@cryer/star-cli 0.3.5 → 0.3.7

package/dist/stream-H4BYBVAZ.js

@@ -1,184 +0,0 @@
-#!/usr/bin/env node
-import {
-  starHome,
-  summarizeStreamError
-} from "./chunk-B6P4VUBE.js";
-
-// src/llm/stream.ts
-import { streamText } from "ai";
-
-// src/llm/debug.ts
-import { appendFileSync } from "fs";
-import path from "path";
-function debugStreamLog(event, data) {
-  if (!process.env.STAR_DEBUG) return;
-  try {
-    const line = `${JSON.stringify({ ts: (/* @__PURE__ */ new Date()).toISOString(), event, ...data })}
-`;
-    appendFileSync(path.join(starHome(), "debug.log"), line);
-  } catch {
-  }
-}
-
-// src/llm/stream.ts
-var DEFAULT_IDLE_TIMEOUT_MS = 2e4;
-var DEFAULT_FIRST_PART_TIMEOUT_MS = 12e4;
-var CACHE_READ_KEYS = ["cacheReadInputTokens", "cache_read_input_tokens"];
-var CACHED_PROMPT_KEYS = ["cachedPromptTokens", "cached_prompt_tokens", "cachedTokens"];
-function extractCacheUsage(providerMetadata) {
-  if (!providerMetadata || typeof providerMetadata !== "object") return {};
-  for (const value of Object.values(providerMetadata)) {
-    if (!value || typeof value !== "object") continue;
-    const record = value;
-    for (const key of CACHE_READ_KEYS) {
-      const n = record[key];
-      if (typeof n === "number" && Number.isFinite(n)) return { cacheReadInputTokens: n };
-    }
-    for (const key of CACHED_PROMPT_KEYS) {
-      const n = record[key];
-      if (typeof n === "number" && Number.isFinite(n)) return { cachedPromptTokens: n };
-    }
-  }
-  return {};
-}
-async function* streamChat(opts) {
-  const controller = new AbortController();
-  const onAbort = () => controller.abort();
-  opts.abortSignal?.addEventListener("abort", onAbort);
-  if (opts.abortSignal?.aborted) controller.abort();
-  const result = streamText({
-    model: opts.model,
-    messages: opts.messages,
-    tools: opts.tools,
-    abortSignal: controller.signal,
-    maxTokens: opts.maxTokens,
-    // Disable the SDK's own retries (ai@4 defaults to 2 internal attempts with
-    // fixed backoff that ignores Retry-After and the abort signal, and
-    // exhaustion throws an AI_RetryError stripped of status/headers/body) —
-    // the retry policy in agent/loop.ts + llm/retry.ts owns the full budget
-    // and needs the raw error to classify.
-    maxRetries: 0,
-    experimental_providerMetadata: {
-      // Only the openai provider (our responses-protocol path) reads these;
-      // other providers ignore unknown metadata. The SDK's responses provider
-      // sends every tool with strict:true by default (strictSchemas), and
-      // strict mode requires every schema property to be required — our
-      // classic function-calling schemas use optional properties, which
-      // relays then reject. store:false matches how codex and opencode talk
-      // to relays (no server-side response storage).
-      openai: { strictSchemas: false, store: false }
-    }
-  });
-  const idleTimeoutMs = opts.idleTimeoutMs ?? DEFAULT_IDLE_TIMEOUT_MS;
-  const firstPartTimeoutMs = opts.firstPartTimeoutMs ?? DEFAULT_FIRST_PART_TIMEOUT_MS;
-  const iterator = result.fullStream[Symbol.asyncIterator]();
-  let seenPart = false;
-  let deltas = 0;
-  let hasContent = false;
-  debugStreamLog("request", {
-    messages: opts.messages.length,
-    idleTimeoutMs,
-    firstPartTimeoutMs
-  });
-  try {
-    for (; ; ) {
-      let timer;
-      const idle = new Promise((resolve) => {
-        timer = setTimeout(() => resolve(null), seenPart ? idleTimeoutMs : firstPartTimeoutMs);
-      });
-      let next;
-      try {
-        next = await Promise.race([iterator.next(), idle]);
-      } catch (error) {
-        debugStreamLog("stream-throw", {
-          seenPart,
-          deltas,
-          error: error instanceof Error ? summarizeStreamError(error) : String(error)
-        });
-        throw error;
-      } finally {
-        if (timer) clearTimeout(timer);
-      }
-      if (next === null) {
-        debugStreamLog("idle-timeout", {
-          seenPart,
-          deltas,
-          waitedMs: seenPart ? idleTimeoutMs : firstPartTimeoutMs
-        });
-        yield {
-          type: "finish",
-          finishReason: "idle-timeout",
-          usage: void 0,
-          // Content already streamed means the reply may be cut off
-          // mid-thought; the loop surfaces a notice for that.
-          ...hasContent ? { truncated: true } : {}
-        };
-        return;
-      }
-      if (next.done) {
-        debugStreamLog("end", { seenPart, deltas });
-        return;
-      }
-      seenPart = true;
-      const part = next.value;
-      switch (part.type) {
-        case "text-delta":
-          deltas++;
-          hasContent = true;
-          yield { type: "text-delta", text: part.textDelta };
-          break;
-        case "reasoning":
-          deltas++;
-          hasContent = true;
-          yield { type: "reasoning", text: part.textDelta };
-          break;
-        case "tool-call":
-          hasContent = true;
-          yield {
-            type: "tool-call",
-            id: part.toolCallId,
-            name: part.toolName,
-            args: part.args
-          };
-          break;
-        case "finish": {
-          const finite = (n) => Number.isFinite(n) ? n : 0;
-          const cache = extractCacheUsage(
-            part.providerMetadata ?? part.experimental_providerMetadata
-          );
-          debugStreamLog("finish", {
-            finishReason: part.finishReason,
-            usage: part.usage,
-            deltas
-          });
-          yield {
-            type: "finish",
-            finishReason: part.finishReason,
-            usage: part.usage ? {
-              promptTokens: finite(part.usage.promptTokens),
-              completionTokens: finite(part.usage.completionTokens),
-              totalTokens: finite(part.usage.totalTokens),
-              ...cache
-            } : void 0
-          };
-          return;
-        }
-        case "error": {
-          if (opts.abortSignal?.aborted) return;
-          const error = part.error instanceof Error ? part.error : new Error(String(part.error));
-          debugStreamLog("error-part", { seenPart, deltas, error: summarizeStreamError(error) });
-          yield { type: "error", error };
-          break;
-        }
-      }
-    }
-  } finally {
-    opts.abortSignal?.removeEventListener("abort", onAbort);
-    controller.abort();
-  }
-}
-export {
-  extractCacheUsage,
-  streamChat
-};
-//# sourceMappingURL=stream-H4BYBVAZ.js.map

package/dist/stream-H4BYBVAZ.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/llm/stream.ts","../src/llm/debug.ts"],"sourcesContent":["import { type LanguageModel, type ToolSet, streamText } from \"ai\";\nimport type { StreamEvent, TokenUsage } from \"../core/events\";\nimport type { CoreMessage } from \"../core/messages\";\nimport { debugStreamLog } from \"./debug\";\nimport { summarizeStreamError } from \"./retry\";\n\nexport interface StreamChatOptions {\n  model: LanguageModel;\n  messages: CoreMessage[];\n  tools?: Record<string, unknown>;\n  abortSignal?: AbortSignal;\n  maxTokens?: number;\n  // Idle watchdog: some relays deliver the final content but never send the\n  // terminal chunks (or never close the socket). If no stream part arrives\n  // within this window once streaming has started, the stream is ended\n  // gracefully with what we have.\n  idleTimeoutMs?: number;\n  // Separate, longer allowance for the very first part: thinking models and\n  // slow relays can take a while before producing anything, and a false\n  // timeout here would kill a healthy request.\n  firstPartTimeoutMs?: number;\n}\n\nconst DEFAULT_IDLE_TIMEOUT_MS = 20_000;\nconst DEFAULT_FIRST_PART_TIMEOUT_MS = 120_000;\n\nconst CACHE_READ_KEYS = [\"cacheReadInputTokens\", \"cache_read_input_tokens\"];\nconst CACHED_PROMPT_KEYS = [\"cachedPromptTokens\", \"cached_prompt_tokens\", \"cachedTokens\"];\n\n// Pulls prompt-cache token counts out of a finish part's providerMetadata.\n// Anthropic-style fields (cacheReadInputTokens) report cache reads on top of\n// promptTokens; OpenAI-style fields (cachedPromptTokens) are a subset of it.\n// Only finite numbers are accepted; anything else means \"not provided\".\nexport function extractCacheUsage(\n  providerMetadata: unknown,\n): Pick<TokenUsage, \"cachedPromptTokens\" | \"cacheReadInputTokens\"> {\n  if (!providerMetadata || typeof providerMetadata !== \"object\") return {};\n  for (const value of Object.values(providerMetadata)) {\n    if (!value || typeof value !== \"object\") continue;\n    const record = value as Record<string, unknown>;\n    for (const key of CACHE_READ_KEYS) {\n      const n = record[key];\n      if (typeof n === \"number\" && Number.isFinite(n)) return { cacheReadInputTokens: n };\n    }\n    for (const key of CACHED_PROMPT_KEYS) {\n      const n = record[key];\n      if (typeof n === \"number\" && Number.isFinite(n)) return { cachedPromptTokens: n };\n    }\n  }\n  return {};\n}\n\nexport async function* streamChat(opts: StreamChatOptions): AsyncGenerator<StreamEvent> {\n  const controller = new AbortController();\n  const onAbort = () => controller.abort();\n  opts.abortSignal?.addEventListener(\"abort\", onAbort);\n  // Abort listeners only fire on future aborts: mirror an abort that already\n  // happened before this listener was registered (Esc landing while the loop\n  // was still preparing the request), or the request would run on until the\n  // idle watchdog despite having been cancelled.\n  if (opts.abortSignal?.aborted) controller.abort();\n  const result = streamText({\n    model: opts.model,\n    messages: opts.messages,\n    tools: opts.tools as ToolSet | undefined,\n    abortSignal: controller.signal,\n    maxTokens: opts.maxTokens,\n    // Disable the SDK's own retries (ai@4 defaults to 2 internal attempts with\n    // fixed backoff that ignores Retry-After and the abort signal, and\n    // exhaustion throws an AI_RetryError stripped of status/headers/body) —\n    // the retry policy in agent/loop.ts + llm/retry.ts owns the full budget\n    // and needs the raw error to classify.\n    maxRetries: 0,\n    experimental_providerMetadata: {\n      // Only the openai provider (our responses-protocol path) reads these;\n      // other providers ignore unknown metadata. The SDK's responses provider\n      // sends every tool with strict:true by default (strictSchemas), and\n      // strict mode requires every schema property to be required — our\n      // classic function-calling schemas use optional properties, which\n      // relays then reject. store:false matches how codex and opencode talk\n      // to relays (no server-side response storage).\n      openai: { strictSchemas: false, store: false },\n    },\n  });\n  const idleTimeoutMs = opts.idleTimeoutMs ?? DEFAULT_IDLE_TIMEOUT_MS;\n  const firstPartTimeoutMs = opts.firstPartTimeoutMs ?? DEFAULT_FIRST_PART_TIMEOUT_MS;\n  const iterator = result.fullStream[Symbol.asyncIterator]();\n  let seenPart = false;\n  let deltas = 0;\n  // Any visible content streamed (text, reasoning, or a tool call) — decides\n  // whether an idle-watchdog cutoff marks the finish event as truncated.\n  let hasContent = false;\n  debugStreamLog(\"request\", {\n    messages: opts.messages.length,\n    idleTimeoutMs,\n    firstPartTimeoutMs,\n  });\n  try {\n    for (;;) {\n      let timer: ReturnType<typeof setTimeout> | undefined;\n      const idle = new Promise<null>((resolve) => {\n        timer = setTimeout(() => resolve(null), seenPart ? idleTimeoutMs : firstPartTimeoutMs);\n      });\n      let next: Awaited<ReturnType<typeof iterator.next>> | null;\n      try {\n        next = await Promise.race([iterator.next(), idle]);\n      } catch (error) {\n        debugStreamLog(\"stream-throw\", {\n          seenPart,\n          deltas,\n          error: error instanceof Error ? summarizeStreamError(error) : String(error),\n        });\n        throw error;\n      } finally {\n        // Clear in finally: when iterator.next() rejects, skipping this would\n        // leave the watchdog timer pending for up to the full first-part\n        // allowance.\n        if (timer) clearTimeout(timer);\n      }\n      if (next === null) {\n        // Note: the AI SDK holds the model's finish part until the source\n        // stream closes, so a relay that stops sending without closing can\n        // only be detected by this watchdog.\n        debugStreamLog(\"idle-timeout\", {\n          seenPart,\n          deltas,\n          waitedMs: seenPart ? idleTimeoutMs : firstPartTimeoutMs,\n        });\n        yield {\n          type: \"finish\",\n          finishReason: \"idle-timeout\",\n          usage: undefined,\n          // Content already streamed means the reply may be cut off\n          // mid-thought; the loop surfaces a notice for that.\n          ...(hasContent ? { truncated: true } : {}),\n        };\n        return;\n      }\n      if (next.done) {\n        debugStreamLog(\"end\", { seenPart, deltas });\n        return;\n      }\n      seenPart = true;\n      const part = next.value;\n      switch (part.type) {\n        case \"text-delta\":\n          deltas++;\n          hasContent = true;\n          yield { type: \"text-delta\", text: part.textDelta };\n          break;\n        case \"reasoning\":\n          deltas++;\n          hasContent = true;\n          yield { type: \"reasoning\", text: part.textDelta };\n          break;\n        case \"tool-call\":\n          hasContent = true;\n          yield {\n            type: \"tool-call\",\n            id: part.toolCallId,\n            name: part.toolName,\n            args: part.args,\n          };\n          break;\n        case \"finish\": {\n          const finite = (n: number) => (Number.isFinite(n) ? n : 0);\n          const cache = extractCacheUsage(\n            part.providerMetadata ?? part.experimental_providerMetadata,\n          );\n          debugStreamLog(\"finish\", {\n            finishReason: part.finishReason,\n            usage: part.usage,\n            deltas,\n          });\n          yield {\n            type: \"finish\",\n            finishReason: part.finishReason,\n            usage: part.usage\n              ? {\n                  promptTokens: finite(part.usage.promptTokens),\n                  completionTokens: finite(part.usage.completionTokens),\n                  totalTokens: finite(part.usage.totalTokens),\n                  ...cache,\n                }\n              : undefined,\n          };\n          return;\n        }\n        case \"error\": {\n          // A user-initiated abort surfacing as a stream error is not a failure.\n          if (opts.abortSignal?.aborted) return;\n          const error = part.error instanceof Error ? part.error : new Error(String(part.error));\n          debugStreamLog(\"error-part\", { seenPart, deltas, error: summarizeStreamError(error) });\n          yield { type: \"error\", error };\n          break;\n        }\n      }\n    }\n  } finally {\n    opts.abortSignal?.removeEventListener(\"abort\", onAbort);\n    // Cancels the upstream request if it is still open (e.g. a relay that\n    // sent its last byte but kept the connection alive).\n    controller.abort();\n  }\n}\n","import { appendFileSync } from \"node:fs\";\nimport path from \"node:path\";\nimport { starHome } from \"../config/paths\";\n\n// STAR_DEBUG-gated JSONL log of stream internals (~/.star-cli/debug.log):\n// what the provider actually sent — finish reasons, usage, error details —\n// so a misbehaving relay can be diagnosed from the client side. All write\n// failures are silent; logging must never break a turn.\nexport function debugStreamLog(event: string, data?: Record<string, unknown>): void {\n  if (!process.env.STAR_DEBUG) return;\n  try {\n    const line = `${JSON.stringify({ ts: new Date().toISOString(), event, ...data })}\\n`;\n    appendFileSync(path.join(starHome(), \"debug.log\"), line);\n  } catch {\n    // best-effort only\n  }\n}\n"],"mappings":";;;;;;;AAAA,SAA2C,kBAAkB;;;ACA7D,SAAS,sBAAsB;AAC/B,OAAO,UAAU;AAOV,SAAS,eAAe,OAAe,MAAsC;AAClF,MAAI,CAAC,QAAQ,IAAI,WAAY;AAC7B,MAAI;AACF,UAAM,OAAO,GAAG,KAAK,UAAU,EAAE,KAAI,oBAAI,KAAK,GAAE,YAAY,GAAG,OAAO,GAAG,KAAK,CAAC,CAAC;AAAA;AAChF,mBAAe,KAAK,KAAK,SAAS,GAAG,WAAW,GAAG,IAAI;AAAA,EACzD,QAAQ;AAAA,EAER;AACF;;;ADOA,IAAM,0BAA0B;AAChC,IAAM,gCAAgC;AAEtC,IAAM,kBAAkB,CAAC,wBAAwB,yBAAyB;AAC1E,IAAM,qBAAqB,CAAC,sBAAsB,wBAAwB,cAAc;AAMjF,SAAS,kBACd,kBACiE;AACjE,MAAI,CAAC,oBAAoB,OAAO,qBAAqB,SAAU,QAAO,CAAC;AACvE,aAAW,SAAS,OAAO,OAAO,gBAAgB,GAAG;AACnD,QAAI,CAAC,SAAS,OAAO,UAAU,SAAU;AACzC,UAAM,SAAS;AACf,eAAW,OAAO,iBAAiB;AACjC,YAAM,IAAI,OAAO,GAAG;AACpB,UAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,EAAG,QAAO,EAAE,sBAAsB,EAAE;AAAA,IACpF;AACA,eAAW,OAAO,oBAAoB;AACpC,YAAM,IAAI,OAAO,GAAG;AACpB,UAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,EAAG,QAAO,EAAE,oBAAoB,EAAE;AAAA,IAClF;AAAA,EACF;AACA,SAAO,CAAC;AACV;AAEA,gBAAuB,WAAW,MAAsD;AACtF,QAAM,aAAa,IAAI,gBAAgB;AACvC,QAAM,UAAU,MAAM,WAAW,MAAM;AACvC,OAAK,aAAa,iBAAiB,SAAS,OAAO;AAKnD,MAAI,KAAK,aAAa,QAAS,YAAW,MAAM;AAChD,QAAM,SAAS,WAAW;AAAA,IACxB,OAAO,KAAK;AAAA,IACZ,UAAU,KAAK;AAAA,IACf,OAAO,KAAK;AAAA,IACZ,aAAa,WAAW;AAAA,IACxB,WAAW,KAAK;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAMhB,YAAY;AAAA,IACZ,+BAA+B;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,MAQ7B,QAAQ,EAAE,eAAe,OAAO,OAAO,MAAM;AAAA,IAC/C;AAAA,EACF,CAAC;AACD,QAAM,gBAAgB,KAAK,iBAAiB;AAC5C,QAAM,qBAAqB,KAAK,sBAAsB;AACtD,QAAM,WAAW,OAAO,WAAW,OAAO,aAAa,EAAE;AACzD,MAAI,WAAW;AACf,MAAI,SAAS;AAGb,MAAI,aAAa;AACjB,iBAAe,WAAW;AAAA,IACxB,UAAU,KAAK,SAAS;AAAA,IACxB;AAAA,IACA;AAAA,EACF,CAAC;AACD,MAAI;AACF,eAAS;AACP,UAAI;AACJ,YAAM,OAAO,IAAI,QAAc,CAAC,YAAY;AAC1C,gBAAQ,WAAW,MAAM,QAAQ,IAAI,GAAG,WAAW,gBAAgB,kBAAkB;AAAA,MACvF,CAAC;AACD,UAAI;AACJ,UAAI;AACF,eAAO,MAAM,QAAQ,KAAK,CAAC,SAAS,KAAK,GAAG,IAAI,CAAC;AAAA,MACnD,SAAS,OAAO;AACd,uBAAe,gBAAgB;AAAA,UAC7B;AAAA,UACA;AAAA,UACA,OAAO,iBAAiB,QAAQ,qBAAqB,KAAK,IAAI,OAAO,KAAK;AAAA,QAC5E,CAAC;AACD,cAAM;AAAA,MACR,UAAE;AAIA,YAAI,MAAO,cAAa,KAAK;AAAA,MAC/B;AACA,UAAI,SAAS,MAAM;AAIjB,uBAAe,gBAAgB;AAAA,UAC7B;AAAA,UACA;AAAA,UACA,UAAU,WAAW,gBAAgB;AAAA,QACvC,CAAC;AACD,cAAM;AAAA,UACJ,MAAM;AAAA,UACN,cAAc;AAAA,UACd,OAAO;AAAA;AAAA;AAAA,UAGP,GAAI,aAAa,EAAE,WAAW,KAAK,IAAI,CAAC;AAAA,QAC1C;AACA;AAAA,MACF;AACA,UAAI,KAAK,MAAM;AACb,uBAAe,OAAO,EAAE,UAAU,OAAO,CAAC;AAC1C;AAAA,MACF;AACA,iBAAW;AACX,YAAM,OAAO,KAAK;AAClB,cAAQ,KAAK,MAAM;AAAA,QACjB,KAAK;AACH;AACA,uBAAa;AACb,gBAAM,EAAE,MAAM,cAAc,MAAM,KAAK,UAAU;AACjD;AAAA,QACF,KAAK;AACH;AACA,uBAAa;AACb,gBAAM,EAAE,MAAM,aAAa,MAAM,KAAK,UAAU;AAChD;AAAA,QACF,KAAK;AACH,uBAAa;AACb,gBAAM;AAAA,YACJ,MAAM;AAAA,YACN,IAAI,KAAK;AAAA,YACT,MAAM,KAAK;AAAA,YACX,MAAM,KAAK;AAAA,UACb;AACA;AAAA,QACF,KAAK,UAAU;AACb,gBAAM,SAAS,CAAC,MAAe,OAAO,SAAS,CAAC,IAAI,IAAI;AACxD,gBAAM,QAAQ;AAAA,YACZ,KAAK,oBAAoB,KAAK;AAAA,UAChC;AACA,yBAAe,UAAU;AAAA,YACvB,cAAc,KAAK;AAAA,YACnB,OAAO,KAAK;AAAA,YACZ;AAAA,UACF,CAAC;AACD,gBAAM;AAAA,YACJ,MAAM;AAAA,YACN,cAAc,KAAK;AAAA,YACnB,OAAO,KAAK,QACR;AAAA,cACE,cAAc,OAAO,KAAK,MAAM,YAAY;AAAA,cAC5C,kBAAkB,OAAO,KAAK,MAAM,gBAAgB;AAAA,cACpD,aAAa,OAAO,KAAK,MAAM,WAAW;AAAA,cAC1C,GAAG;AAAA,YACL,IACA;AAAA,UACN;AACA;AAAA,QACF;AAAA,QACA,KAAK,SAAS;AAEZ,cAAI,KAAK,aAAa,QAAS;AAC/B,gBAAM,QAAQ,KAAK,iBAAiB,QAAQ,KAAK,QAAQ,IAAI,MAAM,OAAO,KAAK,KAAK,CAAC;AACrF,yBAAe,cAAc,EAAE,UAAU,QAAQ,OAAO,qBAAqB,KAAK,EAAE,CAAC;AACrF,gBAAM,EAAE,MAAM,SAAS,MAAM;AAC7B;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAAA,EACF,UAAE;AACA,SAAK,aAAa,oBAAoB,SAAS,OAAO;AAGtD,eAAW,MAAM;AAAA,EACnB;AACF;","names":[]}