A per-step thinking level in Settings, the highest for all by default #420

Closed
opened 2026-11-08 00:11:07 -05:00 by cmoriarty · 1 comment
Owner

What happens. Every step thinks at the model's highest level (braid#407 set the built-in map to high for all). In run 181 the judge rounds took 156 min at xhigh where run 174's took 50 min at medium, with sound criteria. The only way to lower a step today is a repository's own pipeline file.

Fix. A per-step thinking level in Settings: a table of the built-in steps (apply, fix, review, code review, Visual check, summarize, …) each with the model's levels, default the highest for all, saved with the profile and kept by a run for its life as the effort is now. The hover keeps showing the map in force.

Done when. The operator sets the Visual check to medium in Settings, the next run's judge requests carry it, apply's still carry xhigh, and a profile left alone sends the highest everywhere.

From run 181 (the owner, 2026-11-08: "open to adding settings for changing the thinking level per step, but by default, it is the highest for all").

**What happens.** Every step thinks at the model's highest level (braid#407 set the built-in map to high for all). In run 181 the judge rounds took 156 min at `xhigh` where run 174's took 50 min at `medium`, with sound criteria. The only way to lower a step today is a repository's own pipeline file. **Fix.** A per-step thinking level in Settings: a table of the built-in steps (apply, fix, review, code review, Visual check, summarize, …) each with the model's levels, default the highest for all, saved with the profile and kept by a run for its life as the effort is now. The hover keeps showing the map in force. **Done when.** The operator sets the Visual check to `medium` in Settings, the next run's judge requests carry it, apply's still carry `xhigh`, and a profile left alone sends the highest everywhere. From run 181 (the owner, 2026-11-08: "open to adding settings for changing the thinking level per step, but by default, it is the highest for all").
Author
Owner

Shipped in PR #421 (ff9381d8), change per-step-thinking-level. A profile keeps step_efforts, a thinking level per built-in step family (explore and propose, review and revise, apply, code review, fix rounds, Visual check, summarize, plan), as a table under Reasoning effort with "Same as the profile" first; a family left alone thinks at the profile's effort, so a profile left alone sends the model's highest level to every step. A run keeps the table for its life (in its settings copy); the field's ⓘ lists the levels in force. Tests: tests/api/test_settings.py, tests/agent/test_effort.py, ui/src/lib/effort.test.ts, and a journey that sets the Visual check and reloads. Done when the next run's judge requests carry the level set: set it in Settings before the Scratch run.

**Shipped in PR #421 (ff9381d8), change `per-step-thinking-level`.** A profile keeps `step_efforts`, a thinking level per built-in step family (explore and propose, review and revise, apply, code review, fix rounds, Visual check, summarize, plan), as a table under Reasoning effort with "Same as the profile" first; a family left alone thinks at the profile's effort, so a profile left alone sends the model's highest level to every step. A run keeps the table for its life (in its settings copy); the field's ⓘ lists the levels in force. Tests: `tests/api/test_settings.py`, `tests/agent/test_effort.py`, `ui/src/lib/effort.test.ts`, and a journey that sets the Visual check and reloads. Done when the next run's judge requests carry the level set: set it in Settings before the Scratch run.
Sign in to join this conversation.
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
cmoriarty/braid#420
No description provided.