baro-ai 0.34.0 → 0.35.0

package/dist/cli.mjs

@@ -7672,17 +7672,6 @@ var AgenticEnvironment = class {
     this.isActive = false;
   }
 };
-var OpenAIInferenceRunner = class {
-  constructor() {
-    this.runtime = new OpenAIResponses();
-  }
-  async *run(context, model, signal) {
-    const response = await this.runtime.infer(new InferenceRequest(model, context));
-    for (const item of response.contextItems) {
-      yield item;
-    }
-  }
-};
 var Gpt55 = class {
   constructor() {
     this.specification = {
@@ -9876,6 +9865,58 @@ function extractVerdictJson(text) {
   throw new Error(`unbalanced JSON object in critic response: ${trimmed.slice(0, 200)}`);
 }
 
+// ../baro-orchestrator/src/planning/openai-runtime.ts
+var runtime = new OpenAIResponses();
+async function runInferenceRound(context, model) {
+  const response = await runtime.infer(new InferenceRequest(model, context));
+  return {
+    items: response.contextItems,
+    usage: response.tokenUsage
+  };
+}
+var UsageAccumulator = class {
+  input = 0;
+  output = 0;
+  total = 0;
+  cached = 0;
+  reasoning = 0;
+  rounds = 0;
+  add(usage) {
+    if (!usage) return;
+    this.rounds += 1;
+    this.input += usage.inputTokens ?? 0;
+    this.output += usage.outputTokens ?? 0;
+    this.total += usage.totalTokens ?? 0;
+    this.cached += usage.inputTokenDetails?.cached_tokens ?? 0;
+    this.reasoning += usage.outputTokenDetails?.reasoning_tokens ?? 0;
+  }
+  get isEmpty() {
+    return this.rounds === 0;
+  }
+  /**
+   * Plain-object snapshot suitable for embedding in
+   * `AgentResultItem.usage` (which is typed `any` to allow per-
+   * provider shapes). Keys are snake_case to line up with what the
+   * Claude side's stream-json mapper produces from Anthropic
+   * usage frames.
+   */
+  toJSON() {
+    return {
+      input_tokens: this.input,
+      output_tokens: this.output,
+      total_tokens: this.total,
+      cached_input_tokens: this.cached,
+      reasoning_tokens: this.reasoning,
+      rounds: this.rounds
+    };
+  }
+  /** One-line summary for the stderr / log path. */
+  summary() {
+    if (this.isEmpty) return "(no token usage reported)";
+    return `${this.total} total tokens (${this.input} in, ${this.output} out${this.cached ? `, ${this.cached} cached` : ""}${this.reasoning ? `, ${this.reasoning} reasoning` : ""}) across ${this.rounds} round(s)`;
+  }
+};
+
 // ../baro-orchestrator/src/participants/critic-openai.ts
 function pickModel(name) {
   switch (name) {
@@ -9896,7 +9937,6 @@ function pickModel(name) {
 var CriticOpenAI = class extends BaroParticipant {
   opts;
   model;
-  runner = new OpenAIInferenceRunner();
   emissions = /* @__PURE__ */ new Map();
   turnCount = /* @__PURE__ */ new Map();
   pending = /* @__PURE__ */ new Set();
@@ -9967,9 +10007,12 @@ var CriticOpenAI = class extends BaroParticipant {
     const userPrompt = buildEvalPrompt(criteria, resultText);
     const context = ModelContext.create("critic").addContextItem(SystemMessageItem.create(VERDICT_SYSTEM_PROMPT)).addContextItem(UserMessageItem.create(userPrompt));
     try {
+      const round = await runInferenceRound(context, this.model);
+      const usage = new UsageAccumulator();
+      usage.add(round.usage);
       let assistantText = "";
-      for await (const item of this.runner.run(context, this.model)) {
-        if (item.type === "message" && item.role === "assistant") {
+      for (const item of round.items) {
+        if (item.type === "message") {
           const json = item.toJSON();
           assistantText += json.content?.[0]?.text ?? "";
         }
@@ -9977,6 +10020,8 @@ var CriticOpenAI = class extends BaroParticipant {
       if (!assistantText.trim()) {
         throw new Error("OpenAI returned empty assistant text");
       }
+      process.stderr.write(`[critic-openai] ${usage.summary()}
+`);
       const verdictJson = extractVerdictJson(assistantText);
       const parsed = JSON.parse(verdictJson);
       return {
@@ -11266,7 +11311,6 @@ var OpenAIStoryAgent = class extends BaroParticipant {
   spec;
   opts;
   model;
-  runner = new OpenAIInferenceRunner();
   tools;
   envRef = null;
   currentPhase = "idle";
@@ -11381,6 +11425,7 @@ var OpenAIStoryAgent = class extends BaroParticipant {
     for (let turn = 1; turn <= this.spec.maxTurns; turn++) {
       const turnResult = await this.runOneTurn(context);
       context = turnResult.context;
+      const usageJson = turnResult.usage.isEmpty ? null : turnResult.usage.toJSON();
       this.envRef?.deliverBusEvent(
         this,
         new AgentResultItem(
@@ -11390,14 +11435,17 @@ var OpenAIStoryAgent = class extends BaroParticipant {
           // session id — not applicable for OpenAI
           !turnResult.success,
           turnResult.assistantText,
-          null,
-          // usage info — not surfaced this phase
+          usageJson,
           null,
           null,
           null,
           {}
         )
       );
+      process.stderr.write(
+        `[story-openai/${this.spec.id}] turn ${turn}: ${turnResult.usage.summary()}
+`
+      );
       if (!turnResult.success) {
         this.transition("failed", turnResult.error ?? "turn failed");
         return;
@@ -11428,19 +11476,27 @@ var OpenAIStoryAgent = class extends BaroParticipant {
     let context = initialContext;
     let assistantText = null;
     const perRoundMs = this.opts.perRoundTimeoutSecs * 1e3;
+    const usage = new UsageAccumulator();
     for (let round = 1; round <= this.opts.maxRoundsPerTurn; round++) {
-      const ac = new AbortController();
-      const timer = setTimeout(() => ac.abort(), perRoundMs);
       const calls = [];
       let sawMessage = false;
       let lastMessageText = null;
       try {
-        for await (const item of this.runner.run(context, this.model, ac.signal)) {
+        const roundPromise = runInferenceRound(context, this.model);
+        const timeoutPromise = new Promise(
+          (_, rej) => setTimeout(
+            () => rej(new Error(`round ${round} timed out after ${perRoundMs}ms`)),
+            perRoundMs
+          )
+        );
+        const result = await Promise.race([roundPromise, timeoutPromise]);
+        usage.add(result.usage);
+        for (const item of result.items) {
           if (item.type === "function_call") {
             await this.envRef?.deliverFunctionCall(this, item);
             context = context.addContextItem(item);
             calls.push(item);
-          } else if (item.type === "message" && item.role === "assistant") {
+          } else if (item.type === "message") {
             await this.envRef?.deliverModelMessage(this, item);
             context = context.addContextItem(item);
             const json = item.toJSON();
@@ -11452,15 +11508,13 @@ var OpenAIStoryAgent = class extends BaroParticipant {
           }
         }
       } catch (e) {
-        clearTimeout(timer);
         return {
           context,
           success: false,
           assistantText,
+          usage,
           error: `inference round ${round} failed: ${e?.message ?? String(e)}`
         };
-      } finally {
-        clearTimeout(timer);
       }
       for (const call of calls) {
         const tool = this.tools.find((t) => t.name === call.name);
@@ -11473,7 +11527,8 @@ var OpenAIStoryAgent = class extends BaroParticipant {
         return {
           context,
           success: true,
-          assistantText: lastMessageText
+          assistantText: lastMessageText,
+          usage
         };
       }
       if (!sawMessage && calls.length === 0) {
@@ -11481,6 +11536,7 @@ var OpenAIStoryAgent = class extends BaroParticipant {
           context,
           success: false,
           assistantText,
+          usage,
           error: `round ${round} returned no items`
         };
       }
@@ -11490,6 +11546,7 @@ var OpenAIStoryAgent = class extends BaroParticipant {
       context,
       success: false,
       assistantText,
+      usage,
       error: `exceeded maxRoundsPerTurn=${this.opts.maxRoundsPerTurn}`
     };
   }
@@ -11852,7 +11909,6 @@ function pickModel3(name) {
 var SurgeonOpenAI = class extends BaroParticipant {
   opts;
   model;
-  runner = new OpenAIInferenceRunner();
   replansEmitted = 0;
   pending = /* @__PURE__ */ new Set();
   constructor(opts) {
@@ -11896,9 +11952,12 @@ var SurgeonOpenAI = class extends BaroParticipant {
     const userPrompt = buildSurgeonPrompt(snap, failure);
     const context = ModelContext.create("surgeon").addContextItem(SystemMessageItem.create(SURGEON_SYSTEM_PROMPT)).addContextItem(UserMessageItem.create(userPrompt));
     try {
+      const round = await runInferenceRound(context, this.model);
+      const usage = new UsageAccumulator();
+      usage.add(round.usage);
       let assistantText = "";
-      for await (const item of this.runner.run(context, this.model)) {
-        if (item.type === "message" && item.role === "assistant") {
+      for (const item of round.items) {
+        if (item.type === "message") {
           const json = item.toJSON();
           assistantText += json.content?.[0]?.text ?? "";
         }
@@ -11906,6 +11965,8 @@ var SurgeonOpenAI = class extends BaroParticipant {
       if (!assistantText.trim()) {
         throw new Error("OpenAI returned empty assistant text");
       }
+      process.stderr.write(`[surgeon-openai] ${usage.summary()}
+`);
       const verdictJson = extractJsonObject(assistantText);
       const parsed = JSON.parse(verdictJson);
       if (parsed.action === "abort") return null;