fix: resolve silent agent termination on stream drop

[Varuu-0]

Context

This PR addresses a critical bug where Kilo agents would prematurely stop generating text mid-response without any error shown to the user. This silent failure occurs when the upstream Server-Sent Events (SSE) connection drops unexpectedly before a proper terminal event or finish_reason is received from the provider. Previously, the system defaulted to an "other" finish reason, which the agent loop interpreted as a successful completion, skipping the error handling and retry logic entirely.

fixes

#8476
#9544

Implementation

To fix this, the system must properly distinguish between a clean generation completion and a premature stream termination:

1. Provider Tracking: Updated openai-compatible-chat-language-model.ts and openai-responses-language-model.ts to actively track whether a terminal finish chunk or reason was successfully received. If the stream flushes prematurely, they now emit an unknown finish reason instead of defaulting to other.
2. Mapper Updates: Modified map-openai-compatible-finish-reason.ts and map-openai-responses-finish-reason.ts to map unrecognized states to unknown instead of other.
3. Triggering Retries: Updated the session processor (processor.ts) to intercept the unknown finish reason immediately after the stream drains. If no tool parts were executed, it throws a retryable MessageV2.APIError ("Stream terminated prematurely without a finish reason"), which triggers the existing Effect.retry pipeline.
4. Defense in Depth: Modified the agent loop condition in prompt.ts to explicitly exclude other as a valid completion signal, ensuring no ambiguous states can cause a silent loop exit.

Screenshots

File	Before	After
openai-compatible-chat-language-model.ts	typescript<br/>let finishReason = {<br/> unified: "other",<br/> raw: undefined<br/>}<br/>	typescript<br/>let finishReason = { ... }<br/>let receivedFinishReason = false;<br/><br/>/* in flush() */<br/>if (!receivedFinishReason && !isFirstChunk) {<br/> finishReason = { unified: "unknown", raw: undefined }<br/>}<br/>
openai-responses-language-model.ts	typescript<br/>let finishReason = {<br/> unified: "other",<br/> raw: undefined<br/>}<br/>	typescript<br/>let finishReason = { ... }<br/>let receivedFinishChunk = false;<br/><br/>/* in flush() */<br/>if (!receivedFinishChunk && responseId !== null) {<br/> finishReason = { unified: "unknown", raw: undefined }<br/>}<br/>
processor.ts	typescript<br/>yield* stream.pipe(<br/> Stream.tap(...),<br/> Stream.takeUntil(...),<br/> Stream.runDrain<br/>)<br/><br/><br/>	typescript<br/>yield* stream.pipe(...)<br/><br/>if (ctx.assistantMessage.finish === "unknown") {<br/> const parts = MessageV2.parts(ctx.assistantMessage.id)<br/> if (!parts.some(p => p.type === "tool")) {<br/> yield* Effect.fail(new MessageV2.APIError({<br/> message: "Stream terminated prematurely...",<br/> isRetryable: true<br/> }).toObject())<br/> }<br/>}<br/>
prompt.ts	typescript<br/>const finished = handle.message.finish &&<br/> !["tool-calls", "unknown"].includes(handle.message.finish)<br/>	typescript<br/>const finished = handle.message.finish &&<br/> !["tool-calls", "unknown", "other"].includes(handle.message.finish)<br/>
Mappers	typescript<br/>default:<br/> return "other"<br/>	typescript<br/>default:<br/> return "unknown"<br/>

How to Test

1. Start an agent generation that produces a long response.
2. Artificially simulate a network drop or prematurely kill the provider stream before it sends the final finish_reason chunk.
3. Observe that instead of the agent stopping its execution with a half-written response, it catches the interruption, logs a retryable API Error, and automatically retries the generation.

Get in Touch

Discord @varuu

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (7 files)

• packages/opencode/src/provider/sdk/copilot/chat/map-openai-compatible-finish-reason.ts
• packages/opencode/src/provider/sdk/copilot/chat/openai-compatible-chat-language-model.ts
• packages/opencode/src/provider/sdk/copilot/responses/map-openai-responses-finish-reason.ts
• packages/opencode/src/provider/sdk/copilot/responses/openai-responses-language-model.ts
• packages/opencode/src/session/message-v2.ts
• packages/opencode/src/session/processor.ts
• packages/opencode/src/session/prompt.ts

---

_{Reviewed by gpt-5.5-20260423 · 1,392,325 tokens}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 313b3ba3c0

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".