fix(provider): prefer per-model temperature over agent override#27777
Closed
Alezander9 wants to merge 1 commit into
Closed
fix(provider): prefer per-model temperature over agent override#27777Alezander9 wants to merge 1 commit into
Alezander9 wants to merge 1 commit into
Conversation
`ProviderTransform.temperature(model)` encodes per-model-family temperature values that the model's provider requires (e.g. Moonshot's kimi-k2.* returns HTTP 400 "invalid temperature: only 1 is allowed for this model" if any other value is sent). The session/llm.ts call site consulted it in the wrong order: input.agent.temperature ?? ProviderTransform.temperature(input.model) so the built-in title agent's hard-coded `temperature: 0.5` (agent.ts:253) won over the kimi-required 1.0, and Moonshot rejected the request. Swap the operands. The transform value, when present, is the model-aware answer and should beat a generic agent default. For models that aren't in the transform map, the function returns `undefined` and the agent's preference still applies through `??` fallthrough. Affects: kimi-k2.* (incl. k2.6, thinking, k2-5), gemini, glm-4.6/4.7, minimax-m2, qwen - anywhere the transform has an entry.
Contributor
|
This PR doesn't fully meet our contributing guidelines and PR template. What needs to be fixed:
Please edit this PR description to address the above within 2 hours, or it will be automatically closed. If you believe this was flagged incorrectly, please let a maintainer know. |
Contributor
|
Thanks for your contribution! This PR doesn't have a linked issue. All PRs must reference an existing issue. Please:
See CONTRIBUTING.md for details. |
Contributor
|
This pull request has been automatically closed because it was not updated to meet our contributing guidelines within the 2-hour window. Feel free to open a new pull request that follows our guidelines. |
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.
Problem
ProviderTransform.temperature(model)insrc/provider/transform.tsreturns the temperature value a given model family requires (kimi-k2.* -> 1.0 or 0.6; gemini / glm-4.6 / glm-4.7 / minimax-m2 -> 1.0; qwen -> 0.55). Several of these are hard requirements: Moonshot rejects any other value for the kimi-k2 family withHTTP 400 invalid temperature: only 1 is allowed for this model.The call site in
packages/opencode/src/session/llm.ts:172consults that function in the wrong order:The built-in
titleagent (packages/opencode/src/agent/agent.ts:253) hard-codestemperature: 0.5. Because??short-circuits on the first non-nullish operand, the agent default always wins and a temperature of 0.5 is sent to e.g. kimi-k2.6, which then 400s.I observed this in production telemetry across multiple users on kimi-k2.6 (the most common case). The same code path will trip whenever an agent sets a temperature on any of the strict-required model families.
Fix
Swap the operands at the single call site:
The transform function exists precisely because it encodes model-aware knowledge of what each family expects; that knowledge should beat a generic agent default. For models not covered by the transform map (claude, openai, etc.), the function returns
undefinedand the agent preference still applies through normal??fallthrough - so the only behavior change is on the families already flagged as needing specific values.This also aligns temperature with the adjacent
topK(no agent override) and brings it closer totopP(stillagent ?? transform, buttopPhas no known strict-required cases today).Diff
Notes
This was originally fixed in a downstream fork (browser-use/browsercode#64) where it was hit in production telemetry. The same code is unchanged on
devhere, so opening the equivalent upstream PR.