Problem
The gap detector found only 1 skill gap (exit used 1x) in a 52-session, 467-Agent-call workload. It missed that Agent dispatches ARE the workflow — they're the equivalent of skills for orchestration sessions.
A session with 467 Agent calls, 20+ parallel worktree dispatches, and 5 distinct agent types (Explore, Plan, general-purpose, worktree-isolated, background) has rich workflow patterns that the gap detector ignores because they're tool calls, not slash commands.
What's needed
- Detect Agent tool calls and classify by subagent_type (Explore, Plan, general-purpose)
- Track agent dispatch patterns: parallel batches, sequential chains, isolation modes
- Report agent workflow statistics alongside skill usage:
## Agent Workflows
- Explore agents: 12 dispatches (codebase research)
- Plan agents: 3 dispatches (architecture decisions)
- Implementation agents: 15 dispatches (12 worktree-isolated, 3 inline)
- Background agents: 8 dispatches
- Average agents per user message: 2.3
- Identify gaps: "User manually did X 5 times that could be an agent dispatch"
Evidence
52 sessions with 467 Agent calls — gap detector reported nothing useful. The real patterns (parallel implementation agents, explore-before-implement, background monitoring) were invisible.
Problem
The gap detector found only 1 skill gap (exit used 1x) in a 52-session, 467-Agent-call workload. It missed that Agent dispatches ARE the workflow — they're the equivalent of skills for orchestration sessions.
A session with 467 Agent calls, 20+ parallel worktree dispatches, and 5 distinct agent types (Explore, Plan, general-purpose, worktree-isolated, background) has rich workflow patterns that the gap detector ignores because they're tool calls, not slash commands.
What's needed
Evidence
52 sessions with 467 Agent calls — gap detector reported nothing useful. The real patterns (parallel implementation agents, explore-before-implement, background monitoring) were invisible.