@ubiquity-os/plugin-sdk 3.8.4 → 3.10.0

package/dist/llm.d.mts

@@ -1,5 +1,6 @@
-import { ChatCompletionMessageParam, ChatCompletion, ChatCompletionChunk } from 'openai/resources/chat/completions';
-import { C as Context } from './context-sqbr2o6i.mjs';
+import { ChatCompletionMessageParam, ChatCompletionCreateParamsNonStreaming, ChatCompletion, ChatCompletionChunk } from 'openai/resources/chat/completions';
+import { C as Context } from './context-Dwl3aRX-.mjs';
+import { PluginInput } from './signature.mjs';
 import '@octokit/webhooks';
 import '@ubiquity-os/ubiquity-os-logger';
 import '@octokit/plugin-rest-endpoint-methods';
@@ -9,52 +10,15 @@ import '@octokit/plugin-paginate-graphql';
 import '@octokit/plugin-paginate-rest';
 import '@octokit/webhooks/node_modules/@octokit/request-error';
 import '@octokit/core';
+import '@sinclair/typebox';
 
-type LlmResponseFormat = {
-    type: "json_object" | "text";
-} | {
-    type: string;
-    [key: string]: unknown;
-};
-type LlmPayload = {
-    repository?: {
-        owner?: {
-            login?: string;
-        };
-        name?: string;
-    };
-    installation?: {
-        id?: number;
-    };
-};
-type LlmAuthContext = {
-    authToken?: string;
-    ubiquityKernelToken?: string;
-    payload?: LlmPayload;
-    eventPayload?: LlmPayload;
-};
 type LlmCallOptions = {
     baseUrl?: string;
     model?: string;
     stream?: boolean;
     messages: ChatCompletionMessageParam[];
-    max_tokens?: number;
-    max_completion_tokens?: number;
-    temperature?: number;
-    top_p?: number;
-    frequency_penalty?: number;
-    presence_penalty?: number;
-    response_format?: LlmResponseFormat;
-    stop?: string | string[];
-    n?: number;
-    logit_bias?: Record<string, number>;
-    seed?: number;
-    user?: string;
-    metadata?: Record<string, unknown>;
-    tools?: unknown[];
-    tool_choice?: string | Record<string, unknown>;
-    [key: string]: unknown;
-};
-declare function callLlm(options: LlmCallOptions, input: Context | LlmAuthContext): Promise<ChatCompletion | AsyncIterable<ChatCompletionChunk>>;
+    aiAuthToken?: string;
+} & Partial<Omit<ChatCompletionCreateParamsNonStreaming, "model" | "messages" | "stream">>;
+declare function callLlm(options: LlmCallOptions, input: PluginInput | Context): Promise<ChatCompletion | AsyncIterable<ChatCompletionChunk>>;
 
 export { type LlmCallOptions, callLlm };

package/dist/llm.d.ts

@@ -1,5 +1,6 @@
-import { ChatCompletionMessageParam, ChatCompletion, ChatCompletionChunk } from 'openai/resources/chat/completions';
-import { C as Context } from './context-BbEmsEct.js';
+import { ChatCompletionMessageParam, ChatCompletionCreateParamsNonStreaming, ChatCompletion, ChatCompletionChunk } from 'openai/resources/chat/completions';
+import { C as Context } from './context-zLHgu52i.js';
+import { PluginInput } from './signature.js';
 import '@octokit/webhooks';
 import '@ubiquity-os/ubiquity-os-logger';
 import '@octokit/plugin-rest-endpoint-methods';
@@ -9,52 +10,15 @@ import '@octokit/plugin-paginate-graphql';
 import '@octokit/plugin-paginate-rest';
 import '@octokit/webhooks/node_modules/@octokit/request-error';
 import '@octokit/core';
+import '@sinclair/typebox';
 
-type LlmResponseFormat = {
-    type: "json_object" | "text";
-} | {
-    type: string;
-    [key: string]: unknown;
-};
-type LlmPayload = {
-    repository?: {
-        owner?: {
-            login?: string;
-        };
-        name?: string;
-    };
-    installation?: {
-        id?: number;
-    };
-};
-type LlmAuthContext = {
-    authToken?: string;
-    ubiquityKernelToken?: string;
-    payload?: LlmPayload;
-    eventPayload?: LlmPayload;
-};
 type LlmCallOptions = {
     baseUrl?: string;
     model?: string;
     stream?: boolean;
     messages: ChatCompletionMessageParam[];
-    max_tokens?: number;
-    max_completion_tokens?: number;
-    temperature?: number;
-    top_p?: number;
-    frequency_penalty?: number;
-    presence_penalty?: number;
-    response_format?: LlmResponseFormat;
-    stop?: string | string[];
-    n?: number;
-    logit_bias?: Record<string, number>;
-    seed?: number;
-    user?: string;
-    metadata?: Record<string, unknown>;
-    tools?: unknown[];
-    tool_choice?: string | Record<string, unknown>;
-    [key: string]: unknown;
-};
-declare function callLlm(options: LlmCallOptions, input: Context | LlmAuthContext): Promise<ChatCompletion | AsyncIterable<ChatCompletionChunk>>;
+    aiAuthToken?: string;
+} & Partial<Omit<ChatCompletionCreateParamsNonStreaming, "model" | "messages" | "stream">>;
+declare function callLlm(options: LlmCallOptions, input: PluginInput | Context): Promise<ChatCompletion | AsyncIterable<ChatCompletionChunk>>;
 
 export { type LlmCallOptions, callLlm };

package/dist/llm.js

@@ -23,6 +23,7 @@ __export(llm_exports, {
   callLlm: () => callLlm
 });
 module.exports = __toCommonJS(llm_exports);
+var EMPTY_STRING = "";
 function normalizeBaseUrl(baseUrl) {
   let normalized = baseUrl.trim();
   while (normalized.endsWith("/")) {
@@ -30,87 +31,218 @@ function normalizeBaseUrl(baseUrl) {
   }
   return normalized;
 }
+var MAX_LLM_RETRIES = 2;
+var RETRY_BACKOFF_MS = [250, 750];
+function getRetryDelayMs(attempt) {
+  return RETRY_BACKOFF_MS[Math.min(attempt, RETRY_BACKOFF_MS.length - 1)] ?? 750;
+}
+function sleep(ms) {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}
 function getEnvString(name) {
-  if (typeof process === "undefined" || !process?.env) return "";
-  return String(process.env[name] ?? "").trim();
+  if (typeof process === "undefined" || !process?.env) return EMPTY_STRING;
+  return String(process.env[name] ?? EMPTY_STRING).trim();
+}
+function normalizeToken(value) {
+  return typeof value === "string" ? value.trim() : EMPTY_STRING;
+}
+function isGitHubToken(token) {
+  return token.trim().startsWith("gh");
+}
+function getEnvTokenFromInput(input) {
+  if ("env" in input) {
+    const envValue = input.env;
+    if (envValue && typeof envValue === "object") {
+      const token = normalizeToken(envValue.UOS_AI_TOKEN);
+      if (token) return token;
+    }
+  }
+  return getEnvString("UOS_AI_TOKEN");
+}
+function resolveAuthToken(input, aiAuthToken) {
+  const explicit = normalizeToken(aiAuthToken);
+  if (explicit) return { token: explicit, isGitHub: isGitHubToken(explicit) };
+  const envToken = getEnvTokenFromInput(input);
+  if (envToken) return { token: envToken, isGitHub: isGitHubToken(envToken) };
+  const fallback = normalizeToken(input.authToken);
+  if (!fallback) {
+    const err = new Error("Missing auth token; set UOS_AI_TOKEN, pass aiAuthToken, or provide authToken in input");
+    err.status = 401;
+    throw err;
+  }
+  return { token: fallback, isGitHub: isGitHubToken(fallback) };
 }
 function getAiBaseUrl(options) {
   if (typeof options.baseUrl === "string" && options.baseUrl.trim()) {
     return normalizeBaseUrl(options.baseUrl);
   }
-  const envBaseUrl = getEnvString("UBQ_AI_BASE_URL") || getEnvString("UBQ_AI_URL");
+  const envBaseUrl = getEnvString("UOS_AI_URL");
   if (envBaseUrl) return normalizeBaseUrl(envBaseUrl);
-  return "https://ai.ubq.fi";
+  return "https://ai-ubq-fi.deno.dev";
 }
 async function callLlm(options, input) {
-  const inputPayload = input;
-  const authToken = inputPayload.authToken;
-  const ubiquityKernelToken = inputPayload.ubiquityKernelToken;
-  const payload = inputPayload.payload ?? inputPayload.eventPayload;
-  const owner = payload?.repository?.owner?.login ?? "";
-  const repo = payload?.repository?.name ?? "";
-  const installationId = payload?.installation?.id;
-  if (!authToken) throw new Error("Missing authToken in inputs");
-  const isKernelTokenRequired = authToken.trim().startsWith("gh");
-  if (isKernelTokenRequired && !ubiquityKernelToken) {
-    throw new Error("Missing ubiquityKernelToken in inputs (kernel attestation is required for GitHub auth)");
-  }
-  const { baseUrl, model, stream: isStream, messages, ...rest } = options;
-  const url = `${getAiBaseUrl({ ...options, baseUrl })}/v1/chat/completions`;
+  const { baseUrl, model, stream: isStream, messages, aiAuthToken, ...rest } = options;
+  const { token: authToken, isGitHub } = resolveAuthToken(input, aiAuthToken);
+  const kernelToken = "ubiquityKernelToken" in input ? input.ubiquityKernelToken : void 0;
+  const payload = getPayload(input);
+  const { owner, repo, installationId } = getRepoMetadata(payload);
+  ensureMessages(messages);
+  const url = buildAiUrl(options, baseUrl);
   const body = JSON.stringify({
     ...rest,
     ...model ? { model } : {},
     messages,
     stream: isStream ?? false
   });
-  const headers = {
-    Authorization: `Bearer ${authToken}`,
-    "Content-Type": "application/json"
-  };
-  if (owner) headers["X-GitHub-Owner"] = owner;
-  if (repo) headers["X-GitHub-Repo"] = repo;
-  if (typeof installationId === "number" && Number.isFinite(installationId)) {
-    headers["X-GitHub-Installation-Id"] = String(installationId);
-  }
-  if (ubiquityKernelToken) {
-    headers["X-Ubiquity-Kernel-Token"] = ubiquityKernelToken;
-  }
-  const response = await fetch(url, { method: "POST", headers, body });
-  if (!response.ok) {
-    const err = await response.text();
-    throw new Error(`LLM API error: ${response.status} - ${err}`);
-  }
+  const headers = buildHeaders(authToken, {
+    owner,
+    repo,
+    installationId,
+    ubiquityKernelToken: isGitHub ? kernelToken : void 0
+  });
+  const response = await fetchWithRetry(url, { method: "POST", headers, body }, MAX_LLM_RETRIES);
   if (isStream) {
     if (!response.body) {
       throw new Error("LLM API error: missing response body for streaming request");
     }
     return parseSseStream(response.body);
   }
-  return response.json();
+  const rawText = await response.text();
+  try {
+    return JSON.parse(rawText);
+  } catch (err) {
+    const preview = rawText ? rawText.slice(0, 1e3) : EMPTY_STRING;
+    const message = "LLM API error: failed to parse JSON response from server" + (preview ? `; response body (truncated): ${preview}` : EMPTY_STRING);
+    const error = new Error(message);
+    error.cause = err;
+    error.status = response.status;
+    throw error;
+  }
+}
+function ensureMessages(messages) {
+  if (!Array.isArray(messages) || messages.length === 0) {
+    const err = new Error("messages must be a non-empty array");
+    err.status = 400;
+    throw err;
+  }
+}
+function buildAiUrl(options, baseUrl) {
+  return `${getAiBaseUrl({ ...options, baseUrl })}/v1/chat/completions`;
+}
+async function fetchWithRetry(url, options, maxRetries) {
+  let attempt = 0;
+  let lastError;
+  while (attempt <= maxRetries) {
+    try {
+      const response = await fetch(url, options);
+      if (response.ok) return response;
+      throw await buildResponseError(response);
+    } catch (error) {
+      lastError = error;
+      if (!shouldRetryError(error, attempt, maxRetries)) throw error;
+      await sleep(getRetryDelayMs(attempt));
+      attempt += 1;
+    }
+  }
+  throw lastError ?? new Error("LLM API error: request failed after retries");
+}
+async function buildResponseError(response) {
+  const errText = await response.text();
+  const error = new Error(`LLM API error: ${response.status} - ${errText}`);
+  error.status = response.status;
+  return error;
+}
+function shouldRetryError(error, attempt, maxRetries) {
+  if (attempt >= maxRetries) return false;
+  const status = getErrorStatus(error);
+  if (typeof status === "number") {
+    return status >= 500;
+  }
+  return true;
+}
+function getErrorStatus(error) {
+  return typeof error?.status === "number" ? error.status : void 0;
+}
+function getPayload(input) {
+  if ("payload" in input) {
+    return input.payload;
+  }
+  return input.eventPayload;
+}
+function getRepoMetadata(payload) {
+  const repoPayload = payload;
+  return {
+    owner: repoPayload?.repository?.owner?.login ?? EMPTY_STRING,
+    repo: repoPayload?.repository?.name ?? EMPTY_STRING,
+    installationId: repoPayload?.installation?.id
+  };
+}
+function buildHeaders(authToken, options) {
+  const headers = {
+    Authorization: `Bearer ${authToken}`,
+    "Content-Type": "application/json"
+  };
+  if (options.owner) headers["X-GitHub-Owner"] = options.owner;
+  if (options.repo) headers["X-GitHub-Repo"] = options.repo;
+  if (typeof options.installationId === "number" && Number.isFinite(options.installationId)) {
+    headers["X-GitHub-Installation-Id"] = String(options.installationId);
+  }
+  if (options.ubiquityKernelToken) {
+    headers["X-Ubiquity-Kernel-Token"] = options.ubiquityKernelToken;
+  }
+  return headers;
 }
 async function* parseSseStream(body) {
   const reader = body.getReader();
   const decoder = new TextDecoder();
-  let buffer = "";
+  let buffer = EMPTY_STRING;
   try {
     while (true) {
       const { value, done: isDone } = await reader.read();
       if (isDone) break;
       buffer += decoder.decode(value, { stream: true });
-      const events = buffer.split("\n\n");
-      buffer = events.pop() || "";
+      const { events, remainder } = splitSseEvents(buffer);
+      buffer = remainder;
       for (const event of events) {
-        if (event.startsWith("data: ")) {
-          const data = event.slice(6);
-          if (data === "[DONE]") return;
-          yield JSON.parse(data);
-        }
+        const data = getEventData(event);
+        if (!data) continue;
+        if (data.trim() === "[DONE]") return;
+        yield parseEventData(data);
       }
     }
   } finally {
     reader.releaseLock();
   }
 }
+function splitSseEvents(buffer) {
+  const normalized = buffer.replace(/\r\n/g, "\n").replace(/\r/g, "\n");
+  const parts = normalized.split("\n\n");
+  const remainder = parts.pop() ?? EMPTY_STRING;
+  return { events: parts, remainder };
+}
+function getEventData(event) {
+  if (!event.trim()) return null;
+  const dataLines = event.split("\n").filter((line) => line.startsWith("data:"));
+  if (!dataLines.length) return null;
+  const data = dataLines.map((line) => line.startsWith("data: ") ? line.slice(6) : line.slice(5).replace(/^ /, EMPTY_STRING)).join("\n");
+  return data || null;
+}
+function parseEventData(data) {
+  try {
+    return JSON.parse(data);
+  } catch (error) {
+    if (data.includes("\n")) {
+      const collapsed = data.replace(/\n/g, EMPTY_STRING);
+      try {
+        return JSON.parse(collapsed);
+      } catch {
+      }
+    }
+    const message = error instanceof Error ? error.message : String(error);
+    const preview = data.length > 200 ? `${data.slice(0, 200)}...` : data;
+    throw new Error(`LLM stream parse error: ${message}. Data: ${preview}`);
+  }
+}
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
   callLlm

package/dist/llm.mjs

@@ -1,4 +1,5 @@
 // src/llm/index.ts
+var EMPTY_STRING = "";
 function normalizeBaseUrl(baseUrl) {
   let normalized = baseUrl.trim();
   while (normalized.endsWith("/")) {
@@ -6,87 +7,218 @@ function normalizeBaseUrl(baseUrl) {
   }
   return normalized;
 }
+var MAX_LLM_RETRIES = 2;
+var RETRY_BACKOFF_MS = [250, 750];
+function getRetryDelayMs(attempt) {
+  return RETRY_BACKOFF_MS[Math.min(attempt, RETRY_BACKOFF_MS.length - 1)] ?? 750;
+}
+function sleep(ms) {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}
 function getEnvString(name) {
-  if (typeof process === "undefined" || !process?.env) return "";
-  return String(process.env[name] ?? "").trim();
+  if (typeof process === "undefined" || !process?.env) return EMPTY_STRING;
+  return String(process.env[name] ?? EMPTY_STRING).trim();
+}
+function normalizeToken(value) {
+  return typeof value === "string" ? value.trim() : EMPTY_STRING;
+}
+function isGitHubToken(token) {
+  return token.trim().startsWith("gh");
+}
+function getEnvTokenFromInput(input) {
+  if ("env" in input) {
+    const envValue = input.env;
+    if (envValue && typeof envValue === "object") {
+      const token = normalizeToken(envValue.UOS_AI_TOKEN);
+      if (token) return token;
+    }
+  }
+  return getEnvString("UOS_AI_TOKEN");
+}
+function resolveAuthToken(input, aiAuthToken) {
+  const explicit = normalizeToken(aiAuthToken);
+  if (explicit) return { token: explicit, isGitHub: isGitHubToken(explicit) };
+  const envToken = getEnvTokenFromInput(input);
+  if (envToken) return { token: envToken, isGitHub: isGitHubToken(envToken) };
+  const fallback = normalizeToken(input.authToken);
+  if (!fallback) {
+    const err = new Error("Missing auth token; set UOS_AI_TOKEN, pass aiAuthToken, or provide authToken in input");
+    err.status = 401;
+    throw err;
+  }
+  return { token: fallback, isGitHub: isGitHubToken(fallback) };
 }
 function getAiBaseUrl(options) {
   if (typeof options.baseUrl === "string" && options.baseUrl.trim()) {
     return normalizeBaseUrl(options.baseUrl);
   }
-  const envBaseUrl = getEnvString("UBQ_AI_BASE_URL") || getEnvString("UBQ_AI_URL");
+  const envBaseUrl = getEnvString("UOS_AI_URL");
   if (envBaseUrl) return normalizeBaseUrl(envBaseUrl);
-  return "https://ai.ubq.fi";
+  return "https://ai-ubq-fi.deno.dev";
 }
 async function callLlm(options, input) {
-  const inputPayload = input;
-  const authToken = inputPayload.authToken;
-  const ubiquityKernelToken = inputPayload.ubiquityKernelToken;
-  const payload = inputPayload.payload ?? inputPayload.eventPayload;
-  const owner = payload?.repository?.owner?.login ?? "";
-  const repo = payload?.repository?.name ?? "";
-  const installationId = payload?.installation?.id;
-  if (!authToken) throw new Error("Missing authToken in inputs");
-  const isKernelTokenRequired = authToken.trim().startsWith("gh");
-  if (isKernelTokenRequired && !ubiquityKernelToken) {
-    throw new Error("Missing ubiquityKernelToken in inputs (kernel attestation is required for GitHub auth)");
-  }
-  const { baseUrl, model, stream: isStream, messages, ...rest } = options;
-  const url = `${getAiBaseUrl({ ...options, baseUrl })}/v1/chat/completions`;
+  const { baseUrl, model, stream: isStream, messages, aiAuthToken, ...rest } = options;
+  const { token: authToken, isGitHub } = resolveAuthToken(input, aiAuthToken);
+  const kernelToken = "ubiquityKernelToken" in input ? input.ubiquityKernelToken : void 0;
+  const payload = getPayload(input);
+  const { owner, repo, installationId } = getRepoMetadata(payload);
+  ensureMessages(messages);
+  const url = buildAiUrl(options, baseUrl);
   const body = JSON.stringify({
     ...rest,
     ...model ? { model } : {},
     messages,
     stream: isStream ?? false
   });
-  const headers = {
-    Authorization: `Bearer ${authToken}`,
-    "Content-Type": "application/json"
-  };
-  if (owner) headers["X-GitHub-Owner"] = owner;
-  if (repo) headers["X-GitHub-Repo"] = repo;
-  if (typeof installationId === "number" && Number.isFinite(installationId)) {
-    headers["X-GitHub-Installation-Id"] = String(installationId);
-  }
-  if (ubiquityKernelToken) {
-    headers["X-Ubiquity-Kernel-Token"] = ubiquityKernelToken;
-  }
-  const response = await fetch(url, { method: "POST", headers, body });
-  if (!response.ok) {
-    const err = await response.text();
-    throw new Error(`LLM API error: ${response.status} - ${err}`);
-  }
+  const headers = buildHeaders(authToken, {
+    owner,
+    repo,
+    installationId,
+    ubiquityKernelToken: isGitHub ? kernelToken : void 0
+  });
+  const response = await fetchWithRetry(url, { method: "POST", headers, body }, MAX_LLM_RETRIES);
   if (isStream) {
     if (!response.body) {
       throw new Error("LLM API error: missing response body for streaming request");
     }
     return parseSseStream(response.body);
   }
-  return response.json();
+  const rawText = await response.text();
+  try {
+    return JSON.parse(rawText);
+  } catch (err) {
+    const preview = rawText ? rawText.slice(0, 1e3) : EMPTY_STRING;
+    const message = "LLM API error: failed to parse JSON response from server" + (preview ? `; response body (truncated): ${preview}` : EMPTY_STRING);
+    const error = new Error(message);
+    error.cause = err;
+    error.status = response.status;
+    throw error;
+  }
+}
+function ensureMessages(messages) {
+  if (!Array.isArray(messages) || messages.length === 0) {
+    const err = new Error("messages must be a non-empty array");
+    err.status = 400;
+    throw err;
+  }
+}
+function buildAiUrl(options, baseUrl) {
+  return `${getAiBaseUrl({ ...options, baseUrl })}/v1/chat/completions`;
+}
+async function fetchWithRetry(url, options, maxRetries) {
+  let attempt = 0;
+  let lastError;
+  while (attempt <= maxRetries) {
+    try {
+      const response = await fetch(url, options);
+      if (response.ok) return response;
+      throw await buildResponseError(response);
+    } catch (error) {
+      lastError = error;
+      if (!shouldRetryError(error, attempt, maxRetries)) throw error;
+      await sleep(getRetryDelayMs(attempt));
+      attempt += 1;
+    }
+  }
+  throw lastError ?? new Error("LLM API error: request failed after retries");
+}
+async function buildResponseError(response) {
+  const errText = await response.text();
+  const error = new Error(`LLM API error: ${response.status} - ${errText}`);
+  error.status = response.status;
+  return error;
+}
+function shouldRetryError(error, attempt, maxRetries) {
+  if (attempt >= maxRetries) return false;
+  const status = getErrorStatus(error);
+  if (typeof status === "number") {
+    return status >= 500;
+  }
+  return true;
+}
+function getErrorStatus(error) {
+  return typeof error?.status === "number" ? error.status : void 0;
+}
+function getPayload(input) {
+  if ("payload" in input) {
+    return input.payload;
+  }
+  return input.eventPayload;
+}
+function getRepoMetadata(payload) {
+  const repoPayload = payload;
+  return {
+    owner: repoPayload?.repository?.owner?.login ?? EMPTY_STRING,
+    repo: repoPayload?.repository?.name ?? EMPTY_STRING,
+    installationId: repoPayload?.installation?.id
+  };
+}
+function buildHeaders(authToken, options) {
+  const headers = {
+    Authorization: `Bearer ${authToken}`,
+    "Content-Type": "application/json"
+  };
+  if (options.owner) headers["X-GitHub-Owner"] = options.owner;
+  if (options.repo) headers["X-GitHub-Repo"] = options.repo;
+  if (typeof options.installationId === "number" && Number.isFinite(options.installationId)) {
+    headers["X-GitHub-Installation-Id"] = String(options.installationId);
+  }
+  if (options.ubiquityKernelToken) {
+    headers["X-Ubiquity-Kernel-Token"] = options.ubiquityKernelToken;
+  }
+  return headers;
 }
 async function* parseSseStream(body) {
   const reader = body.getReader();
   const decoder = new TextDecoder();
-  let buffer = "";
+  let buffer = EMPTY_STRING;
   try {
     while (true) {
       const { value, done: isDone } = await reader.read();
       if (isDone) break;
       buffer += decoder.decode(value, { stream: true });
-      const events = buffer.split("\n\n");
-      buffer = events.pop() || "";
+      const { events, remainder } = splitSseEvents(buffer);
+      buffer = remainder;
       for (const event of events) {
-        if (event.startsWith("data: ")) {
-          const data = event.slice(6);
-          if (data === "[DONE]") return;
-          yield JSON.parse(data);
-        }
+        const data = getEventData(event);
+        if (!data) continue;
+        if (data.trim() === "[DONE]") return;
+        yield parseEventData(data);
       }
     }
   } finally {
     reader.releaseLock();
   }
 }
+function splitSseEvents(buffer) {
+  const normalized = buffer.replace(/\r\n/g, "\n").replace(/\r/g, "\n");
+  const parts = normalized.split("\n\n");
+  const remainder = parts.pop() ?? EMPTY_STRING;
+  return { events: parts, remainder };
+}
+function getEventData(event) {
+  if (!event.trim()) return null;
+  const dataLines = event.split("\n").filter((line) => line.startsWith("data:"));
+  if (!dataLines.length) return null;
+  const data = dataLines.map((line) => line.startsWith("data: ") ? line.slice(6) : line.slice(5).replace(/^ /, EMPTY_STRING)).join("\n");
+  return data || null;
+}
+function parseEventData(data) {
+  try {
+    return JSON.parse(data);
+  } catch (error) {
+    if (data.includes("\n")) {
+      const collapsed = data.replace(/\n/g, EMPTY_STRING);
+      try {
+        return JSON.parse(collapsed);
+      } catch {
+      }
+    }
+    const message = error instanceof Error ? error.message : String(error);
+    const preview = data.length > 200 ? `${data.slice(0, 200)}...` : data;
+    throw new Error(`LLM stream parse error: ${message}. Data: ${preview}`);
+  }
+}
 export {
   callLlm
 };