oc-pilot 0.5.1 → 0.88.1

package/docs/integrations.md

@@ -22,7 +22,7 @@ provider configured for the selected engine.
 
 ```sh
 # Works now, before the first npm release (pin the tag):
-npm install -g https://github.com/MyNameIsNikki/oc-pilot/archive/refs/tags/v0.5.1.tar.gz
+npm install -g https://github.com/MyNameIsNikki/oc-pilot/archive/refs/tags/v0.5.2.tar.gz
 oc-pilot --version
 oc-pilot mcp --version
 ```
@@ -75,6 +75,12 @@ or set `engine`, `baseUrl`, and `apiKeyEnv` in `.oc-pilot.json`. Run
 The key is read from `OC_PILOT_API_KEY` or `OPENAI_API_KEY`; configure another
 variable with `--api-key-env`. Values are never printed in logs or errors.
 
+With `--engine openai`, a diff larger than one prompt (300 000 characters) is
+split into per-file slices; each slice gets its own API call and the responses
+are merged into a single analysis (`meta.slices` shows the call count). No
+files are dropped by truncation. `fix` always uses a single prompt because it
+must return one patch.
+
 `setup` preserves existing servers and unrelated settings. Use `--http` and
 `--token` to point a client at a remote server instead of a local stdio process.
 The token itself is never written to the config: the generated entry references
@@ -102,6 +108,26 @@ All tools return:
 A policy failure is **not** an MCP tool error. The agent should read
 `structuredContent.findings` and decide what to do.
 
+## MCP resources
+
+The server also exposes read-only context through MCP resources and resource
+templates, so hosts can pull project context without invoking a model:
+
+| Resource | Contents |
+| --- | --- |
+| `oc-pilot://config` | Resolved configuration and config path, JSON |
+| `oc-pilot://rules` | `.oc-pilot/rules.md` with secrets redacted, or a notice |
+| `oc-pilot://version` | Version and MCP protocol version, JSON |
+| `oc-pilot://prompt/en`, `oc-pilot://prompt/ru` | The built-in system prompts |
+| `oc-pilot://diff/{mode}` | Git diff by `staged`, `unstaged` or `head`, secrets redacted |
+| `oc-pilot://scan/{mode}` | Offline secret scan findings (masked previews), JSON |
+| `oc-pilot://file/{+path}` | Any project file up to 2 MB with secrets redacted |
+
+Resources are listed with `resources/list` and `resources/templates/list` and
+read with `resources/read`. File paths must stay inside the project directory;
+binary files are rejected; detected secrets are always masked before the text
+reaches the host model.
+
 ### Argument conventions
 
 - `diff` is the **unified diff text**, not a path.
@@ -177,7 +203,7 @@ mounted workspace.
 ## Agent skills
 
 `skills/oc-pilot/SKILL.md` is a portable Agent Skill. It tells an agent when to
-call `scan`, `review`, `test_fix`, `fix`, and `docgen`, prefers MCP tools, and
+call `scan`, `review`, `invariants`, `test_fix`, `fix`, `test_gen`, `docgen`, `slop`, `mutate`, `fuzz`, `bankruptcy`, `query_ast`, `sandbox`, `hivemind_status`, `flaky`, `impact`, `deps`, `coverage`, `reviewers`, `test_guard`, `env_sync`, `mock_leak`, `ecosystem_verify`, `pack`, `pack_hologram`, `zoom_symbol`, `contract_drift`, `anti_regression`, `boundaries`, `test_plan`, `compliance_verify`, `brag`, `battle`, `cleanup`, `arch_diff`, `roast`, `cinema`, `bounty`, `autopsy`, `echo`, `sast`, `authz`, `preflight`, `ready`, and `receipt_verify`, prefers MCP tools, and
 documents safety rules for `fix.apply`. Install it by copying
 `skills/oc-pilot` into the client's skills directory (for Claude Code, usually
 `.claude/skills/oc-pilot`).
@@ -205,8 +231,8 @@ Or register it from the CLI:
 claude mcp add oc-pilot -- oc-pilot-mcp --cwd /absolute/path/to/project
 ```
 
-Claude Code will expose `review`, `test_fix`, `fix`, `docgen`, `custom`, and
-`scan` as MCP tools. A typical instruction:
+Claude Code will expose `review`, `invariants`, `test_fix`, `fix`, `test_gen`, `docgen`, `custom`,
+`scan`, `slop`, `mutate`, `fuzz`, `bankruptcy`, `query_ast`, `sandbox`, `hivemind_status`, `flaky`, `impact`, `deps`, `coverage`, `reviewers`, `test_guard`, `env_sync`, `mock_leak`, `ecosystem_verify`, `pack`, `pack_hologram`, `zoom_symbol`, `contract_drift`, `anti_regression`, `boundaries`, `test_plan`, `compliance_verify`, `brag`, `battle`, `cleanup`, `arch_diff`, `roast`, `cinema`, `bounty`, `autopsy`, `echo`, `sast`, `authz`, `preflight`, `ready`, and `receipt_verify` as MCP tools. A typical instruction:
 
 > Before proposing a fix, call the `oc-pilot` `review` tool with
 > `fail_on: "warning"` and include its critical findings in your answer.
@@ -368,7 +394,7 @@ server. Add it to `~/.dsh/profiles/web/cordis.patch.yml`:
         serverName: oc_pilot
         transport: stdio
         command: npx
-        args: ['-y', 'oc-pilot@0.5.1', 'mcp', '--cwd', '/absolute/path/to/project']
+        args: ['-y', 'oc-pilot@0.5.2', 'mcp', '--cwd', '/absolute/path/to/project']
         toolCallTimeoutMs: 300000
 ```
 
@@ -432,13 +458,27 @@ In GitHub Actions, `--github` prints inline annotations instead of JSON:
 oc-pilot review --base origin/main --fail-on warning --github
 ```
 
+For GitHub Code Scanning, produce a SARIF 2.1.0 report and upload it:
+
+```yaml
+- run: oc-pilot scan --base origin/${{ github.base_ref }} --format sarif -o oc-pilot.sarif
+- uses: github/codeql-action/upload-sarif@v3
+  with:
+    sarif_file: oc-pilot.sarif
+```
+
+`--format sarif` is supported by `scan` and `review`. Rules are stable ids:
+`oc-pilot/secret` (secrets, error), `oc-pilot/critical` (error),
+`oc-pilot/warning` (warning; style findings are `note`). Secret previews stay
+redacted in the report.
+
 The repository also ships a composite action, so the whole gate is:
 
 ```yaml
 - uses: actions/checkout@v4
   with:
     fetch-depth: 0
-- uses: MyNameIsNikki/oc-pilot@v0.5.1
+- uses: MyNameIsNikki/oc-pilot@v0.88.1
   with:
     command: scan
     args: --base origin/${{ github.base_ref }} --fail-on-secrets
@@ -450,7 +490,111 @@ A tested patch loop is also available, but it writes to the checkout:
 oc-pilot fix --diff changes.diff --apply --command "npm test"
 ```
 
-The patch is reverted when tests fail. Do not run it on a shared working tree.
+Before applying, `fix` saves a snapshot of tracked changes (binary patch) and
+untracked files under `.git/oc-pilot/snapshots/`. The patch is still reverted
+when tests fail. Restore a snapshot with `oc-pilot restore` (or
+`oc-pilot restore --list` to pick one). Snapshots never enter commits;
+`--no-snapshot` skips the safety copy.
+
+## Git hooks
+
+`oc-pilot install-hook` installs native POSIX sh hooks without husky:
+
+```sh
+oc-pilot install-hook              # pre-commit: oc-pilot preflight --staged
+oc-pilot install-hook --scan       # lighter pre-commit: oc-pilot scan --staged
+oc-pilot install-hook --pre-push   # also pre-push: oc-pilot review --fail-on critical
+oc-pilot install-hook --uninstall  # remove managed hooks, restore backups
+```
+
+A foreign hook file is never overwritten silently: without `--force` the
+install fails, with `--force` the old file is kept as `<hook>.oc-pilot.bak`.
+`OC_PILOT_SKIP=1` bypasses the hooks for a single command. `core.hooksPath`
+and worktrees are respected. `oc-pilot doctor --fix` refreshes outdated
+managed hooks (keeping a backup) and prunes expired cache entries; foreign
+hooks are never touched.
+
+## pre-commit framework
+
+The repository ships `.pre-commit-hooks.yaml`, so the standard
+[pre-commit](https://pre-commit.com) framework works without local setup:
+
+```yaml
+repos:
+  - repo: https://github.com/MyNameIsNikki/oc-pilot
+    rev: v0.88.1
+    hooks:
+      - id: oc-pilot-secrets  # oc-pilot scan --staged
+      - id: oc-pilot-review   # oc-pilot review --staged --fail-on critical --fail-on-secrets
+```
+
+Both hooks use `language: system` and expect `oc-pilot` on the `PATH`
+(`npm i -g oc-pilot`). After changing the pinned `rev`, run
+`pre-commit install-hooks` once.
+
+## Secret ignore rules
+
+False positives can be silenced per path or per line. `.oc-pilot/ignore`
+holds one glob per line (`#` starts a comment); the inline marker
+`oc-pilot:ignore-secret` silences a single source line:
+
+```
+# .oc-pilot/ignore
+fixtures/**
+docs/examples/**
+
+const TOKEN = "sk-proj-..."; // oc-pilot:ignore-secret
+```
+
+The rules apply to every scanning point: diffs, project rules, docgen, and
+full-file context. High-entropy assignments (20+ chars, Shannon entropy over
+4 bits) are reported as critical `High-entropy secret`.
+
+## Baseline and PR comments
+
+Known findings can be grandfathered so `review` fails only on what is new:
+
+```sh
+oc-pilot review --staged --update-baseline   # write .oc-pilot/baseline.json, exit 0
+oc-pilot review --staged --baseline --fail-on critical
+```
+
+A fingerprint is `severity | file | normalized text` — the line number is
+deliberately excluded, so a finding survives code shifts above it. Baseline
+filtering applies to the exit code, text output, `--json`, SARIF, and PR
+comments; hidden findings are counted in `meta.baseline` (`known`, `total`).
+
+`review --pr` posts the review to the current pull request. Inside GitHub
+Actions it posts through the review API with `GITHUB_TOKEN` — the workflow needs
+`permissions: pull-requests: write` and the default `GITHUB_TOKEN` secret; no
+`gh` install or extra credentials are required. Up to 25 findings become inline
+comments on the diff line they point at; findings that do not map to a changed
+line (or whole runs outside Actions) fall back to a single summary comment,
+which locally goes through the GitHub CLI (`gh pr comment --body-file -`) and
+must be installed and authenticated. `--no-pr-inline` forces the summary
+comment. `--pr-status` additionally posts a commit status under the `oc-pilot`
+context: `failure` while critical findings remain, `success` otherwise, so a
+branch protection rule can require it (needs `statuses: write`). On failure the
+command exits 1 with a hint.
+
+## Response cache
+
+With `engine: openai`, identical requests are served from a local cache under
+`.git/oc-pilot/cache/` — re-running review on the same diff costs zero tokens.
+The key is `sha256(model + base URL + system prompt + input)`; entries live
+for 24 hours by default (`cacheTtlSeconds`, env `OC_PILOT_CACHE_TTL`, `0`
+disables caching). Triage, main, and slice calls all use the cache; the
+opencode engine is never cached. Cache hits are reported in `meta.cache`
+(`hit` | `miss`), and `--no-cache` bypasses the cache for a single run.
+
+```sh
+oc-pilot cache              # stats: entries, size, expired
+oc-pilot cache --prune      # drop expired entries
+oc-pilot cache --clear      # drop everything
+```
+
+`oc-pilot doctor` reports cache health and `oc-pilot doctor --fix` prunes
+expired entries.
 
 ## Troubleshooting
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oc-pilot",
-  "version": "0.5.1",
+  "version": "0.88.1",
   "description": "Local AI code review, tested fixes, test-failure analysis and secret scanning for agent CLIs",
   "type": "module",
   "repository": {
@@ -43,6 +43,7 @@
     "dev": "npm run build && node dist/cli.js",
     "mcp": "node dist/mcp.js",
     "test": "npm run build && node --test \"dist/**/*.test.js\"",
+    "test:coverage": "npm run build && node --test --experimental-test-coverage \"dist/**/*.test.js\"",
     "check": "tsc -p tsconfig.json --noEmit",
     "prompt:export": "npm run build && node scripts/export-prompt.mjs",
     "prepublishOnly": "npm test"

package/prompts/oc-pilot.en.md

@@ -40,12 +40,19 @@ Follow the recipe in the input:
 - Follow the instruction in ADDITIONAL_CONTEXT while respecting these principles and the output schema.
 - If code or signatures are missing, list exactly what to request.
 
+#### 5. `commit` (commit message generation)
+- Analyze GIT_DIFF and produce EXACTLY ONE Conventional Commits message; when RECENT_COMMITS is provided, match the project's style.
+- Format: a line like `type(scope)?: subject` (max 72 characters, imperative mood), then a blank line and an optional body (what and why, wrapped at ~72). For breaking changes append `!` after the type and/or a `BREAKING CHANGE: description` footer.
+- Types: feat, fix, docs, style, refactor, perf, test, build, ci, chore.
+- Output ONLY the raw commit message: no markdown, no code fences, no explanations or quotes.
+
 ---
 
 ### INPUT SCHEMA
 PROJECT_TYPE: <language/stack>
-RECIPE: <review | test-fix | docgen | custom>
+RECIPE: <review | test-fix | docgen | custom | commit>
 ADDITIONAL_CONTEXT: <signatures, AST excerpt, test output, or empty>
+SECRET_SCAN: <secret scanner report>
 GIT_DIFF:
 <diff content>
 

package/prompts/oc-pilot.ru.md

@@ -40,14 +40,21 @@
 - Выполни инструкцию из ADDITIONAL_CONTEXT, соблюдая принципы и формат вывода.
 - Если для ответа не хватает кода или сигнатур, перечисли, что именно нужно запросить.
 
+#### 5. `commit` (Генерация commit-сообщения)
+- Проанализируй GIT_DIFF и сгенерируй РОВНО ОДНО commit-сообщение в стиле Conventional Commits; если передан RECENT_COMMITS, ориентируйся на стиль проекта.
+- Формат: строка вида `тип(область)?: краткое описание` (до 72 символов, повелительное наклонение), затем пустая строка и необязательное тело (что и почему, ширина ~72). Для ломающих изменений добавь `!` после типа и/или футер `BREAKING CHANGE: описание`.
+- Типы: feat, fix, docs, style, refactor, perf, test, build, ci, chore.
+- Выведи ТОЛЬКО готовое commit-сообщение: без markdown, без блоков кода, без пояснений и кавычек.
+
 ---
 
 ### ФОРМАТ ВХОДНЫХ ДАННЫХ (INPUT SCHEMA)
 Пользователь передает данные в следующем виде:
 
 PROJECT_TYPE: <язык/стек>
-RECIPE: <review | test-fix | docgen | custom>
+RECIPE: <review | test-fix | docgen | custom | commit>
 ADDITIONAL_CONTEXT: <сигнатуры, AST-выжимка, вывод тестов или пусто>
+SECRET_SCAN: <отчёт сканера секретов>
 GIT_DIFF:
 <содержимое diff>
 

package/server.json

@@ -2,7 +2,7 @@
   "$schema": "https://static.modelcontextprotocol.io/schemas/2025-07-09/server.schema.json",
   "name": "io.github.mynameisnikki/oc-pilot",
   "description": "Local AI code review, tested fixes, test-failure analysis and secret scanning for agent CLIs",
-  "version": "0.5.1",
+  "version": "0.5.2",
   "repository": {
     "url": "https://github.com/MyNameIsNikki/oc-pilot",
     "source": "github"
@@ -12,7 +12,7 @@
     {
       "registry_type": "npm",
       "identifier": "oc-pilot",
-      "version": "0.5.1",
+      "version": "0.5.2",
       "transport": {
         "type": "stdio"
       }

package/skills/agent-self-review/SKILL.md

@@ -0,0 +1,87 @@
+---
+name: agent-self-review
+description: Use before committing or handing off AI-generated code. Runs a local secret scan and an independent review of the staged diff, fixes blocking findings, and re-verifies until the diff is clean or the remaining issues are explicitly documented.
+---
+
+# Agent self-review
+
+A coding agent should not ask a human to review code it never checked itself.
+This skill turns "I think it works" into a verified claim: secrets scanned,
+diff reviewed by a second model, blocking findings fixed, exit code checked.
+
+## Setup
+
+```bash
+npm i -g https://github.com/MyNameIsNikki/oc-pilot/archive/refs/tags/v0.88.1.tar.gz
+oc-pilot --version   # requires 0.11+
+```
+
+Works with no API key when OpenCode is the configured engine. For a hosted
+model, export a key and pass `--engine openai --model <id>`.
+
+## Workflow
+
+1. **Scan secrets first.** This is offline and instant, so it costs nothing to
+   run before any model call.
+
+   ```bash
+   oc-pilot scan --staged --json
+   ```
+
+   Exit `3` means critical secrets are present. Stop and remove them; do not
+   "fix" them by rewriting history you have not inspected.
+
+2. **Review the exact diff that will be committed.**
+
+   ```bash
+   oc-pilot review --staged --fail-on critical --fail-on-secrets
+   ```
+
+   Read the `#### Замечания / Findings` section. Every bullet has a severity
+   and a `file:line` location. Treat `critical` as blocking. Warnings are a
+   judgement call; styles are optional.
+
+3. **Fix only what was found.** Feed the findings back into the same code you
+   just wrote, not into a rewrite. Prefer the smallest change that removes the
+   defect.
+
+   ```bash
+   oc-pilot fix --apply --command "npm test"
+   ```
+
+   `fix` proposes a unified diff, checks it with `git apply --check`, applies
+   it, runs tests, and reverts the patch if tests fail. A snapshot is written
+   before applying, so `oc-pilot restore` can roll back.
+
+4. **Re-verify.** A fix is not done until the gate is green again.
+
+   ```bash
+   oc-pilot review --staged --fail-on critical --fail-on-secrets
+   ```
+
+   Repeat steps 3–4 at most twice. If the same finding survives two attempts,
+   stop and report it to the user with the file, line, and why it is hard to
+   fix. Do not loop silently.
+
+5. **Report.** In the final message state: what was reviewed, what was fixed,
+   what remains, and the exit code of the last run.
+
+## Reading the result
+
+| Exit | Meaning | Agent action |
+| --- | --- | --- |
+| `0` | No findings at or above the threshold | Proceed to commit |
+| `3` | Policy gate failed (findings or secrets) | Fix, then re-run |
+| `1` | Runtime error (engine, network, git) | Report, do not guess |
+| `2` | Bad arguments | Fix the command |
+
+## Rules of engagement
+
+- Never disable the gate to make it pass. No `--no-secret-scan`, no
+  `--fail-on none`, no deleting `.oc-pilot/ignore` entries.
+- If a finding is a false positive, hide it deliberately with
+  `oc-pilot review --staged --interactive` and say so in the report.
+- Do not commit generated lockfiles, vendored trees, or fixtures that the
+  review was told to ignore.
+- If the repository has project rules in `.oc-pilot/rules.md`, do not edit
+  them to silence a finding.

package/skills/commit-discipline/SKILL.md

@@ -0,0 +1,86 @@
+---
+name: commit-discipline
+description: Use when an agent has finished a task and needs to commit. Generates a Conventional Commit message from the staged diff, matches the repository's existing history, blocks commits that contain secrets, and commits with a single reviewed command.
+---
+
+# Commit discipline
+
+"I updated the code" is not a commit message, and `git add -A && git commit -m
+"fix"` hides the one thing a reviewer needs: why. This skill produces a message
+derived from the actual diff, in the repository's own style.
+
+## Setup
+
+```bash
+npm i -g https://github.com/MyNameIsNikki/oc-pilot/archive/refs/tags/v0.88.1.tar.gz
+```
+
+## Workflow
+
+1. **Stage deliberately.** One logical change per commit.
+
+   ```bash
+   git add path/to/changed path/to/changed.test
+   git status --short
+   ```
+
+   Never `git add -A` from an agent session: it picks up scratch files, logs,
+   and local config.
+
+2. **Generate the message.**
+
+   ```bash
+   oc-pilot commit
+   ```
+
+   The command reads the staged diff and the last ten commit subjects, then
+   returns a single Conventional Commit message: `type(scope): imperative
+   subject` with an optional body. Types: `feat`, `fix`, `docs`, `style`,
+   `refactor`, `perf`, `test`, `build`, `ci`, `chore`.
+
+3. **Inspect before applying.** The message must state the behaviour change,
+   not the file list. Reject messages that only say "update", "improve", or
+   restate the diff.
+
+4. **Commit with the gate.**
+
+   ```bash
+   oc-pilot commit --apply
+   ```
+
+   `--apply` is allowed only with staged changes. If the staged diff contains a
+   critical secret, the command exits `3` and does not commit. Use
+   `--allow-secrets` only when you have confirmed the value is public, and say
+   so in the PR.
+
+5. **Verify.**
+
+   ```bash
+   git log -1 --stat
+   ```
+
+   Confirm the commit contains exactly the intended files.
+
+6. **Machine-readable output** for scripts and CI:
+
+   ```bash
+   oc-pilot commit --json
+   ```
+
+   Returns `message`, `subject`, `body`, `committed`, `diffSource`, `stats`,
+   `secrets`, `engine`, `model`, `durationMs`.
+
+## Message rules
+
+- Subject in the imperative mood, 72 characters or fewer, no trailing period.
+- Scope names the module, not the file extension.
+- Body explains the reason and any migration, wrapped at 72 columns.
+- One concern per commit. If the diff mixes a refactor with a behaviour
+  change, split it: `git add -p`, then repeat.
+
+## Guardrails
+
+- Do not commit generated artifacts unless the repository tracks them.
+- Do not amend or force-push on a shared branch from an agent session.
+- If the message does not fit the change, the change is too large or too
+  vague — split it instead of inflating the subject.

package/skills/oc-pilot/SKILL.md

@@ -16,7 +16,47 @@ If the `oc-pilot` MCP server is configured, call its tools directly:
 - `review` — bugs, security issues, idioms, and structured findings.
 - `test_fix` — root cause of a failing test plus a minimal fix.
 - `fix` — propose a unified diff; with `apply: true`, apply it and run tests, reverting on failure.
+- `test_gen` — write tests for changed code; with `apply: true`, apply them and run the suite, reverting failures unless `keep_failing`.
 - `docgen` — GoDoc/RustDoc/JSDoc/docstrings.
+- `slop` — offline Slop Index.
+- `mutate` — mutation testing: flips operators, runs tests, lists survivors; needs `--allow-write` because it rewrites sources temporarily.
+- `flaky` — repeated and shuffled test runs that flag tests passing in some runs and failing in others.
+- `impact` — declarations changed by the diff and every untouched file that still references them.
+- `deps` - dependency audit: phantom and missing imports, unused packages, lockfile drift, OSV vulnerabilities and licenses; `fix` and `sbom` need `--allow-write`.
+- `coverage` - changed lines missing from a local lcov, Istanbul, Cobertura or Go coverage report, with an optional gate.
+- `reviewers` — authors who touched the changed files plus CODEOWNERS owners; accepts `limit`, `depth` and `exclude`.
+- `test_guard` — offline test-weakening guard: skipped/focused/deleted tests, removed assertions, matcher downgrades, `@ts-ignore`, empty catch.
+- `env_sync` — environment variables added by the diff against `.env` and `.env.example`: missing definitions and undocumented keys.
+- `mock_leak` - forgotten stubs and hardcode in production code: local paths, test tokens, always-true conditions, mock identifiers and localhost URLs.
+- `ecosystem_verify` - named imports against installed packages and `.d.ts`, plus age and downloads of newly added npm packages.
+- `pack` - compact task context: rules, changed-file signatures, related references and an optional test log, capped by `budget` characters; read-only.
+- `pack --hologram` - layered context: LOD 0 full diff files, LOD 1 hidden bodies with EFFECTS markers and matching tests, LOD 2 export names; `zoom-symbol` returns one hidden body on demand.
+- `invariants` - semantic lock file `.oc-pilot/invariants.md` checked against the diff with a cheap model; `[CRITICAL]` and `--strict` gate with exit code 3, out-of-scope rules are skipped without a model call.
+- `fuzz` - safe JIT fuzzing of exports added by the diff in separate processes; side-effect files are blocked, crashes and timeouts carry the exact input.
+- `bankruptcy` - deterministic tech-debt forecast from commit history: smoothed score, slope, forecast date and an SVG chart; test weakening and security dominate the weights.
+- `query` - chainable AQL over a lightweight code graph: find, where, sort, limit, select and count; MCP tool `query_ast`.
+- `sandbox run` - execute one JS/ESM file in node:vm with imports, require, fetch and process.exit blocked; output and exact status are reported.
+- `hivemind` - encrypted TCP mesh: ed25519 identity, trusted peers, ECDH X25519 handshake and AES-256-GCM payloads; only diffs, AST and LLM requests are allowed.
+- `contract_drift` - removed or newly required fields and type changes in TypeScript, Zod, Prisma, Pydantic and JSON Schema, plus untouched consumers of a removed field.
+- `anti_regression` - lines the diff re-adds after a human explicitly deleted them in git history, with the deleting commit cited.
+- `boundaries` - added imports checked against `.oc-pilot/architecture.json` layer rules, with file, line and rule reason.
+- `test_plan` — select the test files affected by the diff from the import graph; report uncovered changed files and gate with `fail_on_uncovered`.
+- `compliance_verify` — verify the signed audit manifest for HEAD: hash, commit, previous-manifest chain and Ed25519 signature.
+- `brag` - self-contained HTML/SVG session report from stats, git history, last review and receipt; inline assets only, no model or network.
+- `battle` - experimental coder/hacker loop: the coder patches, the hacker reviews and runs SAST, authz and test-weakening checks until no high findings remain and tests pass.
+- `fix --shadow` / `battle --shadow` - run apply in a detached `.oc-pilot/shadow-tree` worktree and land the diff only after a successful run.
+- `cleanup` - offline cleanup plan from dependency issues, Slop Index, environment variables and weakened tests; `apply` performs only safe dependency fixes and needs `--allow-write`.
+- `arch_diff` — Mermaid diagram of added (green) and removed (red) internal imports, with an optional `fail_on_removed` gate.
+- `roast` — model-backed comedy roast with linus/ramsay/corporate/default personas; every joke cites file:line and ends with one real fix.
+- `cinema` — terminal timelapse of the session from commits, runs, API calls, last review and receipt; self-contained HTML player or asciinema v2 cast.
+- `autopsy` — forensic post-mortem for a commit: diff stats, SAST on added lines, weakened tests, mock leaks, ranked suspects and a before/after timeline; `--fail-on-risk` exits 1 on a dangerous verdict.
+- `echo` — self-contained SVG badges from local stats/git/receipt, with opt-in anonymous aggregate submit to an explicit endpoint; no network by default.
+- `bounty` — ASCII WANTED poster for the worst file scored from SAST, duplicates, dead exports, aged task comments and churn; deterministic and offline.
+- `sast` — deterministic static analysis: injection, path traversal, SSRF, XSS, TLS and other dangerous patterns.
+- `authz` — route authorization: unauthenticated routes, IDOR lookups, admin paths without a role.
+- `preflight` — one offline verdict: secrets, SAST, route authorization, dependency health, test weakening, Slop Index delta.
+- `ready` — offline agent readiness score 0-100 across tests, lockfile, hooks, slop, receipt, docs, CI, config and security.
+- `receipt_verify` — verify the receipt for HEAD: its hash, recorded commit and whether its tests passed.
 - `custom` — any other code question with an optional diff.
 
 Pass `diff` as unified diff **text**. Omit it to let the server use the
@@ -29,12 +69,58 @@ call should fail the build.
 oc-pilot review --fail-on warning --fail-on-secrets
 oc-pilot scan --json
 oc-pilot test-fix
+oc-pilot test-gen --staged --apply
 oc-pilot fix --apply --command "npm test"
 oc-pilot docgen "src/**/*.ts"
+oc-pilot mutate --staged --min-score 80
+oc-pilot flaky --runs 5 --fail-on-flaky
+oc-pilot impact --staged
+oc-pilot deps --audit
+oc-pilot coverage --staged --coverage-file coverage/lcov.info --fail-under 80
+oc-pilot reviewers --base main
+oc-pilot test-guard --staged --fail-on-weakening
+oc-pilot env-sync --staged --fail-on-missing
+oc-pilot mock-leak --staged --fail-on-leak high
+oc-pilot ecosystem-verify --staged --fail-on-issue high
+oc-pilot pack "fix the login redirect" --staged --budget 12000
+oc-pilot pack --hologram --max-tokens 30000 --output context.md
+oc-pilot zoom-symbol src/core/pack.ts buildPack
+oc-pilot invariants --staged --fail-on violation
+oc-pilot fuzz --duration 2s --memory-limit 128MB
+oc-pilot bankruptcy --history 30d --format json
+oc-pilot query "find(CallExpression).where(name, 'stripe').select(file, line)"
+oc-pilot sandbox run scripts/demo.mjs --timeout 500ms
+oc-pilot hivemind init
+oc-pilot contract-drift --staged --fail-on-drift high
+oc-pilot anti-regression --staged --fail-on-regression
+oc-pilot boundaries --staged --fail-on-violation
+oc-pilot test-plan --staged --fail-on-uncovered
+oc-pilot compliance --verify
+oc-pilot brag --since-days 30 --output brag.html
+oc-pilot battle "fix the retry logic" --apply --rounds 3
+oc-pilot battle "harden auth" --apply --shadow
+oc-pilot fix --apply --shadow
+oc-pilot cleanup --staged --fail-on-task high
+oc-pilot arch-diff --staged --fail-on-removed
+oc-pilot roast --staged --persona linus
+oc-pilot cinema --since-days 30 --output cinema.html
+oc-pilot bounty --limit 10 --since-days 90
+oc-pilot autopsy HEAD~1 --depth 10
+oc-pilot echo --name my-team --svg --output badges.svg
 ```
 
 Use `--lang ru` for Russian model output. The CLI UI itself may be Russian.
 
+## Related skills
+
+Focused workflows built on the same CLI:
+
+- `agent-self-review` — pre-commit gate loop with a verified exit code.
+- `test-failure-triage` — root cause first, then a patch with a snapshot.
+- `secure-diff-gate` — secrets, security rules, injection and dependency risks.
+- `commit-discipline` — Conventional Commit message from the staged diff.
+- `stack-review-packs` — rules packs for 23 languages and frameworks.
+
 ## Safety
 
 - Never paste real secrets into a prompt. `oc-pilot` redacts common secrets in