Why

Alignment (kind words and phrases), rule proposal and the ontology agent used a fixed low reasoning effort, whatever the workspace set. That was decided on 2026-09-24 from a gemini measurement. On deepseek-flash low thinks seven or eight thousand tokens a call: in the warm-base run (#1065) those steps were six tenths of the tokens per document, and four fifths of that was thinking.

What changes

• One setting for every call. chat_client_thinking and JUDGEMENT_EFFORT are gone; alignment, rule proposal and the ontology agent use the workspace's reasoning effort, like extraction, errata and time interpretation already did.
• none is a value. The settings API and the column's CHECK (migration 0105) accept it. DeepSeek ignores minimal; only none turns its thinking off. The ninth bench run had to drop the CHECK by hand for this.
• The settings page offers none, and its hint says every call uses the setting and that the value that turns thinking off depends on the endpoint.

Behaviour change for existing workspaces: a workspace set to minimal or none no longer gets low on alignment. One that wants the old behaviour sets low, which then also applies to extraction.

Measured

20 Re-DocRED documents, deepseek-flash, --judge 200 --errata, fresh base each. Thinking off: this branch, workspace none, two groups. Alignment on low: the 2026-09-30 run (dev @ 798fb24, extraction none).

• Tokens more than halve and the run is five times faster. Phrase alignment's output goes from 323k tokens to 4k with the input unchanged.
• Precision stays above 0044's 75.3% threshold in both groups. Misworded goes up, in the direction gemini showed, by much less (gemini: 4% to 31%).
• The two groups differ more than usual: 80 against 52 signatures bound. Without thinking, the two votes on a signature agree less from run to run. The low side is one run, so its own spread is not measured.
• Recall on 20 documents is noisy and half of the gap is at extraction (entity-pair recall 14.0% against 11.0% and 12.9%), which alignment does not touch.

So this trades some precision and stability for tokens and time, and the workspace chooses. Against 0044's token threshold, 19k to 22k is still more than three times over.

Checked

• The settings API accepted none against a freshly migrated database, and the model answered with it set.
• cargo clippy --workspace --all-targets -- -D warnings, cargo fmt --all --check, cargo test -p utopia-cli (schema version 95), -p utopia-llm; pnpm typecheck, pnpm test (228).
• The bench README has the section with the full table.

Notes for landing

• Migration 0105 and CURRENT_SCHEMA_VERSION 95. A deleted document's date no longer attests: evidence records when its passage speaks, and attestation is rebuilt from what remains #1059 holds 0104 and also sets 95: whichever lands second sets 96.
• The typed-graph bench measures a warm base on deepseek-flash: 42k tokens per document against 48k on a fresh one #1065 appends to the same end of the bench README; the second to land needs a rebase.

🤖 Generated with Claude Code