crickets

Release discipline and mechanics in one skill — pre-release checklist (CI green on every OS, version bump committed, CHANGELOG authored, dist/ committed, paired-release order locked for cross-repo releases), then the mechanical cut (conventional-commit semver auto-sizing, CHANGELOG prepend, tag, push, `gh release create`).

Two-step cross-model prose pass for design docs and other authored prose — Gemini simplifies via the headless prose_pass.py script, then Claude verifies every fact-guard held and applies. The prose sibling of code-review's cross-review.sh.

Author + maintain a Diátaxis-style wiki for any repo. Live authoring guidance (mode selection + template-fill + filename style), ongoing drift detection + repair, one-shot migration of legacy audience-based wikis to the six-section documentation layout (how-to · reference · architecture · designs · explanation · operational; onboarding folds into how-to, marked with a mode-tutorial hint), and single-page mode classification with sub-agent fallback on ambiguous cases. Reads operator conventions from AgentMemory `_always-load/diataxis-*.md`; composes a base style-guide ⊕ on-demand voice overlay into authored drafts, and learns generalizable voice lessons from the operator's own edits (edit-driven capture, operator-gated for generality + scope); offers to capture judgment calls back as new conventions (operator-confirmed via permeable-boundary helper). Dispatches the existing `documenter` sub-agent for mechanical-write work; never auto-forks into wiki/ without preview. Subsumes the predecessor `migrate-to-diatax

Update or create a Diátaxis wiki page for the current repo (or another registered repo). Triggers when the operator says 'update the wiki', 'document this in the wiki', 'add this to the wiki', 'create a wiki page for X', or 'update <repo>'s wiki'. Resolves target repo via cwd (default) or explicit repo name (cross-repo via repo_registry from V4 #30 plan 1). Dispatches the documenter sub-agent for the structural edit with preview-before-write. Defers to diataxis-author skill for mode selection when needed. Honors per-repo .diataxis-conventions.md override. Does NOT auto-generate content — agent gathers context from the conversation; the skill handles dispatch + write contract.

Test coverage audit via the Beyonce Rule (uncovered behavior is behavior you've agreed to change silently), DAMP over DRY (tests read like specs, not production code), and the Prove-It pattern (every behavioral claim needs a falsifying test). Use when reviewing a PR for test gaps or when a feature lands without coverage.

The harness discipline every phase workflow and every persona operates under — phase-gated sessions, state-on-disk-not-conversation, single-threaded implementation with read-only sub-agent fan-out, the PLAN.md shape, wake-on-CI, no parallel implementers, and the single-cycle shape for background primitives. A re-home of standards that previously lived only in agentm's AGENTS.md / harness/principles.md and the operator's global ~/.claude/CLAUDE.md — this domain owns the standard now; those files keep a pointer.

Autonomy doctrine — classify every action as recoverable or unrecoverable, proceed on the former, stop on the latter. The stop-gate is reversibility, not destructiveness or blast-radius.

Fix breakage on a Dependabot PR. Trigger when (a) the current branch matches `dependabot/*` and CI is red, (b) the user asks to "fix the dependabot PR" / "make this dependency update pass", or (c) the user invokes `/dependabot-fix [pr-number]`. Reads failing CI logs and upstream CHANGELOG, applies a bounded fix loop, pushes commits to the Dependabot branch, comments residual risks on the PR. Never merges. Aborts honestly when the fix needs human judgment.

Structured privacy-engineering review of a diff/PR, adapted from OWASP ASVS V8 Data Protection, the OWASP Top 10 Privacy Risks, the NIST Privacy Framework, and ICO data-protection-by-design guidance. Runs a deterministic Semgrep pre-pass for the four mechanizable checks (PII-in-URL, PII-in-client-storage, plaintext-columns, third-party-sinks), then LLM-judges the non-mechanical practices (consent scoping, retention adequacy, erasure completeness, purpose limitation). Reports falsifiable PRIVACY-RISK <file>:<line> [ASVS-id] findings; never fixes.

Day-to-day testing principles — tests-are-sacred, verification-first, and the 3-layer pyramid — for Developer-tier users.

ADR authoring via the adr-shape convention — when to write, the 5-section shape with re-audit triggers, and the "why not the alternative" requirement per decision call.

Run the wiki-watcher (W1) — one idempotent cycle that detects doc-worthy changes in a watched repo (+ its active PLAN.md/design/ROADMAP.md), judges significance, and dispatches the documenter to author the update PR-default. Triggers when the operator says 'run the wiki watcher', 'watch the wiki', 'check for doc-worthy changes', or drives /wiki-watch on a loop (/loop or cron). One invocation = ONE cycle (NOT a daemon — DC-W3): cooldown-gated + cursor-backed so repeated runs never drop or double-dispatch. Opt-in: device toggle in .agentm-config.json + a per-repo .harness/wiki-watch.json marker. PR-default (a human merges); direct-commit opt-in per trusted repo. Graceful-skip when disabled / unconfigured / unregistered / gh unavailable. The engine is cross-host; the scheduling wiring is Claude-first (see Antigravity-Limitations).

Scan the current git diff (or working tree) for personal information — emails, personal paths, API keys, phone numbers — before commit or push. Surfaces findings as file:line with redaction suggestions; loops until clean. Invoke before any git push.

Diagnose the Obsidian/Google-Drive vault storage backend (the obsidian-vault plugin). Confirms the configured vault_path is a real MemoryVault, the plugin is the backend selection resolves to, and no unresolved GDrive/DriveFS sync-conflict files remain. Read-only — never writes the vault or the engine config. Invoke when the vault backend misbehaves, after installing the plugin, or on Antigravity (which has no session-start conflict nudge).

Structured security review via the three-tier boundary model: LLM API boundary (prompt injection / untrusted model output), persistence boundary (durable writes that are hard to roll back), and system execution boundary (shell commands, lateral movement risk). Work each tier in order; visual scanning is not this skill.