[NEW AGENT] claude-code-audit-stack — 3 adversarial verification subagents (bot-deploy-verifier, claim-auditor, remote-agent-dispatcher)

## Preliminary Checks - [x] Read Code of Conduct - [x] Reviewed existing subagents — this collection is not a duplicate (no equivalent post-deploy verifier, quantitative claim auditor, or PID-correct remote agent dispatcher in the current set) - [x] Proposal is for legitimate, constructive use cases only - [x] I have domain expertise (3+ years building & operating production trading bots and quant research pipelines, where these specific failure modes occur) ## Proposal scope This is a proposal for a **collection of three tightly-coupled subagents** that together form an adversarial-verification stack, plus one PostToolUse hook. They are submitted as one collection because they share a unifying design principle (assume the agent above lied unless every check independently passes) and because at least one — `claim-auditor` — is auto-fired by a hook in the collection. If maintainers prefer, this could be filed as 3 separate proposals; let me know which fits your queue better. The collection is already public, MIT-licensed, schema-aligned, and production-tested at https://github.com/LaterKidsXD/claude-code-audit-stack. ## Proposed Agent Names 1. `bot-deploy-verifier` 2. `claim-auditor` 3. `remote-agent-dispatcher` Plus a `PostToolUse` hook (`audit-on-report-write`) that auto-fires `claim-auditor` on every `*.report.md` write. ## Domain Expertise The collection specializes in **catching silent failures that other agents miss** — failures where the upstream agent reports `DONE` or `PASS` but the actual end-state of the system is wrong: 1. **`bot-deploy-verifier`** — systemd / journalctl / git-driven verification. Catches the silent-skip pattern where an agent edits a config file but never restarts the service (the file is correct on disk, the running process is on old in-memory config), plus accidental cascade restarts of `BindsTo=`/`Requires=` siblings. 2. **`claim-auditor`** — adversarial quantitative review of Markdown reports. Catches probability stacking (`1−(1−p)^N` vs `N×p`), conditional vs marginal pass-rate confusion, percentage vs percentage-points mixups, best-of-N selection bias, bootstrap-with-replacement implications, sample-size red flags. Read-only (`tools: Read`). 3. **`remote-agent-dispatcher`** — mechanical scp+spawn for autonomous Claude Code agents on remote hosts via SSH. Captures the actual `claude` binary PID via `pgrep` (not the bash-wrapper PID via `$!` — the common trap when wrapping over SSH). ## Unique Value Proposition Unlike the existing breadth-focused subagents in this collection (which excel at domain expertise: Go, security, frontend, DBA, etc.), this stack is **narrow and adversarial-by-design**. Each agent assumes the agent above it might have skipped a step, and re-checks from scratch via `systemctl`/`journalctl`/`git`/`grep`. They surface explicit `BLOCKED` (not `DONE`) when any gate fails — agents are trained to drive toward DONE, and that's exactly the failure mode you want a separate verification process to catch. Specifically, no agent in the current set covers: - Post-deploy state verification of running systemd services (the deployment-engineer/devops-engineer agents do deploy planning, not adversarial post-deploy verification) - Quantitative auditing of Markdown reports with structured P1/P2/P3 severity - Remote agent spawning with PID-correct wrapper-aware capture The "blocked, not done" distinction is the load-bearing design property — it allows these to compose with other agents without false positives propagating downstream. ## Primary Use Cases **`bot-deploy-verifier`:** 1. Verify any `systemctl restart` actually landed (config edit went through, service reloaded, running journal shows new values) 2. Catch accidental cascade restarts of unrelated services (e.g., `BindsTo=`/`Requires=` dependencies firing as side effects) 3. Auto-rollback to a known-good config when verification fails (when caller provides `backup_path`) 4. Refuse to declare a deploy DONE until all gates green; explicit BLOCKED otherwise **`claim-auditor`:** 1. Audit autonomous research output (eval reports, MC results, backtest summaries) before acting on probability/EV claims 2. Surface decision-shaping math errors that human reviewers glide past on the 14th report of the day 3. Run as a post-processing step on any agent-generated quantitative analysis 4. Return structured JSON for CI integration / pipe-into-other-agents **`remote-agent-dispatcher`:** 1. Delegate tasks that exceed Claude Code's interactive timeout to a remote VPS 2. Run long-running autonomous agents on a host different from the human 3. Capture PID correctly across SSH+sudo+nohup+bash wrapping (where `$!` returns the wrapper, not the binary) 4. Survive SSH disconnect via `nohup` + verified PID file ## Required Tools & Capabilities - `bot-deploy-verifier` — `tools: Bash` (executes systemctl, journalctl, git over SSH; reads journal output) - `claim-auditor` — `tools: Read` (read-only — never modifies the report it audits; this is a security property) - `remote-agent-dispatcher` — `tools: Bash, Read` (Bash for scp/ssh/install; Read for spec-header validation only) No external network calls beyond what `Bash` allows the user to invoke. No hardcoded credentials, paths, or hostnames (all user-facing examples use `<your-ssh-host>` placeholders). No filesystem mutations by `claim-auditor`. ## Example Interactions **Example 1 — bot-deploy-verifier catches a silent skip:** ``` User: "Deploy the new trail-stop config to trading-bot.service" Caller agent: "Edited config/live.yaml: trailing_stop_min_dist: 0. Restarted service. DONE." bot-deploy-verifier (invoked after restart): - service active: ✅ - config-drift gate: ❌ FAILED — journal shows trailing_stop_min_dist: 4 (old value) - expected keys verified: 0/4 - untouched siblings: ✅ Returns: BLOCKED, NO ROLLBACK Cause: service was restarted at 14:42 UTC but the journal shows it loaded the OLD config from before the edit (the agent edited the wrong file or the systemd unit references a different path). ``` **Example 2 — claim-auditor catches probability stacking:** ``` User: "Audit eval_results.md before I commit to the strategy" claim-auditor (severity_floor: P2): # Claim Audit — eval_results.md ## P1 (decision-shaping, must address) | Quote | Why wrong | Correct number | |---|---|---| | "3 evals at 35.7% per eval = ~90% chance one passes" | P(≥1 of N) = 1−(1−p)^N, not N×p | 73.4%, not 90% | ## Summary - 1 P1 error / 0 P2 / 0 P3 - Recommended action: Recompute eval-cost economics before committing capital ``` **Example 3 — remote-agent-dispatcher captures correct PID:** ``` User: "Dispatch HANDOFF_LONG_BACKTEST.md to my-vps" remote-agent-dispatcher: - spec validated locally (HANDOFF header present) - scp + install on host: ✅ - spawn pattern executed (nohup + sudo -u daemon + sleep 18s) - pgrep -f basename: 1 claude binary PID found, PID 18432 - PID file written: /opt/agent-work/HANDOFF_LONG_BACKTEST_agent.pid - done-conditions verified (file exists, PID owned by daemon, alive ≥5s) Returns: DISPATCHED as PID 18432 on my-vps Caller schedules first heartbeat check in 15-30 min. ``` ## Your Expertise - 3+ years building and operating production trading bots (futures, crypto, prediction markets) — the failure modes these agents catch are drawn directly from real production incidents - Documented $106K backtest swing caused by a silent-skipped config restart (the originating incident for `bot-deploy-verifier`): https://github.com/LaterKidsXD/claude-code-audit-stack/blob/main/docs/silent-skip-incident.md - Real audit findings (redacted) showing `claim-auditor` catching probability-stacking errors in autonomous research output: https://github.com/LaterKidsXD/claude-code-audit-stack/blob/main/reports/sample-findings.md - All three subagents are in active production use against the live systems they were built for ## Additional Context - **Repository:** https://github.com/LaterKidsXD/claude-code-audit-stack (MIT, public) - **Already-shipped GitHub Action wrapping `claim-auditor`:** https://github.com/marketplace/actions/claim-auditor - **Anthropic Claude Code plugin marketplace:** submitted 2026-05-03, currently under review If maintainers green-light this proposal, I'm happy to submit it as a polished PR following the `plugins/<plugin-name>/` layout used by recent community plugins (e.g., `signed-audit-trails` PR #496, `protect-mcp` PR #503). All schema requirements (subagent frontmatter, tool restrictions, MIT license, no hardcoded paths) are already met in the upstream repo.