@skaile/workspaces 6.1.0 → 6.2.0

package/dist/{chunk-KUQAJVYT.js → chunk-3HACK2Z7.js}

@@ -1,7 +1,7 @@
-import { selectableEntries, entryModelIds, entrySurface, entryBaseUrl } from './chunk-GNZBFSRB.js';
-import { entryCredentialValue } from './chunk-ILLTBKDJ.js';
-import { listenWithProbe, loadHarnessSdkSubpath, loadHarnessSdk } from './chunk-VMVDAZLA.js';
-import { parseModelRef } from './chunk-LSI2IVFT.js';
+import { selectableEntries, entryModelIds, entrySurface, entryBaseUrl } from './chunk-BOT64P3U.js';
+import { entryCredentialValue } from './chunk-X37JUNEW.js';
+import { listenWithProbe, loadHarnessSdkSubpath, loadHarnessSdk } from './chunk-BRK5QZKI.js';
+import { parseModelRef } from './chunk-RXGPFAMN.js';
 import { createHash, timingSafeEqual, randomUUID } from 'crypto';
 import { createServer } from 'http';
 
@@ -177,6 +177,35 @@ function pickSamplingParams(body, keys) {
   }
   return Object.keys(picked).length > 0 ? picked : void 0;
 }
+var PASSTHROUGH_OBJECT_KEYS = ["response_format", "usage", "provider"];
+var RESPONSE_FORMAT_TYPES = ["text", "json_object", "json_schema"];
+function pickPassthroughObjects(body) {
+  const picked = {};
+  for (const key of PASSTHROUGH_OBJECT_KEYS) {
+    const value = body[key];
+    if (value === void 0 || value === null) continue;
+    switch (key) {
+      case "response_format":
+        if (!isRecord(value) || !RESPONSE_FORMAT_TYPES.includes(value.type)) {
+          throw badRequest("response_format.type must be text, json_object or json_schema");
+        }
+        if (value.type === "json_schema" && (!isRecord(value.json_schema) || typeof value.json_schema.name !== "string" || value.json_schema.name.length === 0)) {
+          throw badRequest("response_format.json_schema.name is required");
+        }
+        break;
+      case "usage":
+        if (!isRecord(value) || typeof value.include !== "boolean") {
+          throw badRequest("usage.include must be a boolean");
+        }
+        break;
+      case "provider":
+        if (!isRecord(value)) throw badRequest("provider must be an object");
+        break;
+    }
+    picked[key] = value;
+  }
+  return Object.keys(picked).length > 0 ? picked : void 0;
+}
 function buildOptions(parts) {
   const options = {};
   if (parts.temperature !== void 0) options.temperature = parts.temperature;
@@ -364,6 +393,10 @@ function parseTools(value) {
   }
   return tools;
 }
+function mergeSamplingParams(...parts) {
+  const merged = Object.assign({}, ...parts);
+  return Object.keys(merged).length > 0 ? merged : void 0;
+}
 function parseChatCompletionRequest(body) {
   if (!isRecord(body)) throw badRequest("request body must be a JSON object");
   const model = requireModel(body);
@@ -424,13 +457,10 @@ function parseChatCompletionRequest(body) {
       body.max_completion_tokens ?? body.max_tokens,
       "max_completion_tokens"
     ),
-    samplingParams: pickSamplingParams(body, [
-      "top_p",
-      "frequency_penalty",
-      "presence_penalty",
-      "seed",
-      "stop"
-    ]),
+    samplingParams: mergeSamplingParams(
+      pickSamplingParams(body, ["top_p", "frequency_penalty", "presence_penalty", "seed", "stop"]),
+      pickPassthroughObjects(body)
+    ),
     reasoning: parseReasoningEffort(body.reasoning_effort, "reasoning_effort")
   });
   return {
@@ -454,7 +484,7 @@ function finishReason(stopReason, errorMessage) {
       return "stop";
   }
 }
-function toChatUsage(usage) {
+function toChatUsage(usage, cost) {
   const prompt = promptTokens(usage);
   const result = {
     prompt_tokens: prompt,
@@ -464,9 +494,10 @@ function toChatUsage(usage) {
   };
   if (usage.reasoning !== void 0)
     result.completion_tokens_details = { reasoning_tokens: usage.reasoning };
+  if (cost !== void 0) result.cost = cost;
   return result;
 }
-function toChatCompletion(message, meta) {
+function toChatCompletion(message, meta, cost) {
   const reason = finishReason(message.stopReason, message.errorMessage);
   let text = "";
   let thinking = "";
@@ -491,7 +522,7 @@ function toChatCompletion(message, meta) {
     created: meta.created,
     model: meta.model,
     choices: [{ index: 0, message: out, finish_reason: reason }],
-    usage: toChatUsage(message.usage)
+    usage: toChatUsage(message.usage, cost)
   };
 }
 function createChatStreamState(init) {
@@ -548,7 +579,7 @@ function chatCompletionChunks(event, state) {
           created: state.created,
           model: state.model,
           choices: [],
-          usage: toChatUsage(event.message.usage)
+          usage: toChatUsage(event.message.usage, state.cost)
         });
       }
       return chunks;
@@ -994,8 +1025,70 @@ function responsesEvents(event, state) {
   }
 }
 
+// model-endpoint/src/usage-cost.ts
+function costOf(payload) {
+  if (typeof payload !== "object" || payload === null || !("usage" in payload)) return void 0;
+  const { usage } = payload;
+  if (typeof usage !== "object" || usage === null || !("cost" in usage)) return void 0;
+  const { cost } = usage;
+  return typeof cost === "number" && Number.isFinite(cost) ? cost : void 0;
+}
+async function sseCost(body) {
+  const reader = body.getReader();
+  const decoder = new TextDecoder();
+  let pending = "";
+  let cost;
+  const scan = (line) => {
+    if (!line.startsWith("data:") || !line.includes('"cost"')) return;
+    try {
+      cost = costOf(JSON.parse(line.slice("data:".length))) ?? cost;
+    } catch {
+    }
+  };
+  try {
+    for (; ; ) {
+      const { done, value } = await reader.read();
+      if (done) break;
+      pending += decoder.decode(value, { stream: true });
+      const lines = pending.split("\n");
+      pending = lines.pop() ?? "";
+      for (const line of lines) scan(line.endsWith("\r") ? line.slice(0, -1) : line);
+    }
+    scan(pending + decoder.decode());
+    return cost;
+  } catch {
+    return void 0;
+  }
+}
+function createCostCapture(base = globalThis.fetch) {
+  const reads = [];
+  const capturing = (async (input, init) => {
+    const response = await base(input, init);
+    const type = response.headers.get("content-type") ?? "";
+    if (type.startsWith("text/event-stream") && response.body) {
+      const [forCaller, forCost] = response.body.tee();
+      reads.push(sseCost(forCost));
+      return new Response(forCaller, response);
+    }
+    if (type.startsWith("application/json") && response.body) {
+      reads.push(
+        response.clone().json().then(costOf, () => void 0)
+      );
+    }
+    return response;
+  });
+  return {
+    fetch: capturing,
+    async cost() {
+      const costs = await Promise.all(reads);
+      return costs.reduce((last, cost) => cost ?? last, void 0);
+    }
+  };
+}
+
 // model-endpoint/src/server.ts
 var MAX_BODY_BYTES = 10 * 1024 * 1024;
+var COST_WAIT_MS = 2e3;
 var EMBEDDING_SURFACES = /* @__PURE__ */ new Set(["openai-completions", "openai-responses"]);
 async function startModelEndpoint(options) {
   if (!options.token) throw new Error("model endpoint requires a non-empty token");
@@ -1016,7 +1109,7 @@ async function startModelEndpoint(options) {
       throw new ModelEndpointError(401, "invalid or missing bearer token", "invalid_api_key");
     }
   }
-  async function startUpstream(request, resolved, signal) {
+  async function startUpstream(request, resolved, signal, captureCost = false) {
     const pi = await loadPi(projectDir);
     const surface = resolved.model.api;
     let api;
@@ -1030,33 +1123,46 @@ async function startModelEndpoint(options) {
       );
     }
     const { temperature, maxTokens, samplingParams, reasoning } = request.options;
+    const capture = captureCost && surface === "openai-completions" ? createCostCapture(options.fetch ?? globalThis.fetch) : void 0;
     const common = {
       ...resolved.apiKey !== void 0 ? { apiKey: resolved.apiKey } : {},
       ...resolved.headers ? { headers: resolved.headers } : {},
       signal,
+      ...capture ? { fetch: capture.fetch } : {},
       ...temperature !== void 0 ? { temperature } : {},
       ...maxTokens !== void 0 ? { maxTokens } : {},
       ...samplingParams !== void 0 ? { samplingParams } : {}
     };
     const context = pi.normalizeContext(request.context);
+    const cost = () => capture ? withCeiling(capture.cost(), COST_WAIT_MS) : Promise.resolve(void 0);
     if (reasoning !== void 0 && api.streamSimple) {
-      return api.streamSimple(resolved.model, context, {
-        ...common,
-        reasoning
-      });
+      return {
+        events: api.streamSimple(resolved.model, context, {
+          ...common,
+          reasoning
+        }),
+        cost
+      };
     }
-    return api.stream(resolved.model, context, common);
+    return {
+      events: api.stream(resolved.model, context, common),
+      cost
+    };
   }
   async function handleChat(req, res, signal, trace) {
     const request = parseChatCompletionRequest(await readJson(req));
     const resolved = resolveModel(request.model, deps());
     trace.target = `${resolved.providerId}/${resolved.logicalId}`;
-    const events = await startUpstream(request, resolved, signal);
+    const { events, cost } = await startUpstream(request, resolved, signal, true);
     const id = `chatcmpl-${randomUUID().replaceAll("-", "")}`;
     const created = Math.floor(Date.now() / 1e3);
     if (!request.stream) {
       const message = await collectFinal(events);
-      sendJson(res, 200, toChatCompletion(message, { id, created, model: request.model }));
+      sendJson(
+        res,
+        200,
+        toChatCompletion(message, { id, created, model: request.model }, await cost())
+      );
       return;
     }
     const state = createChatStreamState({
@@ -1066,7 +1172,10 @@ async function startModelEndpoint(options) {
       includeUsage: request.includeUsage
     });
     await pipeSse(res, events, {
-      frames: (event) => chatCompletionChunks(event, state),
+      frames: async (event) => {
+        if (event.type === "done" && state.includeUsage) state.cost = await cost();
+        return chatCompletionChunks(event, state);
+      },
       errorFrame: (message) => ({ error: { message, type: "upstream_error" } }),
       terminator: "[DONE]"
     });
@@ -1075,7 +1184,7 @@ async function startModelEndpoint(options) {
     const request = parseResponsesRequest(await readJson(req));
     const resolved = resolveModel(request.model, deps());
     trace.target = `${resolved.providerId}/${resolved.logicalId}`;
-    const events = await startUpstream(request, resolved, signal);
+    const { events } = await startUpstream(request, resolved, signal);
     const id = `resp_${randomUUID().replaceAll("-", "")}`;
     const created = Math.floor(Date.now() / 1e3);
     if (!request.stream) {
@@ -1257,6 +1366,11 @@ async function collectFinal(events) {
     "upstream_incomplete"
   );
 }
+function withCeiling(promise, ms) {
+  const { promise: timeout, resolve } = Promise.withResolvers();
+  const timer = setTimeout(resolve, ms);
+  return Promise.race([promise, timeout]).finally(() => clearTimeout(timer));
+}
 function upstreamFailure(message) {
   return new ModelEndpointError(
     502,
@@ -1279,7 +1393,7 @@ async function pipeSse(res, events, pipe) {
         await sendAll([pipe.errorFrame(event.error.errorMessage ?? `upstream ${event.reason}`)]);
         break;
       }
-      await sendAll(pipe.frames(event));
+      await sendAll(await pipe.frames(event));
     }
   } catch (err) {
     if (!sse) throw err;
@@ -1385,5 +1499,5 @@ function errorText(err) {
 }
 
 export { ModelEndpointError, checkPiAiInstalled, listModels, resolveModel, startModelEndpoint };
-//# sourceMappingURL=chunk-KUQAJVYT.js.map
-//# sourceMappingURL=chunk-KUQAJVYT.js.map
+//# sourceMappingURL=chunk-3HACK2Z7.js.map
+//# sourceMappingURL=chunk-3HACK2Z7.js.map