{
  "markdown": "# claude-code-audit-stack\n\nAdversarial verification primitives for [Claude Code](https://claude.com/claude-code). Three production-tested subagents that catch the silent failures Claude Code's defaults miss.\n\n```\n$ /plugin install claude-code-audit-stack\n```\n\n> *Pending Anthropic Claude Code plugin marketplace approval (submitted 2026-05-03). Until approved, use [manual install](#install-manually) below.*\n\n## What's in the box\n\n| Subagent | What it catches | When to invoke |\n|---|---|---|\n| **`bot-deploy-verifier`** | Silent skipped restarts. Sibling services bumped by accidental cascade. Service running an old in-memory config because the agent edited the file but forgot to restart. | IMMEDIATELY after any `systemctl restart` or service redeploy |\n| **`claim-auditor`** | Probability stacking (`1−(1−p)^N` vs `N×p`). Conditional vs marginal pass-rate confusion. Percentage vs percentage-points mixups. Bootstrap-with-replacement implications. Best-of-N selection bias. | PROACTIVELY whenever a Markdown report contains probability/EV/pass-rate/MC claims |\n| **`remote-agent-dispatcher`** | Lost-PID daemon spawns (`$!` captures the bash wrapper, not the claude binary). Orphaned long-running agents. SSH-disconnect kills. | When delegating a >15-min task to a remote host |\n\nPlus a **`audit-on-report-write`** PostToolUse hook that auto-fires `claim-auditor` on every `*.report.md` write/edit. Free quality enforcement on every report your agents generate.\n\n## Why this exists — the silent-skip incident\n\nA Claude Code agent was asked to fix a trailing-stop config bug. It read the YAML, edited the value, and reported \"DONE.\" The on-disk file was correct. The running service had **never been restarted** to pick up the change.\n\nThree trading days passed before anyone noticed. A 3-year backtest replay quantified the cost: a **$106K swing** on a 553-day window, almost entirely because the service was running on the wrong config.\n\nThe agent didn't lie. It just didn't verify. `bot-deploy-verifier` exists to be the layer that does verify — adversarially, on every deploy.\n\n[Full incident write-up →](docs/silent-skip-incident.md) · [Sample audit findings (real, redacted) →](reports/sample-findings.md)\n\n## Quick start\n\n### Install via Claude marketplace (recommended once approved)\n\n```bash\n/plugin install claude-code-audit-stack\n```\n\nThis installs all 3 subagents and the hook. Subagents auto-route via their `description` triggers; the hook fires on `Write|Edit` of `*.report.md`.\n\n> **Note:** This command requires the plugin to be live in Anthropic's official Claude Code plugin marketplace. Submitted 2026-05-03, currently under review. Until approved, use the manual install below.\n\n### Install manually\n\n```bash\ngit clone https://github.com/LaterKidsXD/claude-code-audit-stack.git ~/audit-stack\ncp ~/audit-stack/agents/*.md ~/.claude/agents/\n# (optional) copy the hook into your project's .claude/settings.local.json\ncat ~/audit-stack/hooks/hooks.json  # paste the .hooks.PostToolUse block\n```\n\n### Run it in CI as a GitHub Action\n\nTo run `claim-auditor` on every PR — not just locally inside Claude Code — use the [`claude-code-audit-action`](https://github.com/LaterKidsXD/claude-code-audit-action) wrapper. It audits `*.report.md` files in PR diffs, posts a structured findings comment, and fails the check on any P1.\n\n```yaml\n- uses: LaterKidsXD/claude-code-audit-action@v1\n  with:\n    api_key: ${{ secrets.ANTHROPIC_API_KEY }}\n```\n\nThe Action vendors the same `claim-auditor` system prompt this stack ships, so local audits and CI audits stay aligned.\n\n## Examples\n\n- **`bot-deploy-verifier`:** [verify a config swap](examples/bot-deploy-verifier-invocation.md) — shows a real PASS report and a FAILED-WITH-ROLLBACK report\n- **`claim-auditor`:** [audit a Monte Carlo report](examples/claim-auditor-invocation.md) — shows the structured P1/P2/P3 severity table on a buggy report\n- **`remote-agent-dispatcher`:** [dispatch a 3-year backtest to VPS](examples/remote-agent-dispatcher-invocation.md) — shows the PID-capture trap and how the dispatcher avoids it\n\n## How this is positioned\n\nThis stack is **opinionated**. It is NOT trying to be `wshobson/agents` (185 subagents across 25 categories) or `affaan-m/everything-claude-code` (48 subagents + 182 skills). Those collections optimize for breadth.\n\nThis stack optimizes for one thing: **catching the silent failures that bring down Claude Code-driven systems in production.** Each subagent is adversarial by design, ships with a documented failure mode it catches, and refuses to PASS until every check is verified.\n\nIf you want a frontend-developer agent or a SQL query optimizer, install one of the larger collections. If you want layers of verification that assume the agent above lied to you, install this one.\n\n## Verticals\n\nThe general-purpose subagents above (`bot-deploy-verifier`, `claim-auditor`, `remote-agent-dispatcher`) work with any system using `systemd`, `journalctl`, and SSH. The author's home turf is **algorithmic trading bots and quant research pipelines**, and the failure modes catalogued in the subagents (broker-side fills not detected by engine, exit-chain priority bugs, off-spec trail config, claim-auditor catching probability errors in eval reports) come directly from that vertical.\n\nA trading-specific `trading-bot-health` subagent is planned for v0.2 — see the [roadmap](docs/roadmap.md).\n\n## What this stack is NOT\n\n- ❌ Not a general-purpose code review system. Use Anthropic's `/security-review` for security review and `wshobson/agents` for breadth.\n- ❌ Not a Claude Code feature replacement. These subagents complement Claude Code's defaults — they don't replace any.\n- ❌ Not a prompt-versioning or LLM-Ops platform.\n- ❌ Not \"the only auditor you need.\" A real production stack has multiple auditors at different layers; these three are the ones that catch the silent failures most other agents miss.\n\n## Compatibility\n\n- **Claude Code:** any version supporting subagents (Skills + Plugins shipped Oct 2025; Marketplace launched March 2026)\n- **Models:** `bot-deploy-verifier` and `remote-agent-dispatcher` use `sonnet` (cheap, mechanical). `claim-auditor` uses `opus` (heavier reasoning required for the math).\n- **OS:** anywhere `bash`, `systemctl`, `journalctl`, `ssh`, and `jq` work. Tested on Ubuntu 22.04 / 24.04. Should work on any systemd-based Linux. Non-systemd hosts (Docker Compose, supervisord, pm2) are TODO — see [roadmap](docs/roadmap.md).\n\n## License\n\nMIT. Take it, fork it, repackage it.\n\n## Author\n\nBuilt and tested in production on a multi-bot trading infrastructure stack — see the case study in [docs/silent-skip-incident.md](docs/silent-skip-incident.md). Author at [LaterKidsXD](https://github.com/LaterKidsXD).\n\nIf this stack catches something for you, please open an issue or PR. The standard for new subagents is documented in [docs/roadmap.md](docs/roadmap.md).\n",
  "bytes": 6908,
  "sha": "feb803e3b2794f336eda27d1812958e7d685de5e55a83a465e88ccf96a80b488",
  "repo_slug": "laterkidsxd/claude-code-audit-stack",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/plg_laterkidsxd_claude_code_audit_stack_clau_e1d38c1b/readme"
}