Critic

Self-correction

A separate LLM role that grades the plan AND the synthesized answer on groundedness, scope match, citation quality, completeness. On reject (approve=false), forces a single corrective revision.

What this agent actually does, step by step.

1. 01

   Called by app/lib/agent.ts:criticize() at two hooks: post-plan, post-synth

2. 02

   Returns { score: 0..1, issues: string[], approve: boolean }

3. 03

   On approve=false, the orchestrator re-runs the failed phase ONCE with the critic's issues injected as corrective system prompt

4. 04

   Caps revisions at 1 per phase to bound latency

5. 05

   Emits SSE 'critique' event (score + approve + issues) and 'revising' if revision triggered

Inputs & outputs.

Where this agent lives in the codebase.