[khezen]

Summary

dspy 3.4.1 (2.2.0)

• Upgrade dspy 3.4.0 → 3.4.1 and keep the new RLM sub-agents prompt block.
• Double the max_llm_calls defaults so sub-agents have room (default 8 → 16).
• The CodespyRLM reasoning-description override now only applies while the stock description is still present, so it stops applying once dspy PR #10643 ships.

doc, summary, audit on RLM (2.3.0)

• doc, summary and audit now run on RLM instead of ChainOfThought, using the same setup as code_review: build_rlm_agent / run_rlm / log_rlm_outcome, the recall_memory tool when enabled, and the react fallback threshold (0.30).
• Summarizer and Auditor get an async aforward; their sync forward calls it through asyncio.run, so pipeline call sites are unchanged.
• Their budgets now inherit the code_review defaults (max_iters 5 / max_llm_calls 16). Changed in config_dspy.py, codespy.yaml, .env.example and the action.yml descriptions.
• Hippocampus Distiller and Cartographer stay on ChainOfThought.
• Docs updated: architecture.md, CHANGELOG.

Risks

• Higher cost and latency: summary and audit can now make up to 16 LLM calls each, instead of about 1.
• Summary stays on nemotron-super, and RLM output quality on that model has not been checked. Watch for log_rlm_outcome extract-fallback warnings.