mk:agent-detector

Agent Detector

Priority: HIGHEST - Runs FIRST for every message. Automatically detects the correct agent, task complexity level, and model tier for each user message using a multi-layer scoring system. No manual agent selection needed -- the detector analyzes task content, explicit technology mentions, user intent, project context, and file patterns to route to the right agent with the right model.

When to Use

ALWAYS - Every user message, no exceptions. This skill fires before any other skill or agent action. It determines who handles the task, at what complexity level, and with which model.

Workflow

1. Safety baseline precheck (HARD GATE). Before any other step, scan the context window for the cached-sentinel marker emitted by safety-sentinel-inject.cjs:

   ## Safety baseline: verified (cached, session <id>)
   

   If the marker IS present (turns 2..N of the same session): emit Safety baseline: verified (cached) and SKIP the 5-file Read loop below. The 5 rules remain in context from the inner harness's CLAUDE.md auto-load; the sentinel only suppresses redundant re-reads.

   If the marker is ABSENT (turn 1 of a new session, OR MEOWKIT_SKIP_SAFETY_SENTINEL=off): Read each of the 5 always-on safety/baseline rules:

   • .claude/rules/security-rules.md
   • .claude/rules/injection-rules.md
   • .claude/rules/gate-rules.md
   • .claude/rules/core-behaviors.md
   • .claude/rules/development-rules.md

   If any Read returns "file does not exist" or equivalent, ABORT IMMEDIATELY with the exact message:

   SAFETY BASELINE INCOMPLETE: rule <name> not found at .claude/rules/<name>.md
   Refusing to route any task. Restore the rule (git checkout / the installer) before retrying.
   

   Do NOT proceed to detection. Do NOT route to any agent. The 5 rules are the deterministic baseline; their absence indicates either repo corruption or a partial install. Replaces the unverified directory-auto-load assumption with a positive existence check.