Add current local-runtime and inference wire contracts - #3
Conversation
Co-authored-by: Medulla <medulla@tinyhumans.ai>
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. |
|
Warning Review limit reached
On-demand reviews are free for the next 21 days. After that, they cost $0.25 per reviewed file. Or wait 8 minutes for your next included review. View limit detailsLimit details: You’ve used the included review currently available. Review configuration: ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (20)
Warning Your free Security trial is over. An organization admin can activate Security or dismiss this notice. Comment |
How this change flows0 changed behaviours across 2 relationships. 2 surrounding behaviours are shown (60 graph nodes walked). 49 further behaviours left out to keep the diagram readable. flowchart LR
n0["translate_request"]:::impacted
n1["translate_request_with"]:::impacted
n0 -->|calls| n1
n0 -->|tests| n1
classDef changed fill:#0d4429,stroke:#238636,color:#e6edf3
classDef impacted fill:#161b22,stroke:#6e7681,color:#c9d1d9
classDef flagged fill:#5a1e02,stroke:#d93f0b,color:#ffffff
classDef blocking fill:#67060c,stroke:#f85149,color:#ffffff
Green: changed behaviour. Grey: surrounding behaviour. Arrows name the call, use, implementation, or test relationship. Orange: has findings. Red: has a finding that blocks the merge. |
![tinysweeper[bot]](https://avatars.githubusercontent.com/in/4526585?s=60&v=4){kind=small}
![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: 70412ce592
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if degrade.native_tools { | ||
| self.native_tools_on_wire.store(false, Ordering::Relaxed); |
There was a problem hiding this comment.
When a model profile advertises tool calling but the server returns a 400 saying tools are unsupported, this latch makes the retry use prompt-guided <tool_call> output. The response paths at invoke and stream, however, only call prompt_tools::apply_to_response when self.profile.tool_calling is false; the profile remains true here, so the retry can succeed while returning the tool-call markup as ordinary assistant text instead of a normalized call. Base response parsing on this latch as well as the static profile.
AGENTS.md reference: AGENTS.md:L43-L45
Useful? React with 👍 / 👎.
| } else { | ||
| Vec::new() | ||
| }; | ||
| let tool_choice = (!tools.is_empty()).then(|| translate_tool_choice(&request.tool_choice)); |
There was a problem hiding this comment.
For ToolChoice::Tool, this reuses the Chat Completions translator and serializes {"type":"function","function":{"name":...}}. The Responses API uses the flattened {"type":"function","name":...} shape, which the removed responses_tool_choice helper previously produced, so named-tool requests on the Responses path are rejected instead of forcing the requested tool.
Useful? React with 👍 / 👎.
| ModelResponse { | ||
| message: AssistantMessage { | ||
| id: None, | ||
| content: vec![ContentBlock::Text(text)], | ||
| tool_calls, | ||
| content, | ||
| tool_calls: Vec::new(), | ||
| usage, |
There was a problem hiding this comment.
Whenever /responses returns a function_call output item, this now unconditionally emits an empty tool_calls list; the response wire struct also no longer retains the call id, name, or arguments. The previous parser normalized these items, including malformed arguments, so tool-using Responses calls now appear to consumers as empty assistant replies and the requested tool is never executed.
AGENTS.md reference: AGENTS.md:L43-L45
Useful? React with 👍 / 👎.
| Message::User(m) => message_text(&m.content), | ||
| Message::Assistant(m) => message_text(&m.content), |
There was a problem hiding this comment.
For a Responses-primary request containing ContentBlock::Image, message_text retains only text blocks, so this branch silently discards every image; an image-only user turn is skipped altogether at the subsequent empty-text check. The prior translation emitted input_image content parts, so vision requests now run against missing input and can return plausible but incorrect results rather than an error.
Useful? React with 👍 / 👎.
| } else { | ||
| "stop".to_string() | ||
| }), | ||
| finish_reason: Some("stop".to_string()), |
There was a problem hiding this comment.
Every Responses result is now marked stop, including responses whose status is incomplete because max_output_tokens was reached. Consumers therefore cannot distinguish truncated output from a completed answer and may accept or cache partial text/JSON; retain the response status and map incomplete_details.reason as the previous parser did.
AGENTS.md reference: AGENTS.md:L43-L45
Useful? React with 👍 / 👎.
| model_info | ||
| .iter() | ||
| .filter(|(key, _)| key.ends_with(".context_length") || key.as_str() == "context_length") | ||
| .filter_map(|(_, value)| value.as_u64()) | ||
| .filter(|value| *value > 0) | ||
| .min() |
There was a problem hiding this comment.
On multimodal Ollama models, model_info can contain both the language architecture window (for example gemma3.context_length) and a small projector window such as clip.context_length = 77. Taking the minimum advertises 77 tokens as the chat model's input capacity, causing capability checks or context compaction to reject or discard nearly every normal prompt; identify the main architecture's context field instead of minimizing unrelated components.
Useful? React with 👍 / 👎.
| if live.is_empty() { | ||
| return Err(Error::Validation( | ||
| "Ollama embedding batches must not contain blank inputs".into(), | ||
| )); | ||
| return Ok(vec![Vec::new(); texts.len()]); |
There was a problem hiding this comment.
For any nonzero-dimensional OllamaEmbeddingModel, an all-blank batch now returns one empty vector per input, violating EmbeddingModel's fixed-dimension contract. Passing such output through Retriever::index immediately fails in InMemoryVectorStore::add, while direct callers can accidentally persist invalid vectors elsewhere; return a validation error or valid vectors of dimensions() instead.
AGENTS.md reference: AGENTS.md:L53-L58
Useful? React with 👍 / 👎.
| let parsed: ResponsesResponse = | ||
| serde_json::from_value(value.clone()).unwrap_or_else(|_| ResponsesResponse { | ||
| output: Vec::new(), | ||
| output_text: None, | ||
| usage: None, | ||
| }); |
There was a problem hiding this comment.
If a successful HTTP response has an incompatible schema, such as {"output":"not-an-array"}, deserialization now falls back to an empty response and invoke_responses returns a successful blank assistant message. This hides provider incompatibilities and malformed payloads that previously surfaced as serialization errors, making failures indistinguishable from genuine empty completions; keep parsing fallible and propagate the decode error.
Useful? React with 👍 / 👎.
1 participant
Summary
Verification