Make reasoning_effort configurable per stage - #27
Merged
Conversation
qwen3.8-27b defaults to xhigh and produces reasoning blocks that dwarf the answer -- a live run's transcript shows the model re-deriving the check's own constraints (section ordering, the 120-character minimums, which issue numbers are citable) at length before doing any work. There was no way to ask for less: the payload was model + messages + tools and nothing else. reasoning_effort resolves through the same chain as the model -- [harness.profiles.NAME.<stage>] -> [harness.profiles.NAME] -> [harness.<stage>] -> [harness] -- because it is a per-stage decision for the same reason the model is. decompose and judge are single bounded calls where deliberation is cheap; execute runs a tool loop and pays the cost on every turn. Sent only when configured, so an endpoint that rejects unknown keys sees no change by default. Validated at load against a known set, and for every profile rather than only the selected one: an unrecognised value would otherwise be forwarded verbatim, ignored by the server, and leave a run that looks configured and is not. Measurement caveat worth recording: probing an LM Studio endpoint with a short prompt showed no difference between xhigh, medium and low (119/141/ 129 reasoning tokens). That probe was too easy to make the model think at any setting, so it is evidence about the probe, not about the parameter.
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What this changes
qwen3.8-27b defaults to xhigh and produces reasoning blocks that dwarf the answer -- a live run's transcript shows the model re-deriving the check's own constraints (section ordering, the 120-character minimums, which issue numbers are citable) at length before doing any work. There was no way to ask for less: the payload was model + messages + tools and nothing else.
reasoning_effort resolves through the same chain as the model -- [harness.profiles.NAME.] -> [harness.profiles.NAME] -> [harness.] -> [harness] -- because it is a per-stage decision for the same reason the model is. decompose and judge are single bounded calls where deliberation is cheap; execute runs a tool loop and pays the cost on every turn.
Sent only when configured, so an endpoint that rejects unknown keys sees no change by default. Validated at load against a known set, and for every profile rather than only the selected one: an unrecognised value would otherwise be forwarded verbatim, ignored by the server, and leave a run that looks configured and is not.
Checklist
make checkis green locallyare the product)
specs/(spec changes go through an issue first)Generated-by:trailer present if anagent wrote the change