agrag.agents.middleware

Agent middleware for composing models and bounding the research loop.

Classes:

• HideToolsMiddleware – Remove tools by name from every model request.
• RequireVerdictMiddleware – Ask again when the verifier answers in prose instead of with its verdict.
• ResearchAttemptLimiter – Cap how many times the planner may re-delegate after verification.
• RoundRobinModelMiddleware – Rotate across the configured chat models, one model per call.
• VerifierEvidenceMiddleware – Give the verifier the evidence text behind each key a task cites.