fix: prevent Ollama provider from hanging on tool-calling requests#7723
Merged
DOsinga merged 6 commits intoblock:mainfrom Mar 26, 2026
Merged
fix: prevent Ollama provider from hanging on tool-calling requests#7723DOsinga merged 6 commits intoblock:mainfrom
DOsinga merged 6 commits intoblock:mainfrom
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 87518ea85a
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a59f743a52
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
DOsinga
approved these changes
Mar 24, 2026
Collaborator
DOsinga
left a comment
DOsinga
left a comment
There was a problem hiding this comment.
LGTM — clean fix for the Ollama hang. The per-chunk timeout correctly skips the first token (avoiding false positives on slow models), and the apply_ollama_options changes properly strip stream_options and remap max_completion_tokens → options.num_predict. Tests are meaningful and cover the actual bug.
michaelneale
approved these changes
Mar 25, 2026
Collaborator
michaelneale
left a comment
michaelneale
left a comment
There was a problem hiding this comment.
yep - looks good. Do you mind resolving merge and updating - and we can get this in @fresh3nough
Contributor
Author
|
Not at all I'll take care of it today |
Strip unsupported OpenAI-specific fields (stream_options, max_completion_tokens) from the request payload sent to Ollama's /v1/chat/completions endpoint. Convert max_completion_tokens to Ollama's native options.num_predict. Add a 30-second per-chunk streaming timeout so stalled Ollama connections fail fast instead of hanging for the full 600s timeout. Fixes: block#7715 Signed-off-by: fre <[email protected]>
Move the tokio::time::timeout from wrapping message_stream.next() (the output of response_to_streaming_message_ollama) to wrapping the raw FramedRead/LinesCodec line reads via a new with_line_timeout helper. This prevents false stalls during long tool-call generations where response_to_streaming_message_ollama buffers XML content internally without yielding messages, while still timing out dead connections. Signed-off-by: fresh3nough <[email protected]> Signed-off-by: fre <[email protected]>
Update with_line_timeout so the first SSE line is read without the 30s per-chunk timeout, letting time-to-first-token be governed by the request timeout. After the first line arrives, apply tokio::time::timeout to subsequent line reads to preserve dead-connection stall detection. Signed-off-by: fresh3nough <[email protected]> Signed-off-by: fre <[email protected]>
fresh3nough
force-pushed
the
fix/ollama-hang-7715
branch
from
19aec34 to
d483e82
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d483e8269f
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
…options Non-reasoning models emit max_tokens while reasoning models emit max_completion_tokens. Both are unsupported by Ollama and should be converted to options.num_predict. Update tests to match the actual create_request output for non-reasoning models like llama3.1. Signed-off-by: fre <[email protected]>
Signed-off-by: fre <[email protected]>
Contributor
Author
|
should be good now @DOsinga @michaelneale |
github-merge-queue
bot
removed this pull request from the merge queue due to a conflict with the base branch
Signed-off-by: Douwe Osinga <[email protected]>
michaelneale
added a commit
that referenced
this pull request
Mar 26, 2026
* main: (337 commits) fix: replace panics with user-friendly errors in CLI session builder (#7901) fix: read GOOSE_CONTEXT_LIMIT from config.yaml, not just env vars (#7900) fix: deliver truncation notice as separate content block (#7899) fix: use platform-appropriate commands in developer extension instructions (#7898) fix: replace any with proper SVG types in icon components (#7873) chore: remove debug console.log statements, stale comments, and dead code (#8142) feat: Gemini OAuth provider (#8129) chore(deps): bump picomatch from 2.3.1 to 2.3.2 in /documentation (#8123) feat: show installed skills in UI (#7910) fix(deps): gate keyring platform features behind target-specific deps (#8039) chore(deps): bump yaml from 2.8.2 to 2.8.3 in /evals/open-model-gym/suite (#8124) fix: strip message wrapper in CLI session title generation (#7996) fix(providers): fall back to configured models when models endpoint fetch fails (#7530) chore(deps): bump brace-expansion from 5.0.3 to 5.0.5 in /evals/open-model-gym/suite (#8139) fix: prevent Ollama provider from hanging on tool-calling requests (#7723) fix: VMware Tanzu Platform provider - bug fixes, streaming, UI improvements (#8126) feat: allow GOOSE_CLI_SHOW_THINKING to be set in config.yaml (#8097) fix: GitHub Copilot auth fails to open browser in Desktop app (#6957) (#8019) fix(ci): produce .tar.gz archives for Zed ACP registry compatibility (#8054) feat: add GOOSE_SHOW_FULL_OUTPUT config to disable tool output truncation (#7919) ... # Conflicts: # crates/goose/src/providers/formats/openai.rs
hydrosquall
pushed a commit
to hydrosquall/goose
that referenced
this pull request
Mar 31, 2026
…lock#7723) Signed-off-by: fre <[email protected]> Signed-off-by: fresh3nough <[email protected]> Signed-off-by: Douwe Osinga <[email protected]> Co-authored-by: Douwe Osinga <[email protected]> Signed-off-by: Cameron Yick <[email protected]>
blackgirlbytes
pushed a commit
that referenced
this pull request
Apr 2, 2026
* main: (337 commits) fix: replace panics with user-friendly errors in CLI session builder (#7901) fix: read GOOSE_CONTEXT_LIMIT from config.yaml, not just env vars (#7900) fix: deliver truncation notice as separate content block (#7899) fix: use platform-appropriate commands in developer extension instructions (#7898) fix: replace any with proper SVG types in icon components (#7873) chore: remove debug console.log statements, stale comments, and dead code (#8142) feat: Gemini OAuth provider (#8129) chore(deps): bump picomatch from 2.3.1 to 2.3.2 in /documentation (#8123) feat: show installed skills in UI (#7910) fix(deps): gate keyring platform features behind target-specific deps (#8039) chore(deps): bump yaml from 2.8.2 to 2.8.3 in /evals/open-model-gym/suite (#8124) fix: strip message wrapper in CLI session title generation (#7996) fix(providers): fall back to configured models when models endpoint fetch fails (#7530) chore(deps): bump brace-expansion from 5.0.3 to 5.0.5 in /evals/open-model-gym/suite (#8139) fix: prevent Ollama provider from hanging on tool-calling requests (#7723) fix: VMware Tanzu Platform provider - bug fixes, streaming, UI improvements (#8126) feat: allow GOOSE_CLI_SHOW_THINKING to be set in config.yaml (#8097) fix: GitHub Copilot auth fails to open browser in Desktop app (#6957) (#8019) fix(ci): produce .tar.gz archives for Zed ACP registry compatibility (#8054) feat: add GOOSE_SHOW_FULL_OUTPUT config to disable tool output truncation (#7919) ... # Conflicts: # crates/goose/src/providers/formats/openai.rs
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
3 participants
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.
Summary
Strip unsupported OpenAI-specific fields (
stream_options,max_completion_tokens) from the request payload sent to Ollama's/v1/chat/completionsendpoint. Convertmax_completion_tokensto Ollama's nativeoptions.num_predict. Add a 30-second per-chunk streaming timeout so stalled Ollama connections fail fast instead of hanging for the full 600s timeout.Problem
When goose sends a chat completion request to Ollama (especially with tool definitions), Ollama hangs indefinitely at max CPU. Direct
curlandollama runwork fine. The root cause is that goose's Ollama provider reuses the generic OpenAIcreate_requestwhich emits fields Ollama does not support:stream_options: {"include_usage": true}-- not recognized by Ollamamax_completion_tokens-- should benum_predictinside Ollama'soptionsobjectChanges
apply_ollama_options: Now stripsstream_optionsand convertsmax_completion_tokenstooptions.num_predictbefore sending to Ollamastream_ollama: Added a 30-second per-chunk timeout so stalled streams produce a clear error instead of hanging silentlycreate_requestproduces the problematic fieldsapply_ollama_optionsremoves them correctlyFixes #7715