Integrations' ai_brief sent max_tokens equal to the answer's length (40 for a session title, at most 200). Current models think by default and count it against max_tokens: flagon-io's Auto picks Claude Haiku 5.5, which thinks at effort medium, so it spent the whole allowance thinking and answered nothing. ai_brief took the empty text as no answer, so every AI session title fell back to its prompt (no title from the provider since at least 2026-10-18) and the plan's Triage suggestions came back empty (wrangler tail at 05:45: "flagon-io: the plan got no answer from the AI provider"). Thinking can't be turned off portably (Claude Opus 5.5 and Claude Sonnet 5.5 refuse thinking disabled with a 400), so a brief now sends its answer length plus 2,048 tokens of room; callers keep only a short answer whatever comes back. Migration 0059 asks again, once, for Triage items asked before this.
Checked: cargo test -p g1t-integrations (78, new: a brief leaves its model room to think); cargo check -p g1t-integrations --target wasm32-unknown-unknown. G1t-Session: ses_01m5bgsckce7gt943v64h219yy