Skip to main content

Plan Critic

src/goal_orchestrator/plan_critic.py. Feature-flagged via plan_critic_enabled (default true).

What it does​

Plan Critic is a second LLM pass that validates every TaskGraph generated by PlanGenerator before the executor starts running it. The critic looks for structural and policy issues that a single-pass plan generator can miss.

Validation checks​

  • All skills referenced exist in the registered catalog.
  • No circular dependencies between tasks.
  • Each task's input mappings reference real outputs from prior tasks.
  • Side-effect skills (gmail.send, agent_manager.create, task_manager.create) have explicit user intent in the goal objective.
  • Plan stays within MAX_PLAN_STEPS (8).
  • No task uses a placeholder argument (subject="Subject", body="...") that would be blocked by the intent gate at execution time.

Integration​

PlanGenerator → TaskGraph
   │
   ▼ (if plan_critic_enabled)
PlanCritic.validate(task_graph, goal)
   │
   ▼
   Pass: forward to executor
   Fail: regenerate (up to MAX_PLAN_RETRIES = 3)

If the critic rejects 3 times in a row, the goal is flagged FAILED with the latest critic feedback as the failure reason.

Trade-off​

The critic costs one extra LLM call per plan. For simple goals (1–2 step plans), this is wasted. Disable via /config if you want to save tokens at the cost of catching planning errors at execution time instead of pre-flight.

Critic feedback in audit log​

Each critic decision is logged as plan_critic.passed or plan_critic.rejected with the structured reason. Useful for tracing why a goal failed during planning.

See also​