raven

Force-invoke Andie-jr, Raven's brownfield debug assistant, for any bug, error, regression, stack trace, or "not working" case. Use this when the automatic triage router did not fire but you want a structured 2-round triage. For planning or design, use /andie instead.

Force-invoke Andie, Raven's orchestration layer, for planning, architecture, tradeoff analysis, or any non-trivial decision. Use this when the automatic routers did not load Andie but you want her anyway. For bugs, errors, and "not working" cases, use /andie-jr instead.

Use when an architect wants to approve a pending library request. Updates manifest.json and notifies the requesting developer.

Use to run a full Raven boot diagnostic. Checks all files, agents, manifest, and hooks are correctly installed. If no manifest exists, runs first-run discovery instead of erroring.

Reviews security_log.md and promotes accumulated observations into permanent rules. Hardens confirmed patterns into CLAUDE.md. Promotes dynamic specialist profiles with 3+ uses into curated skills. Classifies findings into internal_raven_ops.md or general_security_patterns.md. The weekly learning loop.

Use PROACTIVELY when new code files are created. During coding — advises only, never blocks. At git commit — missing architecture doc becomes a hard block via pre-commit hook after 24 hours grace period. Ensures .raven/architecture.md exists and is versioned.

Use PROACTIVELY at session start and end to manage persistent memory. At start: loads recent session notes and open questions from .raven/memory/. At end: writes session summary with decisions, fixes, and carry-forwards. No external tools required — pure file I/O. Obsidian-compatible output.

Database discipline enforcer. Fires on PostEdit for any source file. Detects inline SQL in non-SQL files, missing ERDs, and broken migration numbering. Spawned as a sub-agent by db-specialist pre-flight. Also runs independently on every file save via PostEdit hook.

Use PROACTIVELY when any skill attempts to read sensitive files or perform restricted actions. Watches for unexpected reads of manifest.secrets.json, .env, SSH keys, or modifications to settings.json or manifest.json by non-approved processes. Hard blocks and alerts.

Use PROACTIVELY before any coding action, file creation, or tool use. Verifies .raven/manifest.json exists, is valid JSON, and has all required fields. Hard stops if manifest missing — this is the only agent that hard blocks during coding. All other agents advise only.

Use when designing or reviewing multi-agent systems, subagent orchestration, parallel vs sequential agent execution, context passing between agents, tool delegation, and agent memory patterns. Platform-neutral — covers production multi-agent architecture across AI runtimes.

Use for any AI/ML Engineering question. Assumes Andrej Karpathy (AI researcher) persona. Deep multi-dimensional analysis. Bullets not prose.

3-phase browser testing skill. Generates UI·UX·Data·Security·Correctness·Speed test cases, drives visible Chrome live, writes Raven-compatible blocker MDs. Human watches and guides throughout. Andie packed inside — no external skill calls.

USE PROACTIVELY whenever the user asks: 'explain', 'what does this mean', 'in plain English', 'break it down', 'simplify', 'translate for non-tech', 'help me understand', or asks for a layman summary of any prior Raven/specialist output. Feynman-style explainer. Returns 50-word Feynman + 100-word summary with Technical, Business, Functional bullets. Skip for direct technical questions — only for explanations.

USE PROACTIVELY whenever the user reports: a bug, error, exception, stack trace, regression, test failure, 'not working', 'broken', 'why is X failing', 'X stopped working', 'doesn't behave', unexpected output. Brownfield debug assistant. Max 2 rounds, max 3 roles, returns problem · root cause · fix · why · audit note · commit suggestion. Do NOT use for greenfield design — that goes to andie.

10 hooks across 5 events