fix(runtime): persist provider-executed streamed tool results

[kojiwakayama]

Summary

• persist provider-executed streamed tool results into runtime conversation state
• forward streamed provider tool errors through the data stream protocol
• avoid re-entering streamed tool-call steps without matching provider tool-result messages
• bump version to 0.1.138 to trigger deployment

Testing

• deno test --allow-all src/agent/runtime/ai-stream-handler.test.ts
• deno check src/agent/runtime/index.ts
• full repo pre-push checks (format, lint, typecheck, unit tests)

Refs veryfront-studio#1239

[chatgpt-codex-connector]

💡 Codex Review

veryfront-code/src/agent/runtime/index.ts

Lines 785 to 786 in 4d02e6d

	currentMessages.push(toolResultMessage);
	await this.memory.add(toolResultMessage);

Keep final assistant text when storing streamed tool results

Appending each toolResultMessage to currentMessages before the finishReason check can make a tool message the final entry of the run, so the method returns an empty response.text (it is derived from the last message’s text parts). This occurs when a streamed step includes both assistant text and provider tool results but does not continue with "tool-calls", causing downstream onFinish consumers to lose the generated assistant text even though it was streamed.

---

veryfront-code/src/agent/runtime/ai-stream-handler.ts

Line 289 in 4d02e6d

errorText: stringifyToolError(part.error),

Preserve Error.message in streamed tool-error events

In the new tool-error path, errorText is built via stringifyToolError(part.error), which JSON-stringifies first; when part.error is an Error object this commonly becomes {} instead of the actual message. That means clients receive unusable tool-output-error payloads for thrown errors, defeating the purpose of forwarding provider tool errors.

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".