Greptile Summary

This PR adds two defensive guards for non-standard Mistral (and similar provider) response shapes: a typeof === \"string\" check before appending choice.delta.content to the text buffer in the OpenAI transport stream, and object-form thinking-block support (item.thinking.text) in extractThinking. Both fixes are minimal, well-targeted, and preserve existing behavior for Anthropic/OpenAI responses.

Confidence Score: 5/5

Safe to merge — changes are narrow defensive guards with no regression risk on existing providers.

Both changes are simple, targeted type-guards that prevent silent coercion bugs. The only finding is a P2 style nit about if (thinkingValue) vs if (thinkingValue !== null) that has no practical impact since Mistral is unlikely to send an empty thinking text.

No files require special attention.

This is a comment left during a code review.
Path: ui/src/ui/chat/message-extract.ts
Line: 60

Comment:
**Empty string thinking silently dropped**

`if (thinkingValue)` is falsy for `""`, so a thinking block whose `.text` property is an empty string is silently discarded. This is likely fine in practice, but if `thinkingValue !== null` is the intended gate (to distinguish "present but empty" from "absent"), the check should be explicit.

```suggestion
        if (thinkingValue !== null) {
```

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "fix: guard against non-string content de..." | Re-trigger Greptile}

Related work from PRtags group golden-werewolf-3vn3

Title: Open PR duplicate: [Bug]: A problem with mistral small thinking enabled

* This PR

Closing this as duplicate or superseded after Codex automated review.

PR #68855 should close as superseded by open PR #68418. Current main still lacks the fix, but #68418 is the broader canonical parser-side fix for the same #68309 Mistral Small thinking bug, while #68855 only skips non-string content and adds UI extraction for one object-shaped thinking form.

Best possible solution:

Close #68855 as superseded and focus maintainer review on #68418. The right product/code path is to fix processOpenAICompletionsStream by unpacking array-shaped delta.content into text and thinking deltas, preserving string behavior, and landing regression coverage for the Mistral Small thinking shape.

What I checked:

• Current main still appends truthy delta.content as text: processOpenAICompletionsStream still checks if (choice.delta.content) and passes that value directly to the text path, with no string guard or typed-block unpacker on current main. (src/agents/openai-transport-stream.ts:1335, d54d2d6b9b8a)
• Current main has no typed-block content unpacker: The completions reasoning helper only handles reasoning_details and fallback reasoning fields; repository search found no unpackOpenAICompletionsContent implementation on main. (src/agents/openai-transport-stream.ts:1441, d54d2d6b9b8a)
• Current UI extractor only supports string thinking blocks: extractThinking appends thinking content only when item.thinking is a string, so the object-form thinking extraction from fix: guard against non-string content delta and thinking blocks #68855 is not present on main. (ui/src/ui/chat/message-extract.ts:53, d54d2d6b9b8a)
• Existing nearby tests cover reasoning_details, not typed-block arrays: The nearby regression coverage exercises OpenRouter/Qwen3 reasoning_details; search found no current-main regression test for Mistral-style typed-block delta.content arrays or [object Object] coercion in this stream parser. (src/agents/openai-transport-stream.test.ts:2514, d54d2d6b9b8a)
• Related PR fix(agents): unpack typed-block delta.content arrays in openai-completions stream #68418 is the canonical duplicate: Provided GitHub context shows fix(agents): unpack typed-block delta.content arrays in openai-completions stream #68418 is open, also closes [Bug]: A problem with mistral small thinking enabled #68309, describes the same Mistral Small thinking [object Object] symptom, adds parser-side typed-block delta.content unpacking, and includes a regression test plan. The PRtags bot grouped fix(agents): unpack typed-block delta.content arrays in openai-completions stream #68418 and fix: guard against non-string content delta and thinking blocks #68855 as duplicate work. (06b26221735b)
• PR fix: guard against non-string content delta and thinking blocks #68855 is narrower than the canonical fix: The fix: guard against non-string content delta and thinking blocks #68855 diff only adds a typeof choice.delta.content === "string" guard and UI object-thinking extraction. For array-shaped typed blocks, that avoids coercion but drops the real text/thinking instead of preserving it in the parser. (8c399d49f20a)

So I’m closing this here and keeping the remaining discussion on the canonical linked item.

Codex Review notes: model gpt-5.5, reasoning high; reviewed against d54d2d6b9b8a.