@odla-ai/harness 0.5.1 → 0.6.0

package/dist/code-runtime-cli.js

@@ -6,7 +6,7 @@ import {
   assertCodeBuildRecipe,
   createCodeRuntimeControlClient,
   runCodeRuntimeHeartbeatLoop
-} from "./chunk-L2T3LPEF.js";
+} from "./chunk-ZYNGL5SC.js";
 import {
   assertPinnedImage,
   selectContainerEngine

package/dist/node.cjs

@@ -1966,6 +1966,9 @@ async function materializeCommandWorkspace(input) {
   }
 }
 
+// src/code-runtime-attempt.ts
+var import_ai2 = require("@odla-ai/ai");
+
 // src/code-agent.ts
 var import_ai = require("@odla-ai/ai");
 
@@ -2017,7 +2020,12 @@ function codeSkill(opts) {
       { lease: opts.lease, workspaceDir: opts.workspaceDir, signal },
       { requestId: `bench-${tool}-${++seq}`, tool, input }
     );
-    opts.onToolCall?.({ tool, ok: response2.ok, durationMs: Date.now() - startedAt });
+    opts.onToolCall?.({
+      tool,
+      ok: response2.ok,
+      durationMs: Date.now() - startedAt,
+      ...response2.ok ? {} : { error: String(response2.content).slice(0, 300) }
+    });
     return { content: response2.content, isError: !response2.ok };
   };
   const read2 = {
@@ -2166,6 +2174,7 @@ async function runCodeAgent(options) {
 // src/code-runtime-attempt.ts
 async function runCodeAgentAttempt(options) {
   try {
+    const surface = options.surface ?? "v2";
     const { run } = await runCodeAgent({
       inference: options.inference,
       broker: options.broker,
@@ -2175,17 +2184,31 @@ async function runCodeAgentAttempt(options) {
       // The brokered route resolves the real model from platform policy; this
       // id only labels the request the control plane is about to rewrite.
       model: "brokered",
-      surface: options.surface ?? "v2",
+      surface,
       ...options.maxSteps === void 0 ? {} : { maxSteps: options.maxSteps },
       ...options.budget ? { budget: options.budget } : {},
       ...options.signal ? { signal: options.signal } : {},
       ...options.onToolCall ? { onToolCall: options.onToolCall } : {}
     });
+    let finalText = run.finalText.trim();
+    if (!finalText && run.stoppedReason !== "refusal") {
+      const closing = await options.inference.chat({
+        model: "brokered",
+        system: `${SYSTEM_PROMPT_FOR[surface]}
+
+Finish with a concise, non-empty answer to the owner. Do not call tools or promise future work.`,
+        messages: [...run.messages, { role: "user", content: "Give the owner the closing answer now, grounded in the repository evidence and tool results above." }],
+        maxTokens: 16384,
+        ...options.signal ? { signal: options.signal } : {}
+      });
+      finalText = (0, import_ai2.extractText)(closing.content).trim();
+    }
+    const missingClosing = !finalText && run.stoppedReason !== "refusal";
     return {
-      status: run.stoppedReason === "refusal" ? "failed" : "completed",
-      finalText: run.finalText,
+      status: run.stoppedReason === "refusal" || missingClosing ? "failed" : "completed",
+      finalText,
       stoppedReason: run.stoppedReason,
-      ...run.stoppedReason === "refusal" ? { error: run.finalText || "the agent refused the task" } : {}
+      ...run.stoppedReason === "refusal" ? { error: finalText || "the agent refused the task" } : missingClosing ? { error: "the Code agent did not produce a closing answer" } : {}
     };
   } catch (cause) {
     const error = (cause instanceof Error ? cause.message : String(cause)).slice(0, 2e3);
@@ -2195,31 +2218,7 @@ async function runCodeAgentAttempt(options) {
 
 // src/code-runtime-inference.ts
 async function handleCodeRuntimeInference(input) {
-  const { command, metadata, request, state } = input;
-  if (state.tokens >= metadata.maxTokensPerInteraction) {
-    if (!state.noticeEmitted) {
-      state.noticeEmitted = true;
-      await input.event({
-        type: "message",
-        actor: "system",
-        body: `The agent paused at the ${metadata.maxTokensPerInteraction.toLocaleString("en-US")}-token per-interaction limit. Send a new instruction to continue.`
-      }).catch(() => void 0);
-    }
-    return {
-      protocolVersion: HARNESS_PROTOCOL_VERSION,
-      type: "inference.response",
-      requestId: request.requestId,
-      response: {
-        id: `budget:${command.commandId}`,
-        provider: "openai",
-        model: "interaction-budget",
-        role: "assistant",
-        content: [{ type: "text", text: "Pause now. The owner-set token limit for this interaction has been reached." }],
-        stopReason: "end_turn",
-        usage: { inputTokens: 0, outputTokens: 0 }
-      }
-    };
-  }
+  const { command, request, state } = input;
   const startedAt = Date.now();
   const response2 = await input.control.infer(command.sessionId, {
     requestId: request.requestId,
@@ -2239,7 +2238,6 @@ async function handleCodeRuntimeInference(input) {
     durationMs: Date.now() - startedAt,
     interactionId: command.commandId,
     interactionTokens: state.tokens,
-    interactionMaxTokens: metadata.maxTokensPerInteraction,
     ...costUsd === void 0 ? {} : { costUsd },
     ...state.costKnown ? { interactionCostUsd: state.costUsd } : {}
   }).catch(() => void 0);
@@ -3370,7 +3368,7 @@ var CodePiRuntimeEngine = class {
       recipeAuthorization: this.options.recipeAuthorization
     }, lease, metadata.role));
     const startedAt = Date.now();
-    const interaction = { tokens: 0, noticeEmitted: false, costUsd: 0, costKnown: true };
+    const interaction = { tokens: 0, costUsd: 0, costKnown: true };
     const inference = createCodeRuntimeInference({
       command,
       metadata,
@@ -3385,31 +3383,30 @@ var CodePiRuntimeEngine = class {
       lease,
       workspaceDir: active.workspace.workspaceDir,
       prompt: metadata.prompt,
-      signal: active.abort.signal,
-      // The owner's per-interaction allowance, enforced by runAgent against
-      // INCREMENTAL usage. The control plane still reserves against the same
-      // ceiling, but this is what stops the loop cleanly at the boundary rather
-      // than letting it discover the limit through a synthesized pause reply.
-      budget: { maxTotalTokens: metadata.maxTokensPerInteraction }
+      signal: active.abort.signal
     });
-    const body = result.finalText.trim() || (result.status === "completed" ? "The agent finished without a closing message." : result.error ?? "The agent failed.");
+    const closing = result.finalText.trim();
+    const completed = result.status === "completed" && Boolean(closing);
+    const detail = result.error?.trim() || (closing ? "the Code agent failed" : "the Code agent did not produce a closing answer");
+    const body = closing || detail;
     await this.#event(command, {
       type: "message",
-      actor: result.status === "completed" ? "agent" : "system",
+      actor: completed ? "agent" : "system",
       body
     }, active.conversationRefs).catch(() => void 0);
     await this.#event(command, {
       type: "status",
-      status: result.status === "completed" ? "idle" : "failed",
+      status: completed ? "idle" : "failed",
       durationMs: Date.now() - startedAt
     }, active.conversationRefs).catch(() => void 0);
-    if (result.status === "failed") {
-      const detail = (result.error ?? "").trim() || "the Code agent failed";
+    if (!completed) {
       await this.#diagnostic(command, active, detail);
       await this.#failure(command, active, detail);
     }
     return {
       ...result,
+      status: completed ? "completed" : "failed",
+      ...!completed ? { error: detail } : {},
       tokens: interaction.tokens,
       ...interaction.costKnown ? { costUsd: interaction.costUsd } : {}
     };