Repository navigation
fix(anthropic): keep Claude tool turns streaming and Claude 5.5 thinking valid - #98
Merged
Merged
Conversation
Haiku 5.5 accepts disabled thinking, but Sonnet 5.5 answers 400 unless it is sent as between_tools, and Opus 5.5 cannot turn thinking off at all (both checked against the API on 2026-10-26). reasoning_effort none and an explicit disabled now map per model; on Opus it thinks adaptively at the lowest effort.
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars.
 |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Running Omni's agent loop against
/leanon production (58c58d6) turned up two translation bugs that break Claude turns on the Chat Completions route.Streaming fails on parallel tool calls. A streamed Haiku 5.5 turn with three parallel tool calls ended with
The upstream stream could not be translated.right after the third call started; the same request unstreamed succeeded. A tool call with empty arguments arrives with an emptyinput_json_delta(or none), and the continuation collector ranjson.loads("")at block stop. Empty arguments now become{}, and the stream translator logs the type of any exception it swallows, never its message.Sonnet and Opus 5.5 reject the thinking parameters we send.
reasoning_effortwith a largemax_tokensbecamethinking: {type: "enabled", budget_tokens}, which both answer 400. They now get adaptive thinking withoutput_config.effort, as Haiku 5.5 does, and client sampling parameters are dropped. Turning thinking off differs per model (checked against the API): Haiku acceptsdisabled, Sonnet needsbetween_tools, and Opus cannot turn it off, sononekeeps adaptive thinking at effortlowthere.Tests: regressions for both fail before and pass after; focused suites and the outbound inventory gate pass, and the full suite passed on each fix (5,403 and 5,478 passed) before they were combined.