Chat-completions turns stream: reasoning models no longer die at the 60s header wall
Root cause of the stuck kimi-for-coding kickoff (task running forever, no
team spawning): runKimiCodeTurn posted stream:false, and a non-streaming
completion returns headers only after the whole generation — once the CEO's
context grew past ~150k chars, a reasoning turn needed >60s and provider-http's
header deadline aborted it ("Provider request timed out."), failing the run;
every heartbeat retry re-read the workspace, regrew the context, and died at
the same wall until the 3-run budget was gone.
Fix at the family level: accumulateChatStream (chat-format.ts) consumes the
standard OpenAI chunk deltas — content appends, indexed tool_call fragments,
reasoning_content keeping the 60s idle deadline fed — and a shared
runChatCompletionsStream drives makeChatCompletionsTurn (kimi, kimi-cn, xai,
groq, mistral, cerebras, openrouter) and runKimiCodeTurn.
Wire shape verified against MoonshotAI/kimi-cli (kosong streams by default)
and a live streamed probe of api.kimi.com/coding/v1 with the stored plan
credential: role opener, reasoning deltas, tool_call fragments (id+name then
argument text), finish_reason, [DONE] — all reassembled correctly.
verify:ci green including e2e.
This commit is contained in:
1 parent
3a7b267ead
commit
0fa99d94a7
3 files changed
+104
-29
No files matched your search
@@ -8,7 +8,7 @@ import { postEventStream, postJson, type JsonRecord } from "../provider-http";
|
||||
import type { AgentTurnMessage, AgentTurnResult, ProviderToolTurnInput, TurnCredential } from "../provider-tool-turn";
|
||||
import { accumulateResponsesStream, parseResponsesOutput, responsesToolDefs, toResponsesInput } from "./responses-format";
|
||||
import { anthropicToolDefs, parseAnthropicContent, toAnthropicMessages } from "./anthropic-format";
|
||||
import { chatToolDefs, parseChatChoice, toChatMessages } from "./chat-format";
|
||||
import { accumulateChatStream, chatToolDefs, parseChatChoice, toChatMessages } from "./chat-format";
|
||||
import { accumulateGeminiStream, antigravityRequestBody, parseGeminiResponse, vertexRequestBody } from "./gemini-format";
|
||||
import { bedrockToolConfig, parseBedrockOutput, toBedrockMessages } from "./bedrock-format";
|
||||
import { signAwsRequest } from "./aws-sigv4";
|
||||
@@ -51,21 +51,29 @@ export async function runOpenAiResponsesTurn(input: ProviderToolTurnInput): Prom
|
||||
return { ...parsed, rawResponseJson: JSON.stringify(response.data) };
|
||||
}
|
||||
|
||||
// ---- OpenAI Chat Completions (xAI, Groq, Mistral, Cerebras, OpenRouter) ----
|
||||
// ---- OpenAI Chat Completions, streamed (xAI, Groq, Mistral, Cerebras, OpenRouter, Kimi, Kimi CN) ----
|
||||
|
||||
/**
|
||||
* Streaming is required, not an optimization: a non-streaming completion
|
||||
* sends its response headers only after the WHOLE generation finishes, so a
|
||||
* reasoning model that thinks past the 60s header deadline gets aborted with
|
||||
* "Provider request timed out." even though it is working. Streamed turns
|
||||
* emit chunks continuously and only a truly dead connection times out.
|
||||
*/
|
||||
async function runChatCompletionsStream(endpoint: string, headers: Record<string, string>, input: ProviderToolTurnInput): Promise<AgentTurnResult> {
|
||||
const body: JsonRecord = {
|
||||
model: input.model,
|
||||
messages: toChatMessages(input.instructions, input.messages),
|
||||
tools: chatToolDefs(input.tools),
|
||||
tool_choice: "auto",
|
||||
stream: true,
|
||||
};
|
||||
const response = await postEventStream(endpoint, headers, body);
|
||||
return accumulateChatStream(response.messages);
|
||||
}
|
||||
|
||||
export function makeChatCompletionsTurn(endpoint: string, label: string): (input: ProviderToolTurnInput) => Promise<AgentTurnResult> {
|
||||
return async (input) => {
|
||||
const body: JsonRecord = {
|
||||
model: input.model,
|
||||
messages: toChatMessages(input.instructions, input.messages),
|
||||
tools: chatToolDefs(input.tools),
|
||||
tool_choice: "auto",
|
||||
stream: false,
|
||||
};
|
||||
const response = await postJson(endpoint, { authorization: `Bearer ${requireApiKey(input.credential, label)}` }, body);
|
||||
const parsed = parseChatChoice(response.data);
|
||||
return { ...parsed, rawResponseJson: JSON.stringify(response.data) };
|
||||
};
|
||||
return async (input) => runChatCompletionsStream(endpoint, { authorization: `Bearer ${requireApiKey(input.credential, label)}` }, input);
|
||||
}
|
||||
|
||||
// ---- Anthropic Messages (api key) ----
|
||||
@@ -109,20 +117,11 @@ export async function runClaudeProTurn(input: ProviderToolTurnInput): Promise<Ag
|
||||
return { ...parsed, rawResponseJson: JSON.stringify(response.data) };
|
||||
}
|
||||
|
||||
// ---- Kimi Code (OAuth plan, chat completions at the coding endpoint) ----
|
||||
// ---- Kimi Code (OAuth plan, streamed chat completions at the coding endpoint) ----
|
||||
|
||||
export async function runKimiCodeTurn(input: ProviderToolTurnInput): Promise<AgentTurnResult> {
|
||||
const access = requireAccessToken(input.credential, "Kimi Code");
|
||||
const body: JsonRecord = {
|
||||
model: input.model,
|
||||
messages: toChatMessages(input.instructions, input.messages),
|
||||
tools: chatToolDefs(input.tools),
|
||||
tool_choice: "auto",
|
||||
stream: false,
|
||||
};
|
||||
const response = await postJson(`${KIMI_CODE_BASE_URL}/chat/completions`, { authorization: `Bearer ${access}` }, body);
|
||||
const parsed = parseChatChoice(response.data);
|
||||
return { ...parsed, rawResponseJson: JSON.stringify(response.data) };
|
||||
return runChatCompletionsStream(`${KIMI_CODE_BASE_URL}/chat/completions`, { authorization: `Bearer ${access}` }, input);
|
||||
}
|
||||
|
||||
// ---- GPT account / OpenAI Codex (OAuth, SSE) ----
|
||||
|
||||
Reference in new issue
Block a user