZero new bugs found. Hunt converged."
Round N:
Bug hunt round N. Previous rounds found and fixed M bugs. Find NEW bugs only.
Previously found (do NOT re-report): [list all bugs from all rounds]
Intentionally deferred (do NOT report): [list deferred items]
If zero new bugs: write "# Bug Hunt Round N\n\nZero new bugs found. Hunt converged." to [output file]
/codex is a single-pass review that reports feedback for the user to act on. /bug-hunt is a multi-pass adversarial loop that fixes bugs and re-hunts until convergence. The reviewer is independent (different model), the loop is automated, and the stopping criterion is empirical (zero new findings), not a fixed number of passes.