@copilotkit/aimock 1.45.1 → 1.47.0

package/dist/sse-writer.js

@@ -1,12 +1,18 @@
+import { httpSink, runFrameLoop } from "./response/frame-loop.js";
+
 //#region src/sse-writer.ts
 function delay(ms, signal) {
 	if (ms <= 0 || signal?.aborted) return Promise.resolve();
 	return new Promise((resolve) => {
-		const timer = setTimeout(resolve, ms);
-		signal?.addEventListener("abort", () => {
+		const onAbort = () => {
 			clearTimeout(timer);
 			resolve();
-		}, { once: true });
+		};
+		const timer = setTimeout(() => {
+			signal?.removeEventListener("abort", onAbort);
+			resolve();
+		}, ms);
+		signal?.addEventListener("abort", onAbort, { once: true });
 	});
 }
 /** OpenRouter's SSE keepalive comment line (colon-prefixed, per SSE syntax). */
@@ -39,35 +45,53 @@ function calculateDelay(chunkIndex, profile, fallbackLatency, recordedTimings, r
 	delayMs = Math.max(0, delayMs);
 	return speed > 0 ? delayMs / speed : delayMs;
 }
+function calculateTotalDelay(contentLength, chunkSize, profile, recordedTimings, replaySpeed) {
+	const chunkCount = Math.max(1, Math.ceil(contentLength / chunkSize));
+	let total = 0;
+	for (let chunkIndex = 0; chunkIndex < chunkCount; chunkIndex++) total += calculateDelay(chunkIndex, profile, void 0, recordedTimings, replaySpeed);
+	return total;
+}
 async function writeSSEStream(res, chunks, optionsOrLatency) {
 	const opts = typeof optionsOrLatency === "number" ? { latency: optionsOrLatency } : optionsOrLatency ?? {};
-	const latency = opts.latency ?? 0;
-	const profile = opts.streamingProfile;
-	const { recordedTimings, replaySpeed } = opts;
-	const signal = opts.signal;
-	const onChunkSent = opts.onChunkSent;
-	if (res.writableEnded) return true;
-	res.setHeader("Content-Type", "text/event-stream");
-	res.setHeader("Cache-Control", "no-cache");
-	res.setHeader("Connection", "keep-alive");
-	if (opts.openRouterProcessing && !res.writableEnded) res.write(OPENROUTER_PROCESSING_COMMENT);
-	let chunkIndex = 0;
-	for (const chunk of chunks) {
-		const chunkDelay = calculateDelay(chunkIndex, profile, latency, recordedTimings, replaySpeed);
-		if (chunkDelay > 0) await delay(chunkDelay, signal);
-		if (signal?.aborted) return false;
-		if (res.writableEnded) return true;
-		res.write(`data: ${JSON.stringify(chunk)}\n\n`);
-		onChunkSent?.();
-		if (signal?.aborted) return false;
-		chunkIndex++;
-	}
-	if (!res.writableEnded) {
-		if (opts.usageChunk) res.write(`data: ${JSON.stringify(opts.usageChunk)}\n\n`);
-		res.write("data: [DONE]\n\n");
-		res.end();
-	}
-	return true;
+	return runFrameLoop(httpSink(res), chunks, sseCodec({
+		done: true,
+		usageChunk: opts.usageChunk,
+		openRouterProcessing: opts.openRouterProcessing
+	}), {
+		latency: opts.latency,
+		streamingProfile: opts.streamingProfile,
+		recordedTimings: opts.recordedTimings,
+		replaySpeed: opts.replaySpeed
+	}, {
+		signal: chunks.length > 0 ? opts.signal : void 0,
+		onChunkSent: opts.onChunkSent
+	});
+}
+const SSE_HEADERS = {
+	"Content-Type": "text/event-stream",
+	"Cache-Control": "no-cache",
+	Connection: "keep-alive"
+};
+function sseCodec(opts) {
+	return {
+		headers: SSE_HEADERS,
+		...opts.openRouterProcessing ? { prelude: () => OPENROUTER_PROCESSING_COMMENT } : {},
+		encode: (frame) => `data: ${JSON.stringify(frame)}\n\n`,
+		trailer: () => {
+			if (!opts.done) return [];
+			const parts = [];
+			if (opts.usageChunk) parts.push(`data: ${JSON.stringify(opts.usageChunk)}\n\n`);
+			parts.push("data: [DONE]\n\n");
+			return parts;
+		}
+	};
+}
+function typedSseCodec(opts) {
+	return {
+		headers: SSE_HEADERS,
+		encode: (frame) => `event: ${frame.type}\ndata: ${JSON.stringify(frame)}\n\n`,
+		trailer: () => opts?.terminalSentinel ? ["data: [DONE]\n\n"] : []
+	};
 }
 /**
 * Default rate-limit response headers matching OpenAI's format.
@@ -92,5 +116,5 @@ function writeErrorResponse(res, status, body, options) {
 }
 
 //#endregion
-export { calculateDelay, delay, writeErrorResponse, writeSSEStream };
+export { calculateDelay, calculateTotalDelay, delay, sseCodec, typedSseCodec, writeErrorResponse, writeSSEStream };
 //# sourceMappingURL=sse-writer.js.map

package/dist/sse-writer.js.map

@@ -1 +1 @@
-{"version":3,"file":"sse-writer.js","names":[],"sources":["../src/sse-writer.ts"],"sourcesContent":["import type * as http from \"node:http\";\nimport type { SSEChunk, StreamingProfile, RecordedTimings } from \"./types.js\";\n\nexport function delay(ms: number, signal?: AbortSignal): Promise<void> {\n  if (ms <= 0 || signal?.aborted) return Promise.resolve();\n  return new Promise((resolve) => {\n    const timer = setTimeout(resolve, ms);\n    signal?.addEventListener(\n      \"abort\",\n      () => {\n        clearTimeout(timer);\n        resolve();\n      },\n      { once: true },\n    );\n  });\n}\n\nexport interface StreamOptions {\n  latency?: number;\n  streamingProfile?: StreamingProfile;\n  recordedTimings?: RecordedTimings;\n  replaySpeed?: number;\n  signal?: AbortSignal;\n  onChunkSent?: () => void;\n  /** When set, emitted as the final chunk before [DONE] (OpenAI stream_options.include_usage). */\n  usageChunk?: SSEChunk;\n  /**\n   * When true, a single OpenRouter `: OPENROUTER PROCESSING` SSE comment\n   * (keepalive) line is emitted FIRST — before the first data chunk — matching\n   * real OpenRouter bytes (or-capture/chat-stream.sse). A faithful client skips\n   * `:`-prefixed lines; opt-in, default off (so strict OpenAI parsers aren't\n   * surprised). Used only by the OpenRouter chat surface.\n   */\n  openRouterProcessing?: boolean;\n}\n\n/** OpenRouter's SSE keepalive comment line (colon-prefixed, per SSE syntax). */\nconst OPENROUTER_PROCESSING_COMMENT = \": OPENROUTER PROCESSING\\n\\n\";\n\nexport function calculateDelay(\n  chunkIndex: number,\n  profile?: StreamingProfile,\n  fallbackLatency?: number,\n  recordedTimings?: RecordedTimings,\n  replaySpeed?: number,\n): number {\n  const speed = replaySpeed ?? 1.0;\n  let delayMs: number;\n\n  if (profile) {\n    // StreamingProfile has highest precedence\n    let fromProfile = true;\n    if (chunkIndex === 0 && profile.ttft !== undefined) {\n      delayMs = profile.ttft;\n    } else if (profile.tps !== undefined && profile.tps > 0) {\n      delayMs = 1000 / profile.tps;\n    } else {\n      delayMs = fallbackLatency ?? 0;\n      fromProfile = false;\n    }\n    // Jitter only applies when the delay came from ttft/tps, not fallback\n    if (fromProfile && profile.jitter && profile.jitter > 0) {\n      delayMs *= 1 + (Math.random() * 2 - 1) * profile.jitter;\n      if (delayMs < 0) delayMs = 0;\n    }\n  } else if (recordedTimings) {\n    // Recorded timings (second precedence)\n    if (chunkIndex === 0) {\n      delayMs = recordedTimings.ttftMs;\n    } else {\n      const idx = chunkIndex - 1;\n      if (idx < recordedTimings.interChunkDelaysMs.length) {\n        delayMs = recordedTimings.interChunkDelaysMs[idx];\n      } else {\n        // Excess chunks: derive average from recorded inter-chunk delays\n        const totalInterChunk = recordedTimings.interChunkDelaysMs.reduce((a, b) => a + b, 0);\n        delayMs =\n          recordedTimings.interChunkDelaysMs.length > 0\n            ? totalInterChunk / recordedTimings.interChunkDelaysMs.length\n            : 0;\n      }\n    }\n  } else {\n    delayMs = fallbackLatency ?? 0;\n  }\n\n  delayMs = Math.max(0, delayMs);\n  return speed > 0 ? delayMs / speed : delayMs;\n}\n\nexport async function writeSSEStream(\n  res: http.ServerResponse,\n  chunks: SSEChunk[],\n  optionsOrLatency?: number | StreamOptions,\n): Promise<boolean> {\n  const opts: StreamOptions =\n    typeof optionsOrLatency === \"number\" ? { latency: optionsOrLatency } : (optionsOrLatency ?? {});\n  const latency = opts.latency ?? 0;\n  const profile = opts.streamingProfile;\n  const { recordedTimings, replaySpeed } = opts;\n  const signal = opts.signal;\n  const onChunkSent = opts.onChunkSent;\n\n  if (res.writableEnded) return true;\n  res.setHeader(\"Content-Type\", \"text/event-stream\");\n  res.setHeader(\"Cache-Control\", \"no-cache\");\n  res.setHeader(\"Connection\", \"keep-alive\");\n\n  // OpenRouter emits exactly ONE `: OPENROUTER PROCESSING` keepalive comment\n  // first — before any data frame (opt-in; see or-capture/chat-stream.sse). A\n  // faithful client discards `:`-prefixed lines before JSON-parsing a `data:`.\n  if (opts.openRouterProcessing && !res.writableEnded) {\n    res.write(OPENROUTER_PROCESSING_COMMENT);\n  }\n\n  let chunkIndex = 0;\n  for (const chunk of chunks) {\n    const chunkDelay = calculateDelay(chunkIndex, profile, latency, recordedTimings, replaySpeed);\n    if (chunkDelay > 0) {\n      await delay(chunkDelay, signal);\n    }\n    if (signal?.aborted) return false;\n    if (res.writableEnded) return true;\n    res.write(`data: ${JSON.stringify(chunk)}\\n\\n`);\n    onChunkSent?.();\n    if (signal?.aborted) return false;\n    chunkIndex++;\n  }\n\n  if (!res.writableEnded) {\n    if (opts.usageChunk) {\n      res.write(`data: ${JSON.stringify(opts.usageChunk)}\\n\\n`);\n    }\n    res.write(\"data: [DONE]\\n\\n\");\n    res.end();\n  }\n  return true;\n}\n\n/**\n * Default rate-limit response headers matching OpenAI's format.\n * Values are static — aimock doesn't track actual request counts.\n */\nconst RATE_LIMIT_HEADERS: Record<string, string> = {\n  \"x-ratelimit-limit-requests\": \"60\",\n  \"x-ratelimit-limit-tokens\": \"150000\",\n  \"x-ratelimit-remaining-requests\": \"0\",\n  \"x-ratelimit-remaining-tokens\": \"0\",\n  \"x-ratelimit-reset-requests\": \"1s\",\n  \"x-ratelimit-reset-tokens\": \"6m0s\",\n};\n\nexport interface ErrorResponseOptions {\n  /** Override the Retry-After header value (seconds). Default: 1. Only applied on 429. */\n  retryAfter?: number;\n}\n\nexport function writeErrorResponse(\n  res: http.ServerResponse,\n  status: number,\n  body: string,\n  options?: ErrorResponseOptions,\n): void {\n  const headers: Record<string, string> = { \"Content-Type\": \"application/json\" };\n  if (status === 429) {\n    headers[\"Retry-After\"] = String(options?.retryAfter ?? 1);\n    Object.assign(headers, RATE_LIMIT_HEADERS);\n  }\n  res.writeHead(status, headers);\n  res.end(body);\n}\n"],"mappings":";AAGA,SAAgB,MAAM,IAAY,QAAqC;AACrE,KAAI,MAAM,KAAK,QAAQ,QAAS,QAAO,QAAQ,SAAS;AACxD,QAAO,IAAI,SAAS,YAAY;EAC9B,MAAM,QAAQ,WAAW,SAAS,GAAG;AACrC,UAAQ,iBACN,eACM;AACJ,gBAAa,MAAM;AACnB,YAAS;KAEX,EAAE,MAAM,MAAM,CACf;GACD;;;AAuBJ,MAAM,gCAAgC;AAEtC,SAAgB,eACd,YACA,SACA,iBACA,iBACA,aACQ;CACR,MAAM,QAAQ,eAAe;CAC7B,IAAI;AAEJ,KAAI,SAAS;EAEX,IAAI,cAAc;AAClB,MAAI,eAAe,KAAK,QAAQ,SAAS,OACvC,WAAU,QAAQ;WACT,QAAQ,QAAQ,UAAa,QAAQ,MAAM,EACpD,WAAU,MAAO,QAAQ;OACpB;AACL,aAAU,mBAAmB;AAC7B,iBAAc;;AAGhB,MAAI,eAAe,QAAQ,UAAU,QAAQ,SAAS,GAAG;AACvD,cAAW,KAAK,KAAK,QAAQ,GAAG,IAAI,KAAK,QAAQ;AACjD,OAAI,UAAU,EAAG,WAAU;;YAEpB,gBAET,KAAI,eAAe,EACjB,WAAU,gBAAgB;MACrB;EACL,MAAM,MAAM,aAAa;AACzB,MAAI,MAAM,gBAAgB,mBAAmB,OAC3C,WAAU,gBAAgB,mBAAmB;OACxC;GAEL,MAAM,kBAAkB,gBAAgB,mBAAmB,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE;AACrF,aACE,gBAAgB,mBAAmB,SAAS,IACxC,kBAAkB,gBAAgB,mBAAmB,SACrD;;;KAIV,WAAU,mBAAmB;AAG/B,WAAU,KAAK,IAAI,GAAG,QAAQ;AAC9B,QAAO,QAAQ,IAAI,UAAU,QAAQ;;AAGvC,eAAsB,eACpB,KACA,QACA,kBACkB;CAClB,MAAM,OACJ,OAAO,qBAAqB,WAAW,EAAE,SAAS,kBAAkB,GAAI,oBAAoB,EAAE;CAChG,MAAM,UAAU,KAAK,WAAW;CAChC,MAAM,UAAU,KAAK;CACrB,MAAM,EAAE,iBAAiB,gBAAgB;CACzC,MAAM,SAAS,KAAK;CACpB,MAAM,cAAc,KAAK;AAEzB,KAAI,IAAI,cAAe,QAAO;AAC9B,KAAI,UAAU,gBAAgB,oBAAoB;AAClD,KAAI,UAAU,iBAAiB,WAAW;AAC1C,KAAI,UAAU,cAAc,aAAa;AAKzC,KAAI,KAAK,wBAAwB,CAAC,IAAI,cACpC,KAAI,MAAM,8BAA8B;CAG1C,IAAI,aAAa;AACjB,MAAK,MAAM,SAAS,QAAQ;EAC1B,MAAM,aAAa,eAAe,YAAY,SAAS,SAAS,iBAAiB,YAAY;AAC7F,MAAI,aAAa,EACf,OAAM,MAAM,YAAY,OAAO;AAEjC,MAAI,QAAQ,QAAS,QAAO;AAC5B,MAAI,IAAI,cAAe,QAAO;AAC9B,MAAI,MAAM,SAAS,KAAK,UAAU,MAAM,CAAC,MAAM;AAC/C,iBAAe;AACf,MAAI,QAAQ,QAAS,QAAO;AAC5B;;AAGF,KAAI,CAAC,IAAI,eAAe;AACtB,MAAI,KAAK,WACP,KAAI,MAAM,SAAS,KAAK,UAAU,KAAK,WAAW,CAAC,MAAM;AAE3D,MAAI,MAAM,mBAAmB;AAC7B,MAAI,KAAK;;AAEX,QAAO;;;;;;AAOT,MAAM,qBAA6C;CACjD,8BAA8B;CAC9B,4BAA4B;CAC5B,kCAAkC;CAClC,gCAAgC;CAChC,8BAA8B;CAC9B,4BAA4B;CAC7B;AAOD,SAAgB,mBACd,KACA,QACA,MACA,SACM;CACN,MAAM,UAAkC,EAAE,gBAAgB,oBAAoB;AAC9E,KAAI,WAAW,KAAK;AAClB,UAAQ,iBAAiB,OAAO,SAAS,cAAc,EAAE;AACzD,SAAO,OAAO,SAAS,mBAAmB;;AAE5C,KAAI,UAAU,QAAQ,QAAQ;AAC9B,KAAI,IAAI,KAAK"}
+{"version":3,"file":"sse-writer.js","names":[],"sources":["../src/sse-writer.ts"],"sourcesContent":["import type * as http from \"node:http\";\nimport type { SSEChunk, StreamingProfile, RecordedTimings } from \"./types.js\";\nimport { httpSink, runFrameLoop, type FrameCodec } from \"./response/frame-loop.js\";\n\nexport function delay(ms: number, signal?: AbortSignal): Promise<void> {\n  if (ms <= 0 || signal?.aborted) return Promise.resolve();\n  return new Promise((resolve) => {\n    const onAbort = (): void => {\n      clearTimeout(timer);\n      resolve();\n    };\n    // Remove the abort listener when the timer fires, so a long stream sharing\n    // one lifecycle signal does not keep one listener per delayed frame.\n    const timer = setTimeout(() => {\n      signal?.removeEventListener(\"abort\", onAbort);\n      resolve();\n    }, ms);\n    signal?.addEventListener(\"abort\", onAbort, { once: true });\n  });\n}\n\nexport interface StreamOptions {\n  latency?: number;\n  streamingProfile?: StreamingProfile;\n  recordedTimings?: RecordedTimings;\n  replaySpeed?: number;\n  signal?: AbortSignal;\n  onChunkSent?: () => void;\n  /** When set, emitted as the final chunk before [DONE] (OpenAI stream_options.include_usage). */\n  usageChunk?: SSEChunk;\n  /**\n   * When true, a single OpenRouter `: OPENROUTER PROCESSING` SSE comment\n   * (keepalive) line is emitted FIRST — before the first data chunk — matching\n   * real OpenRouter bytes (or-capture/chat-stream.sse). A faithful client skips\n   * `:`-prefixed lines; opt-in, default off (so strict OpenAI parsers aren't\n   * surprised). Used only by the OpenRouter chat surface.\n   */\n  openRouterProcessing?: boolean;\n}\n\n/** OpenRouter's SSE keepalive comment line (colon-prefixed, per SSE syntax). */\nconst OPENROUTER_PROCESSING_COMMENT = \": OPENROUTER PROCESSING\\n\\n\";\n\nexport function calculateDelay(\n  chunkIndex: number,\n  profile?: StreamingProfile,\n  fallbackLatency?: number,\n  recordedTimings?: RecordedTimings,\n  replaySpeed?: number,\n): number {\n  const speed = replaySpeed ?? 1.0;\n  let delayMs: number;\n\n  if (profile) {\n    // StreamingProfile has highest precedence\n    let fromProfile = true;\n    if (chunkIndex === 0 && profile.ttft !== undefined) {\n      delayMs = profile.ttft;\n    } else if (profile.tps !== undefined && profile.tps > 0) {\n      delayMs = 1000 / profile.tps;\n    } else {\n      delayMs = fallbackLatency ?? 0;\n      fromProfile = false;\n    }\n    // Jitter only applies when the delay came from ttft/tps, not fallback\n    if (fromProfile && profile.jitter && profile.jitter > 0) {\n      delayMs *= 1 + (Math.random() * 2 - 1) * profile.jitter;\n      if (delayMs < 0) delayMs = 0;\n    }\n  } else if (recordedTimings) {\n    // Recorded timings (second precedence)\n    if (chunkIndex === 0) {\n      delayMs = recordedTimings.ttftMs;\n    } else {\n      const idx = chunkIndex - 1;\n      if (idx < recordedTimings.interChunkDelaysMs.length) {\n        delayMs = recordedTimings.interChunkDelaysMs[idx];\n      } else {\n        // Excess chunks: derive average from recorded inter-chunk delays\n        const totalInterChunk = recordedTimings.interChunkDelaysMs.reduce((a, b) => a + b, 0);\n        delayMs =\n          recordedTimings.interChunkDelaysMs.length > 0\n            ? totalInterChunk / recordedTimings.interChunkDelaysMs.length\n            : 0;\n      }\n    }\n  } else {\n    delayMs = fallbackLatency ?? 0;\n  }\n\n  delayMs = Math.max(0, delayMs);\n  return speed > 0 ? delayMs / speed : delayMs;\n}\n\n// Internal (not part of the public API): the total wait before a non-streaming\n// (`stream: false`) JSON response is written. Sums calculateDelay over the\n// content chunk count that `contentLength` at `chunkSize` would produce, so a\n// fixture with an explicit `streamingProfile` or `recordedTimings` pays roughly\n// the same ttft/tps cost as its streamed content chunks. It is an\n// approximation, not an exact match: a streamed reply also pays delays for its\n// role and finish chunks. There is deliberately no fallback-latency parameter:\n// bare `latency` / `--latency` is an inter-SSE-chunk gap and never delays a\n// non-streaming response. With neither a profile nor recorded timings, this\n// returns 0.\n// Kept as a line comment, not JSDoc: the bundled .d.ts drops this non-public\n// declaration but keeps a JSDoc block, which would then attach to writeSSEStream.\nexport function calculateTotalDelay(\n  contentLength: number,\n  chunkSize: number,\n  profile?: StreamingProfile,\n  recordedTimings?: RecordedTimings,\n  replaySpeed?: number,\n): number {\n  const chunkCount = Math.max(1, Math.ceil(contentLength / chunkSize));\n  let total = 0;\n  for (let chunkIndex = 0; chunkIndex < chunkCount; chunkIndex++) {\n    total += calculateDelay(chunkIndex, profile, undefined, recordedTimings, replaySpeed);\n  }\n  return total;\n}\n\nexport async function writeSSEStream(\n  res: http.ServerResponse,\n  chunks: SSEChunk[],\n  optionsOrLatency?: number | StreamOptions,\n): Promise<boolean> {\n  const opts: StreamOptions =\n    typeof optionsOrLatency === \"number\" ? { latency: optionsOrLatency } : (optionsOrLatency ?? {});\n  // No lifecycle: an external caller gets close-awareness only through its own\n  // `signal` (a gone client still returns true).\n  // Base contract (public API): with no frames, an aborted signal does not\n  // stop the writer; it still finishes the stream (any trailer, then end())\n  // and returns true. The frame loop returns false there, so pass it no\n  // signal when there are no frames (it reads the signal only per frame and\n  // after the loop).\n  return runFrameLoop(\n    httpSink(res),\n    chunks,\n    sseCodec({\n      done: true,\n      usageChunk: opts.usageChunk,\n      openRouterProcessing: opts.openRouterProcessing,\n    }),\n    {\n      latency: opts.latency,\n      streamingProfile: opts.streamingProfile,\n      recordedTimings: opts.recordedTimings,\n      replaySpeed: opts.replaySpeed,\n    },\n    { signal: chunks.length > 0 ? opts.signal : undefined, onChunkSent: opts.onChunkSent },\n  );\n}\n\nconst SSE_HEADERS: Record<string, string> = {\n  \"Content-Type\": \"text/event-stream\",\n  \"Cache-Control\": \"no-cache\",\n  Connection: \"keep-alive\",\n};\n\n// Internal (not part of the public API): data-only SSE (`data: <json>\\n\\n`).\n// `done: true` writes the optional usage chunk then `data: [DONE]` after the\n// last frame (OpenAI chat); `done: false` writes no trailer.\n// `openRouterProcessing` writes the OpenRouter keepalive comment first. Line\n// comments, not JSDoc: see calculateTotalDelay.\nexport function sseCodec<F = SSEChunk>(opts: {\n  done: boolean;\n  usageChunk?: SSEChunk;\n  openRouterProcessing?: boolean;\n}): FrameCodec<F> {\n  return {\n    headers: SSE_HEADERS,\n    ...(opts.openRouterProcessing ? { prelude: () => OPENROUTER_PROCESSING_COMMENT } : {}),\n    encode: (frame) => `data: ${JSON.stringify(frame)}\\n\\n`,\n    trailer: () => {\n      if (!opts.done) return [];\n      const parts: string[] = [];\n      if (opts.usageChunk) parts.push(`data: ${JSON.stringify(opts.usageChunk)}\\n\\n`);\n      parts.push(\"data: [DONE]\\n\\n\");\n      return parts;\n    },\n  };\n}\n\n// Internal: typed SSE (`event: <type>\\ndata: <json>\\n\\n`), used by Anthropic.\n// `terminalSentinel` writes `data: [DONE]` after the last frame, only when the\n// stream was not aborted.\nexport function typedSseCodec<F extends { type: string }>(opts?: {\n  terminalSentinel?: boolean;\n}): FrameCodec<F> {\n  return {\n    headers: SSE_HEADERS,\n    encode: (frame) => `event: ${frame.type}\\ndata: ${JSON.stringify(frame)}\\n\\n`,\n    trailer: () => (opts?.terminalSentinel ? [\"data: [DONE]\\n\\n\"] : []),\n  };\n}\n\n/**\n * Default rate-limit response headers matching OpenAI's format.\n * Values are static — aimock doesn't track actual request counts.\n */\nconst RATE_LIMIT_HEADERS: Record<string, string> = {\n  \"x-ratelimit-limit-requests\": \"60\",\n  \"x-ratelimit-limit-tokens\": \"150000\",\n  \"x-ratelimit-remaining-requests\": \"0\",\n  \"x-ratelimit-remaining-tokens\": \"0\",\n  \"x-ratelimit-reset-requests\": \"1s\",\n  \"x-ratelimit-reset-tokens\": \"6m0s\",\n};\n\nexport interface ErrorResponseOptions {\n  /** Override the Retry-After header value (seconds). Default: 1. Only applied on 429. */\n  retryAfter?: number;\n}\n\nexport function writeErrorResponse(\n  res: http.ServerResponse,\n  status: number,\n  body: string,\n  options?: ErrorResponseOptions,\n): void {\n  const headers: Record<string, string> = { \"Content-Type\": \"application/json\" };\n  if (status === 429) {\n    headers[\"Retry-After\"] = String(options?.retryAfter ?? 1);\n    Object.assign(headers, RATE_LIMIT_HEADERS);\n  }\n  res.writeHead(status, headers);\n  res.end(body);\n}\n"],"mappings":";;;AAIA,SAAgB,MAAM,IAAY,QAAqC;AACrE,KAAI,MAAM,KAAK,QAAQ,QAAS,QAAO,QAAQ,SAAS;AACxD,QAAO,IAAI,SAAS,YAAY;EAC9B,MAAM,gBAAsB;AAC1B,gBAAa,MAAM;AACnB,YAAS;;EAIX,MAAM,QAAQ,iBAAiB;AAC7B,WAAQ,oBAAoB,SAAS,QAAQ;AAC7C,YAAS;KACR,GAAG;AACN,UAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,MAAM,CAAC;GAC1D;;;AAuBJ,MAAM,gCAAgC;AAEtC,SAAgB,eACd,YACA,SACA,iBACA,iBACA,aACQ;CACR,MAAM,QAAQ,eAAe;CAC7B,IAAI;AAEJ,KAAI,SAAS;EAEX,IAAI,cAAc;AAClB,MAAI,eAAe,KAAK,QAAQ,SAAS,OACvC,WAAU,QAAQ;WACT,QAAQ,QAAQ,UAAa,QAAQ,MAAM,EACpD,WAAU,MAAO,QAAQ;OACpB;AACL,aAAU,mBAAmB;AAC7B,iBAAc;;AAGhB,MAAI,eAAe,QAAQ,UAAU,QAAQ,SAAS,GAAG;AACvD,cAAW,KAAK,KAAK,QAAQ,GAAG,IAAI,KAAK,QAAQ;AACjD,OAAI,UAAU,EAAG,WAAU;;YAEpB,gBAET,KAAI,eAAe,EACjB,WAAU,gBAAgB;MACrB;EACL,MAAM,MAAM,aAAa;AACzB,MAAI,MAAM,gBAAgB,mBAAmB,OAC3C,WAAU,gBAAgB,mBAAmB;OACxC;GAEL,MAAM,kBAAkB,gBAAgB,mBAAmB,QAAQ,GAAG,MAAM,IAAI,GAAG,EAAE;AACrF,aACE,gBAAgB,mBAAmB,SAAS,IACxC,kBAAkB,gBAAgB,mBAAmB,SACrD;;;KAIV,WAAU,mBAAmB;AAG/B,WAAU,KAAK,IAAI,GAAG,QAAQ;AAC9B,QAAO,QAAQ,IAAI,UAAU,QAAQ;;AAevC,SAAgB,oBACd,eACA,WACA,SACA,iBACA,aACQ;CACR,MAAM,aAAa,KAAK,IAAI,GAAG,KAAK,KAAK,gBAAgB,UAAU,CAAC;CACpE,IAAI,QAAQ;AACZ,MAAK,IAAI,aAAa,GAAG,aAAa,YAAY,aAChD,UAAS,eAAe,YAAY,SAAS,QAAW,iBAAiB,YAAY;AAEvF,QAAO;;AAGT,eAAsB,eACpB,KACA,QACA,kBACkB;CAClB,MAAM,OACJ,OAAO,qBAAqB,WAAW,EAAE,SAAS,kBAAkB,GAAI,oBAAoB,EAAE;AAQhG,QAAO,aACL,SAAS,IAAI,EACb,QACA,SAAS;EACP,MAAM;EACN,YAAY,KAAK;EACjB,sBAAsB,KAAK;EAC5B,CAAC,EACF;EACE,SAAS,KAAK;EACd,kBAAkB,KAAK;EACvB,iBAAiB,KAAK;EACtB,aAAa,KAAK;EACnB,EACD;EAAE,QAAQ,OAAO,SAAS,IAAI,KAAK,SAAS;EAAW,aAAa,KAAK;EAAa,CACvF;;AAGH,MAAM,cAAsC;CAC1C,gBAAgB;CAChB,iBAAiB;CACjB,YAAY;CACb;AAOD,SAAgB,SAAuB,MAIrB;AAChB,QAAO;EACL,SAAS;EACT,GAAI,KAAK,uBAAuB,EAAE,eAAe,+BAA+B,GAAG,EAAE;EACrF,SAAS,UAAU,SAAS,KAAK,UAAU,MAAM,CAAC;EAClD,eAAe;AACb,OAAI,CAAC,KAAK,KAAM,QAAO,EAAE;GACzB,MAAM,QAAkB,EAAE;AAC1B,OAAI,KAAK,WAAY,OAAM,KAAK,SAAS,KAAK,UAAU,KAAK,WAAW,CAAC,MAAM;AAC/E,SAAM,KAAK,mBAAmB;AAC9B,UAAO;;EAEV;;AAMH,SAAgB,cAA0C,MAExC;AAChB,QAAO;EACL,SAAS;EACT,SAAS,UAAU,UAAU,MAAM,KAAK,UAAU,KAAK,UAAU,MAAM,CAAC;EACxE,eAAgB,MAAM,mBAAmB,CAAC,mBAAmB,GAAG,EAAE;EACnE;;;;;;AAOH,MAAM,qBAA6C;CACjD,8BAA8B;CAC9B,4BAA4B;CAC5B,kCAAkC;CAClC,gCAAgC;CAChC,8BAA8B;CAC9B,4BAA4B;CAC7B;AAOD,SAAgB,mBACd,KACA,QACA,MACA,SACM;CACN,MAAM,UAAkC,EAAE,gBAAgB,oBAAoB;AAC9E,KAAI,WAAW,KAAK;AAClB,UAAQ,iBAAiB,OAAO,SAAS,cAAc,EAAE;AACzD,SAAO,OAAO,SAAS,mBAAmB;;AAE5C,KAAI,UAAU,QAAQ,QAAQ;AAC9B,KAAI,IAAI,KAAK"}

package/dist/stream-collapse.cjs

@@ -1,4 +1,5 @@
 const require_runtime = require('./_virtual/_rolldown/runtime.cjs');
+const require_recording = require('./response/usage/recording.cjs');
 const require_harmony = require('./harmony.cjs');
 let node_zlib = require("node:zlib");
 
@@ -361,7 +362,15 @@ function collapseOpenAISSEWithResponsesTools(rawBody) {
 			transcript = parsed.text;
 			transcriptSeen = true;
 			if (Array.isArray(parsed.languages)) transcriptLanguages = parsed.languages.filter((language) => typeof language === "object" && language !== null && typeof language.code === "string").map((language) => ({ code: language.code }));
-			if (parsed.usage && typeof parsed.usage === "object" && !Array.isArray(parsed.usage) && Object.keys(parsed.usage).length > 0) transcriptUsage = parsed.usage;
+			transcriptUsage = require_recording.captureUsage({
+				provider: "openai",
+				contentType: "text/event-stream",
+				operation: "transcription",
+				family: "transcription",
+				stage: "stream",
+				previous: transcriptUsage,
+				observation: parsed
+			});
 			continue;
 		}
 		if (parsed.type === "response.reasoning_summary_text.delta" && typeof parsed.delta === "string") {
@@ -548,7 +557,15 @@ function collapseOpenAISSEWithResponsesTools(rawBody) {
 			continue;
 		}
 		if (typeof parsed.type === "string" && parsed.type.startsWith("response.")) continue;
-		if (parsed.usage && typeof parsed.usage === "object" && !Array.isArray(parsed.usage) && Object.keys(parsed.usage).length > 0) usage = parsed.usage;
+		usage = require_recording.captureUsage({
+			provider: "openai",
+			contentType: "text/event-stream",
+			operation: "chat",
+			family: "openai",
+			stage: "stream",
+			previous: usage,
+			observation: parsed
+		});
 		const choices = parsed.choices;
 		if (!choices || choices.length === 0) continue;
 		const delta = choices[0].delta;
@@ -664,6 +681,8 @@ function collapseOpenAISSEWithResponsesTools(rawBody) {
 }
 /**
 * Collapse Anthropic Claude Messages SSE stream into a single response.
+* Captures start usage and numeric delta updates; null keeps the start value.
+* Cache TTL objects are captured from message_start, alongside flat counts.
 *
 * Format:
 *   event: message_start\ndata: {...}\n\n
@@ -672,6 +691,7 @@ function collapseOpenAISSEWithResponsesTools(rawBody) {
 function collapseAnthropicSSE(rawBody) {
 	const inputTruncated = isCollapseInputTruncated(rawBody);
 	const blocks = splitSSEEvents(guardCollapseBody(rawBody));
+	let usage;
 	let content = "";
 	let reasoning = "";
 	let reasoningSignature;
@@ -697,6 +717,15 @@ function collapseAnthropicSSE(rawBody) {
 			if (droppedChunks === 1) firstDroppedSample = `parse failed (${err instanceof Error ? err.message : "unknown"}): ${surrogateSafeSlice(payload, 200)}`;
 			continue;
 		}
+		usage = require_recording.captureUsage({
+			provider: "anthropic",
+			contentType: "text/event-stream",
+			operation: "chat",
+			family: "anthropic",
+			stage: "stream",
+			previous: usage,
+			observation: parsed
+		});
 		if (eventType === "content_block_start") {
 			const contentBlock = parsed.content_block;
 			const redactedData = capturedRedactedData(contentBlock);
@@ -754,6 +783,7 @@ function collapseAnthropicSSE(rawBody) {
 			...tc.id ? { id: tc.id } : {}
 		}));
 		return {
+			...usage ? { usage } : {},
 			...orderedBlocks ? { blocks: orderedBlocks } : {},
 			...content ? { content } : {},
 			toolCalls: orderedBlocks ? orderedToolCalls : indexSortedToolCalls,
@@ -766,6 +796,7 @@ function collapseAnthropicSSE(rawBody) {
 		};
 	}
 	return {
+		...usage ? { usage } : {},
 		content,
 		...reasoning ? { reasoning } : {},
 		...reasoningSignature ? { reasoningSignature } : {},
@@ -784,6 +815,7 @@ function collapseAnthropicSSE(rawBody) {
 function collapseGeminiSSE(rawBody) {
 	const inputTruncated = isCollapseInputTruncated(rawBody);
 	const lines = splitSSEEvents(guardCollapseBody(rawBody));
+	let usage;
 	let content = "";
 	let reasoning = "";
 	let droppedChunks = 0;
@@ -804,6 +836,15 @@ function collapseGeminiSSE(rawBody) {
 			if (droppedChunks === 1) firstDroppedSample = `parse failed (${err instanceof Error ? err.message : "unknown"}): ${surrogateSafeSlice(payload, 200)}`;
 			continue;
 		}
+		usage = require_recording.captureUsage({
+			provider: "gemini",
+			contentType: "text/event-stream",
+			operation: "chat",
+			family: "gemini",
+			stage: "stream",
+			previous: usage,
+			observation: parsed
+		});
 		const candidates = parsed.candidates;
 		if (!candidates || candidates.length === 0) continue;
 		const candidateContent = candidates[0].content;
@@ -839,6 +880,7 @@ function collapseGeminiSSE(rawBody) {
 		arguments: normalizeToolArguments(tc.arguments)
 	}));
 	if (audioB64) return {
+		...usage ? { usage } : {},
 		audioB64,
 		audioMimeType,
 		...content ? { content } : {},
@@ -851,6 +893,7 @@ function collapseGeminiSSE(rawBody) {
 	if (toolCalls.length > 0) {
 		const blocks = buildOrderedBlocks(orderAtoms);
 		return {
+			...usage ? { usage } : {},
 			...blocks ? { blocks } : {},
 			...content ? { content } : {},
 			toolCalls: normalizedToolCalls,
@@ -861,6 +904,7 @@ function collapseGeminiSSE(rawBody) {
 		};
 	}
 	return {
+		...usage ? { usage } : {},
 		content,
 		...reasoning ? { reasoning } : {},
 		...droppedChunks > 0 ? { droppedChunks } : {},
@@ -885,6 +929,7 @@ function collapseGeminiSSE(rawBody) {
 function collapseOllamaNDJSON(rawBody) {
 	const inputTruncated = isCollapseInputTruncated(rawBody);
 	const lines = guardCollapseBody(rawBody).split("\n").filter((l) => l.trim().length > 0);
+	let usage;
 	let content = "";
 	let reasoning = "";
 	let droppedChunks = 0;
@@ -902,6 +947,15 @@ function collapseOllamaNDJSON(rawBody) {
 			if (droppedChunks === 1) firstDroppedSample = `parse failed (${err instanceof Error ? err.message : "unknown"}): ${surrogateSafeSlice(line.trim(), 200)}`;
 			continue;
 		}
+		usage = require_recording.captureUsage({
+			provider: "ollama",
+			contentType: "application/x-ndjson",
+			operation: "chat",
+			family: "ollama",
+			stage: "stream",
+			previous: usage,
+			observation: parsed
+		});
 		const message = parsed.message;
 		if (message) {
 			if (typeof message.content === "string") {
@@ -951,6 +1005,7 @@ function collapseOllamaNDJSON(rawBody) {
 			arguments: normalizeToolArguments(tc.arguments)
 		}));
 		return {
+			...usage ? { usage } : {},
 			...blocks ? { blocks } : {},
 			...content ? { content } : {},
 			toolCalls: normalizedToolCalls,
@@ -963,6 +1018,7 @@ function collapseOllamaNDJSON(rawBody) {
 		};
 	}
 	return {
+		...usage ? { usage } : {},
 		content,
 		...reasoning ? { reasoning } : {},
 		...droppedChunks > 0 ? { droppedChunks } : {},
@@ -981,6 +1037,7 @@ function collapseOllamaNDJSON(rawBody) {
 function collapseCohereSSE(rawBody) {
 	const inputTruncated = isCollapseInputTruncated(rawBody);
 	const blocks = splitSSEEvents(guardCollapseBody(rawBody));
+	let usage;
 	let content = "";
 	let reasoning = "";
 	let droppedChunks = 0;
@@ -1004,6 +1061,15 @@ function collapseCohereSSE(rawBody) {
 			if (droppedChunks === 1) firstDroppedSample = `parse failed (${err instanceof Error ? err.message : "unknown"}): ${surrogateSafeSlice(payload, 200)}`;
 			continue;
 		}
+		usage = require_recording.captureUsage({
+			provider: "cohere",
+			contentType: "text/event-stream",
+			operation: "chat",
+			family: "cohere",
+			stage: "stream",
+			previous: usage,
+			observation: parsed
+		});
 		if (eventType === "content-delta") {
 			const contentObj = (parsed.delta?.message)?.content;
 			if (contentObj && contentObj.type === "thinking" && typeof contentObj.thinking === "string") reasoning += contentObj.thinking;
@@ -1064,6 +1130,7 @@ function collapseCohereSSE(rawBody) {
 			...tc.id ? { id: tc.id } : {}
 		}));
 		return {
+			...usage ? { usage } : {},
 			...orderedBlocks ? { blocks: orderedBlocks } : {},
 			...content ? { content } : {},
 			toolCalls: orderedBlocks ? orderedToolCalls : indexSortedToolCalls,
@@ -1074,6 +1141,7 @@ function collapseCohereSSE(rawBody) {
 		};
 	}
 	return {
+		...usage ? { usage } : {},
 		content,
 		...reasoning ? { reasoning } : {},
 		...droppedChunks > 0 ? { droppedChunks } : {},
@@ -1168,6 +1236,8 @@ function decodeEventStreamFrames(buf) {
 }
 /**
 * Collapse Bedrock binary Event Stream into a single response.
+* Claude invoke events preserve usage with the Anthropic start/delta rules;
+* Converse usage remains omitted.
 *
 * Each frame contains a JSON payload with event types like:
 *   contentBlockDelta, contentBlockStart, etc.
@@ -1177,6 +1247,7 @@ function collapseBedrockEventStream(rawBody) {
 	const inputTruncated = rawBody.byteLength > limit;
 	const { frames, truncated: frameTruncated } = decodeEventStreamFrames(inputTruncated ? rawBody.subarray(0, limit) : rawBody);
 	const truncated = frameTruncated || inputTruncated;
+	let usage;
 	let content = "";
 	let reasoning = "";
 	let reasoningSignature;
@@ -1205,6 +1276,15 @@ function collapseBedrockEventStream(rawBody) {
 			if (droppedChunks === 1) firstDroppedSample = `parse failed (${err instanceof Error ? err.message : "unknown"}): ${surrogateSafeSlice(frameStr, 200)}`;
 			continue;
 		}
+		usage = require_recording.captureUsage({
+			provider: "bedrock",
+			contentType: "application/vnd.amazon.eventstream",
+			operation: "chat",
+			family: typeof parsed.type === "string" ? "bedrock-invoke" : "bedrock-converse",
+			stage: "stream",
+			previous: usage,
+			observation: parsed
+		});
 		if (parsed.type === "content_block_delta") {
 			const delta = parsed.delta;
 			if (delta?.type === "text_delta" && typeof delta.text === "string") {
@@ -1312,6 +1392,7 @@ function collapseBedrockEventStream(rawBody) {
 			...tc.id ? { id: tc.id } : {}
 		}));
 		return {
+			...usage ? { usage } : {},
 			...orderedBlocks ? { blocks: orderedBlocks } : {},
 			...content ? { content } : {},
 			toolCalls: orderedBlocks ? orderedToolCalls : indexSortedToolCalls,
@@ -1324,6 +1405,7 @@ function collapseBedrockEventStream(rawBody) {
 		};
 	}
 	return {
+		...usage ? { usage } : {},
 		content,
 		...reasoning ? { reasoning } : {},
 		...reasoningSignature ? { reasoningSignature } : {},
@@ -1350,6 +1432,7 @@ function collapseBedrockEventStream(rawBody) {
 function collapseGeminiInteractionsSSE(rawBody) {
 	const inputTruncated = isCollapseInputTruncated(rawBody);
 	const lines = splitSSEEvents(guardCollapseBody(rawBody));
+	let usage;
 	let content = "";
 	let reasoning = "";
 	let droppedChunks = 0;
@@ -1369,6 +1452,15 @@ function collapseGeminiInteractionsSSE(rawBody) {
 			if (droppedChunks === 1) firstDroppedSample = `parse failed (${err instanceof Error ? err.message : "unknown"}): ${surrogateSafeSlice(payload, 200)}`;
 			continue;
 		}
+		usage = require_recording.captureUsage({
+			provider: "gemini-interactions",
+			contentType: "text/event-stream",
+			operation: "chat",
+			family: "interactions",
+			stage: "stream",
+			previous: usage,
+			observation: parsed
+		});
 		const eventType = parsed.event_type;
 		if (!eventType) continue;
 		const index = typeof parsed.index === "number" ? parsed.index : void 0;
@@ -1426,6 +1518,7 @@ function collapseGeminiInteractionsSSE(rawBody) {
 		});
 	}
 	if (toolCalls.length > 0) return {
+		...usage ? { usage } : {},
 		...content ? { content } : {},
 		toolCalls,
 		...reasoning ? { reasoning } : {},
@@ -1434,6 +1527,7 @@ function collapseGeminiInteractionsSSE(rawBody) {
 		...inputTruncated ? { truncated: true } : {}
 	};
 	return {
+		...usage ? { usage } : {},
 		content,
 		...reasoning ? { reasoning } : {},
 		...droppedChunks > 0 ? { droppedChunks } : {},
@@ -1459,24 +1553,37 @@ function collapseStreamingResponse(contentType, providerKey, body, logger) {
 */
 function collapseStreamingResponseWithResponsesTools(contentType, providerKey, body, logger) {
 	const ct = contentType.toLowerCase();
-	if (ct.includes("application/vnd.amazon.eventstream")) return collapseBedrockEventStream(typeof body === "string" ? Buffer.from(body, "binary") : body);
-	if (ct.includes("application/x-ndjson")) return collapseOllamaNDJSON(typeof body === "string" ? body : body.toString("utf8"));
+	const complete = (result) => {
+		const usage = require_recording.captureUsage({
+			provider: providerKey,
+			contentType,
+			family: "stream-dispatch",
+			stage: "collapsed",
+			observation: result
+		});
+		return usage ? {
+			...result,
+			usage
+		} : result;
+	};
+	if (ct.includes("application/vnd.amazon.eventstream")) return complete(collapseBedrockEventStream(typeof body === "string" ? Buffer.from(body, "binary") : body));
+	if (ct.includes("application/x-ndjson")) return complete(collapseOllamaNDJSON(typeof body === "string" ? body : body.toString("utf8")));
 	if (ct.includes("text/event-stream")) {
 		const str = typeof body === "string" ? body : body.toString("utf8");
 		switch (providerKey) {
 			case "openai":
 			case "azure":
 			case "openrouter":
-			case "byteplus": return collapseOpenAISSEWithResponsesTools(str);
-			case "anthropic": return collapseAnthropicSSE(str);
+			case "byteplus": return complete(collapseOpenAISSEWithResponsesTools(str));
+			case "anthropic": return complete(collapseAnthropicSSE(str));
 			case "gemini":
-			case "vertexai": return collapseGeminiSSE(str);
-			case "gemini-interactions": return collapseGeminiInteractionsSSE(str);
-			case "cohere": return collapseCohereSSE(str);
-			case "bedrock": return collapseAnthropicSSE(str);
+			case "vertexai": return complete(collapseGeminiSSE(str));
+			case "gemini-interactions": return complete(collapseGeminiInteractionsSSE(str));
+			case "cohere": return complete(collapseCohereSSE(str));
+			case "bedrock": return complete(collapseAnthropicSSE(str));
 			default:
 				logger?.warn(`[stream-collapse] unknown SSE provider "${providerKey}", falling back to OpenAI SSE format`);
-				return collapseOpenAISSEWithResponsesTools(str);
+				return complete(collapseOpenAISSEWithResponsesTools(str));
 		}
 	}
 	return null;