agent: a stronger model can write the compaction summary, behind --compaction-summarizer-model

[jaredLunde]

Stacked on #240 (base jared/summary-candidates); it retargets to main once #240 merges. Needs #241 for a Sonnet summarizer to work in direct mode.

--compaction-summarizer-model <MODEL> (env AI_AGENT_COMPACTION_SUMMARIZER_MODEL, on run and serve) has another model write every compaction summary, e.g. claude-sonnet-5-5 for a Haiku or gpt-6-luna session. Swapping only the summarizer moved SWE-bench Verified 6.5 points in one study, and a weak summarizer did worse than no summary in another.

• The summarizer gets a transport of its own, built the way a subagent's is (Agent::with_summarizer), because credentials and routing are model-keyed.
• Every summary is then a re-render: only the session's own model can read its prompt cache, so the cached call is skipped.
• The request runs without the session's thinking and effort settings, which are model-specific.
• In serve, if the summarizer's client can't be built, a warning is logged and the session's model keeps summarizing.
• The Compacted event's usage is the summarizer's, so cost accounting must price it at that model's rates.

Unset, nothing changes. It's a plain flag rather than a --compaction-candidates name because that set only holds on/off values.

Eval results (probe-based replay + continuation) are reported separately; nothing here changes defaults.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JimHGjsfk2Ktm5GxyZJKKk

[jaredLunde]

Closing: the stronger-summarizer arm was inconclusive (+0.056, CI crossed 0) at ~40x the per-summary cost. Not shipping a dead flag.