Verify & document subscription-routed LLM for agentic search (sampling / Agent SDK)

[thorwhalen]

Summary

ir's agentic LLM can run on a Claude subscription instead of a metered key, with zero ir code changes — every seam already injects a plain callable. This issue tracks documentation + the one real unknown.

See the design note: misc/docs/ir_10 -- LLM Routing ....

Facts (SSOT)

• Injectable seams: make_llm_formulator(rewriter=) (formulate.py:66), make_llm_selector(chooser=) (select.py:433), make_llm_synthesizer/with_synopsis(synthesize=) (synopsis.py:97/220), eval_gen (eval_gen.py:264/386). Default backend is aix (not oa — README drift fixed in docs: LLM routing on a Claude subscription (ir_10) + fix oa→aix drift #71).
• Routing targets: connector→Claude.ai via MCP sampling/createMessage (transport exists; client support unverified); Claude Code→host via claude_agent_sdk / coact (exists today); Desktop→sampling (same caveat).

Actions

• Document the seams + routing (ir_10, PR docs: LLM routing on a Claude subscription (ir_10) + fix oa→aix drift #71).
• Fix README oa→aix drift (PR docs: LLM routing on a Claude subscription (ir_10) + fix oa→aix drift #71).
• Verify: do Claude.ai remote connectors honour sampling/createMessage? (the only blocker for the connector path).
• Optional: ship a tiny reference adapter recipe in docs (the actual adapter lives connector-side — see Subscription-routed LLM for agentic connectors (MCP sampling / Agent SDK) enlace_connector#1).

Note

No ir code change is required for any of this; ir stays backend-agnostic. Closing condition: docs merged + the client-support question answered (then connectors can adopt or defer accordingly).

https://claude.ai/code/session_0154TnwYASjN1u1Hw97JQ2vd