Command
Hostile audit of the project's agent enforcement surface: assume every constraint is bypassable until the enforcement path is traced end-to-end and proven to bind, then close each proven loophole and re-run the bypass to prove it blocks.
Command
Audits the project's .claude/rules/ configuration end-to-end: validates every rule file, classifies every rule in CLAUDE.md as correctly placed or misplaced, and generates concrete new-rule suggestions from audit artifacts and project conventions. Assume every rule file has defects until proven otherwise.
Command
Profiles a codebase to identify which architectural patterns are in use, documents structural evidence for each, detects canonical combinations, and infers the compound structural rules that apply. Hostile — every pattern claim must be backed by concrete structural evidence. Writes a profile, not findings: the output…
Command
Hostile architectural audit: assumes violations exist and hunts for proof, validating the codebase against its detected (or declared) architecture.
Command
Adversarial, exhaustive review of the entire repository across code, tests, documentation, and configuration, with optional fixing in the same run. Assumes the code is incorrect until proven otherwise. This is the default command when the invocation names no other.
Command
Snapshot the current open findings as an accepted baseline so release-gate fails only on findings filed after it — adopting the toolkit on a repo with pre-existing debt without deleting, falsifying, or hiding a single real finding.
Command
Hostile audit of caching correctness: assume every cache serves stale data after a write, shares a key across entities that must not share one, grows without bound, and stampedes on expiry — then prove where. A cache is a correctness liability until its key, its invalidation, its bound, and its expiry behavior are all…
Command
Hostile audit of a project's CI/CD pipeline definitions (GitHub Actions first-class; GitLab CI and other YAML pipelines by the same principles): assume every workflow is exploitable or silently non-gating until each file is checked against every defect class.
Command
Hostile audit of every commit message in a range against the diff it labels: assume every label lies until the hunks prove it — release automation (release-please, semantic-release) reads only the message, so a mislabeled commit silently mis-versions the release.
Command
Forces the laziest solution that actually works — simplest, shortest, most minimal — on every coding task, and audits a diff, plan, or whole repo for over-engineering. Channel a lazy senior developer who has seen every over-engineered codebase and been paged at 3am for one: lazy means efficient, not careless, and the…
Command
Hostile single-shot concurrency-safety audit: assume every piece of state reachable from two or more concurrent execution contexts is corrupted until a happens-before edge proves otherwise, and file each defect with the shared state, the contexts that reach it, the corrupting interleaving, and a concrete…
Command
Hostile single-shot application/runtime configuration audit: cross-reference every config value read by code against every declared source of truth and hunt undocumented config, missing validation, unsafe prod defaults, config drift, committed secrets, type-coercion traps, and hardcoded environment values.
Command
Hostile audit of the project's public contract surface: assume the contract has drifted until every declared element is verified against the implementation, and every surface change since the last release tag is verified against the semver bump the commits declare.
Command
Tool-driven implementation of unresolved PR/MR review comments on GitHub, GitLab or Bitbucket: fetch every comment thread plus the out-of-thread notices (review bodies, bot summaries), evaluate each for technical validity, implement valid ones one at a time with a full validation pass after each, and scan the codebase…
Command
Hostile hunt for code that is unreferenced or unreachable — with the deletion-safety discipline that no symbol is called dead until every reachability channel is ruled out. Assume the codebase hides both cruft that should be deleted and live code that only looks dead; prove which is which. Grep-shows-no-callers is a…
Command
Hostile audit of every dependency the project already declares or silently relies on: every manifest entry is dead weight until an actual usage is proven, and every import is undeclared until a manifest names it.
Command
Hostile documentation review: assumes every claim in every doc is wrong until verified against the codebase. Two-directional — docs→code (does code match the claim?) and code→docs (is this code undocumented?).
Command
Hostile audit of application error handling: assume failures are being swallowed and prove where — every path where a real failure becomes silence (lost data, a fabricated success, an outage no operator sees) is a finding; a path that signals is not.
Command
Execute an already-approved implementation plan task by task, verifying each task the moment it is done, stopping when blocked instead of guessing, and finishing behind a gated commit/push menu. Assumes the plan is stale until re-checked against current code and that no task is done until its verification is observed…
Command
Prints the categorized command listing. Writes nothing, audits nothing.
Command
Hostile single-shot internationalization/localization audit against the project's declared locale scope: assume every user-facing string, number, date, and sort is hardcoded to one locale until the code proves it routes through a localization mechanism.
Command
Hostile audit of infrastructure-as-code — container images (Dockerfiles), orchestration (Kubernetes, Compose, Helm), and cloud provisioning (Terraform, CloudFormation, Pulumi): assume every image runs as root, every network is open to the world, every store is unencrypted, and every credential is committed, until each…
Command
Hostile single-shot resource-lifecycle audit: assume every acquired resource leaks on the failure path until a guaranteed-release construct proves otherwise, and file each leak with the acquisition site, the path where release is skipped, and the accumulation driver.
Command
Hostile single-shot audit of database schema and data migrations: assume every migration eats production until proven safe.