Repository navigation
feat(settings): Make the Completer's reasoning effort a Development setting - #656
Merged
Merged
Conversation
omesser
force-pushed
the
feat/638-reasoning-effort-setting
branch
from
September 12, 2026 22:34
72b2e97 to
d8545bd
Compare
…etting The level #617 hardcoded as "low" becomes a Development row: a picker of low/medium/high plus a free-text field, seeded through AI_BUDDY_DIRECTOR_REASONING_EFFORT. Blank still sends "low", on both the chat-completions and the Responses paths. Closes #638 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ENDPOINT_CUSTOM is now PICKER_CUSTOM: the reasoning-effort picker rests on the same title, and it names no endpoint. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
omesser
force-pushed
the
feat/638-reasoning-effort-setting
branch
from
September 13, 2026 09:27
d8545bd to
baba3ef
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The reasoning effort #617 hardcoded as
lowbecomes a Development row.What changed
AI_BUDDY_DIRECTOR_REASONING_EFFORT,Settings::director_reasoning_effortandTextField::DirectorReasoningEffort, seeded throughmodel::env_or_fileindev_flags::seed— aMutex<String>besideMCP_BIN, because the value is astring and not a number.
Compositepopup oflow/medium/highthat fills in theTextFieldbelow it, the same shape as the Completerendpoint row from Pick the Completer endpoint from a list, or type one #465. The field is the setting; the popup is a convenience.
Nothing validates what is typed: on llama.cpp and oMLX the value goes to the
model's chat template, where the valid set belongs to the model file.
Endpointbakes the level in atendpoint_from, like the timeout and thecap, and
request_bodysends it verbatim on both paths —reasoning_efforton chat-completions,reasoning.efforton Responses.low. Not "send nothing", which wouldgive back the 17.5% lost-wake rate feat(director): Ask chat-completions for low reasoning effort, and drop it when refused #617 measured away.
Why only three in the picker
They are the only values every documented host takes. OpenAI adds
none/minimal/xhigh/max, xAI hasxhighbut notmax, Ollama hasmaxbut not
xhigh. The free-text field is what covers the rest.The two traps
completer_retargetshas a clause for the new field, so an Apply reachesthe running Director rather than only the file.
a_changed_reasoning_effort_retargetscovers it. The rebuild is also thewhole of
takes_effort's reset path, which is now recorded in a commentrather than written as code.
lowand refuseshighstill latches the field off forthe session — deliberately, rather than caching a server's validation rules.
The
trace_directorretry line now names the value that cost it:retrying without reasoning_effort=high.AnyCompleter::Httpis boxed: the new field pushedEndpointpast clippy'slarge_enum_variantthreshold.How to verify
cargo test --workspace,cargo clippy --all-targets -- -D warnings,cargo fmt --all -- --check,node --test tests/— all pass. New tests coverthe description and its placeholder, the env-frozen row and picker, seeding and
what blank means, the retarget clause, both request shapes verbatim, the level
on the wire out of
Endpoint::post, and the retry line.By hand: Settings → Development → HTTP limits, pick or type a level, then
AI_BUDDY_TRACE_DIRECTOR=1shows it on the next wake. GTK and Win32 could notbe compiled here (macOS); the Win32 port skips composite popups as it already
does for the Base URL picker, so Windows gets the field and not the picker until
#461.
Closes #638
— Claude Code agent, on @omesser's behalf.
🤖 Generated with Claude Code