菜单
Adversarial multi-agent bug-hunt over a diff or subsystem. Fans out independent reviewers briefed with VPNRouter invariants, verifies each claim against code, and appends survivors to plans/OPEN-DEFECTS.md (the cut-stable gate ledger). Run before a stable cut and after any non-trivial feature.
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
| name | bug-hunt |
| description | Adversarial multi-agent bug-hunt over a diff or subsystem. Fans out independent reviewers briefed with VPNRouter invariants, verifies each claim against code, and appends survivors to plans/OPEN-DEFECTS.md (the cut-stable gate ledger). Run before a stable cut and after any non-trivial feature. |
| when | Before cutting stable, after landing a non-trivial feature/refactor, or when the user asks to "bug-hunt" / "adversarially review" the current work. NOT for ≤5-line hotfixes (the ship review-diff covers those). |
The methodology's review layer (METHODOLOGY §5/§7.1), captured as a repeatable skill so it stops being improvised every session and stops being a post-ship post-mortem. It feeds the open-defect ledger the cut-stable gate reads, closing the loop the auto-failover P0 fell through (it was found, deferred, shipped).
git diff <base>..HEAD) OR a named subsystem (e.g. "the
failover path", "the update extractor", "the Linux kill-switch").docs/REVIEW_AGENT_PROMPT.md (paste verbatim into each agent).Fan out independent reviewers (3-5 for a feature; more for a "thorough
audit"). Each is INDEPENDENT — paste the full diff/subsystem + the
docs/REVIEW_AGENT_PROMPT.md block; never "as discussed". Use the Workflow
tool (one agent per lens/area, then synthesize) when the user has opted into
multi-agent orchestration, else spawn parallel Agent(subagent_type: "general-purpose", ...) calls. Give each a distinct angle so they don't all
find the same thing: e.g. correctness, concurrency/lifetime, security/secrets,
fail-closed/leak, test-coverage.
Verify every claim against the code before recording it. A finding that
isn't grounded in a real file:line (Read/Grep it) is dropped. This is the
difference between a bug-hunt and a hallucination — the synthesizer re-opens
each claimed line and confirms the defect is real.
Triage survivors into severity (P0 = a broken/leaky build can reach users; P1 = real, bounded; P2 = hygiene), each with the concrete fix + blast-radius.
Land them:
plans/OPEN-DEFECTS.md
## Open as - [ ] **P0/P1** — <symptom> — <file:line> — <target>.
The cut-stable gate (tools/check-open-p0.ps1) then physically blocks a
stable cut while it's open (unless explicitly -Waived).severity (P0|P1|P2) · title · problem · evidence (file:line) · fix ·
cost (S|M|L) · risk. Synthesize duplicates raised by multiple reviewers into
one entry, noting which reviewers.
plans/OPEN-DEFECTS.md or the cut gate can't see them.docs/REVIEW_AGENT_PROMPT.md — the per-agent brief + VPNRouter invariants.plans/OPEN-DEFECTS.md — the ledger; tools/check-open-p0.ps1 — the gate.plans/methodology-audit-2026-06-25.md — the 7-lens audit that established this.After every successful ship of a -rN candidate, run this skill to actually verify the new binary works — on the windows-brat TEST VM (192.168.0.106) over WinRM, NEVER the local dev box (see the STOP banner in the body). Downloads the freshly-shipped ZIP, deploys it onto brat, launches + drives it there via UIA over WinRM, screenshots each verification point, tails `vpnrouter*.log` for errors, and produces a PASS/FAIL report. Auto-triggered after `ship-rolling-candidate` completes. DO NOT skip even for "tiny" changes — past sessions shipped 12 candidates without local verification and the user caught it via screenshots. This skill makes "I forgot to test" impossible.
After every successful ship of a -rN candidate, run this skill to actually verify the new binary works — on the windows-brat TEST VM (192.168.0.106) over WinRM, NEVER the local dev box (see the STOP banner in the body). Downloads the freshly-shipped ZIP, deploys it onto brat, launches + drives it there via UIA over WinRM, screenshots each verification point, tails `vpnrouter*.log` for errors, and produces a PASS/FAIL report. Auto-triggered after `ship-rolling-candidate` completes. DO NOT skip even for "tiny" changes — past sessions shipped 12 candidates without local verification and the user caught it via screenshots. This skill makes "I forgot to test" impossible.
Ship a rolling release candidate vX.Y.Z-rN. Bumps AppVersion, builds Windows + triggers Mac/Linux CI, applies notes, marks prerelease, deletes previous candidate per rolling-rN policy.
Ship a rolling release candidate vX.Y.Z-rN. Bumps AppVersion, builds Windows + triggers Mac/Linux CI, applies notes, marks prerelease, deletes previous candidate per rolling-rN policy.
Promote a rolling -rN candidate to stable vX.Y.Z (no suffix). Bumps AppVersion, creates fresh tag without suffix, full rebuild + Mac/Linux CI, restores Latest, deletes -rN.
Diagnose VPNRouter config / connectivity issues by reading the user's config.yaml + current.json + recent vpnrouter*.log. Trace ConfigMode → Vless.Servers → outbounds → flow. Catches silent leaks, ConfigMode mismatches, missing proxy outbounds.