delegation-audit

Audit model delegation and subagent effectiveness for a session — which models handled which subagent types, per-type success rates and average durations, and wasted delegations (heavy models on trivi

SKILL.md

GitHub

Delegation Audit

Audit how a Claude Code session delegated work: model-to-subagent mapping and whether each delegation paid off.

Input

The user provides: $ARGUMENTS

A session ID. If empty, fetch GET /api/sessions?limit=1 and audit the most recent session, stating which one.

Data Sources

Endpoint Returns
GET /api/workflows/{sessionId} The modelDelegation dataset (which models are delegated which subagent types) and the effectiveness dataset (per-type completion/success rate, avg duration, task success)
GET /api/agents Raw subagent records (type, model, status, depth, parent) to corroborate counts and statuses

Report Sections

1. Delegation Matrix

From modelDelegation: a model × subagent-type table of how many agents of each type each model ran.

Model explore code-review debugger ... Total

2. Effectiveness by Subagent Type

From effectiveness: per type, the success rate and average duration.

Subagent type Count Success rate Avg duration Verdict
Mark types below ~70% success as low-yield.

3. Wasted Delegations

Flag, with evidence:

  • A heavy model (e.g. Opus) assigned to a simple/low-stakes subagent type that a cheaper model handled successfully elsewhere — candidate for rebalancing.
  • Subagent types with low success rates (effort spent, task not completed).
  • Duplicate delegations: the same type spawned repeatedly with poor success (retry churn).

4. Rebalancing Suggestions

Concrete model reassignments grounded in the matrix and effectiveness data. State the type, the model used, the success rate, and the suggested model — only where the data supports it.

Output

  • Markdown tables for the matrix and effectiveness.
  • Success rates as percentages; durations in human units (e.g. 1m 12s).
  • Use ▲/▼ when comparing a type's success rate against the session-wide average.
  • Cite only numbers returned by the API; do not infer success rates that the effectiveness dataset does not provide.
  • If the dashboard is unreachable, tell the user to start it with npm start from the repo root.