@apo-ai/sdk 0.7.0 → 0.8.0

package/dist/agent-task/integrations/ai-sdk.d.ts

@@ -1,4 +1,4 @@
-import { A as AgentTaskTraceContext } from '../../tracing-DVLUqD-i.js';
+import { A as AgentTaskTraceContext } from '../../tracing-CevT0hQg.js';
 
 /**
  * Vercel AI SDK tracing integration for apo.

package/dist/agent-task/integrations/ai-sdk.js

@@ -2,8 +2,8 @@
 // Do not edit by hand — edit the src/ originals and rebuild.
 import {
   createApoTracer
-} from "../../chunk-QHRWPVPZ.js";
-import "../../chunk-HWZVL6TS.js";
+} from "../../chunk-CKB3SUDO.js";
+import "../../chunk-HLRXQAZ5.js";
 export {
   createApoTracer
 };

package/dist/agent-task/integrations/anthropic.d.ts

@@ -1,4 +1,4 @@
-import { A as AgentTaskTraceContext } from '../../tracing-DVLUqD-i.js';
+import { A as AgentTaskTraceContext } from '../../tracing-CevT0hQg.js';
 
 /**
  * Anthropic JS SDK tracing integration for apo.

package/dist/agent-task/integrations/anthropic.js

@@ -2,8 +2,8 @@
 // Do not edit by hand — edit the src/ originals and rebuild.
 import {
   createApoAnthropic
-} from "../../chunk-SZDSIPXC.js";
-import "../../chunk-GDB6SDGN.js";
+} from "../../chunk-43OI6GP7.js";
+import "../../chunk-3OQ3JBFJ.js";
 export {
   createApoAnthropic
 };

package/dist/agent-task/integrations/openai.d.ts

@@ -1,4 +1,4 @@
-import { A as AgentTaskTraceContext } from '../../tracing-DVLUqD-i.js';
+import { A as AgentTaskTraceContext } from '../../tracing-CevT0hQg.js';
 
 /**
  * OpenAI JS SDK tracing integration for apo.
@@ -22,7 +22,7 @@ import { A as AgentTaskTraceContext } from '../../tracing-DVLUqD-i.js';
  *     { trace, parentSpanId, taskId: ctx.task.id, turnNumber }
  *   );
  *   const response = await client.chat.completions.create({
- *     model: "google/gemini-2.5-flash-lite",
+ *     model: "deepseek/deepseek-v4.1-flash",
  *     messages: [{ role: "user", content: String(turn) }],
  *     tools: [...],
  *   });

package/dist/agent-task/integrations/openai.js

@@ -2,8 +2,8 @@
 // Do not edit by hand — edit the src/ originals and rebuild.
 import {
   createApoOpenAI
-} from "../../chunk-6ZRW5IW6.js";
-import "../../chunk-GDB6SDGN.js";
+} from "../../chunk-3B3L2H4Y.js";
+import "../../chunk-3OQ3JBFJ.js";
 export {
   createApoOpenAI
 };

package/dist/agent-task/public.d.ts

@@ -1,4 +1,4 @@
-import { A as AgentTaskTraceContext, a as AgentTaskTraceOptions } from '../tracing-DVLUqD-i.js';
+import { A as AgentTaskTraceContext, a as AgentTaskTraceOptions } from '../tracing-CevT0hQg.js';
 export { CreateApoTracerOptions, createApoTracer } from './integrations/ai-sdk.js';
 export { CreateApoOpenAIOptions, createApoOpenAI } from './integrations/openai.js';
 export { CreateApoAnthropicOptions, createApoAnthropic } from './integrations/anthropic.js';
@@ -313,15 +313,18 @@ type TypedAdapterDefinition<TName extends string, TDeliverables extends Record<s
  *   it must never silently pass. Value assertions (``t.check``) and LLM
  *   assertions (``t.judge``) do not consult trace capabilities and retain
  *   their existing pass/fail behavior.
+ * - ``"error"`` → ``pass=false``. An LLM judge produced no verdict: it was
+ *   unreachable, returned an HTTP error, or sent an empty or truncated reply
+ *   (after one retry). The check's quality is unknown, not failed.
  */
-type AssertionOutcome = "pass" | "fail" | "unsupported";
+type AssertionOutcome = "pass" | "fail" | "unsupported" | "error";
 /**
  * Metadata about an LLM judge call. Populated by evaluators that use an
  * LLM to make their pass/fail decision. All fields optional so code-only
  * evaluators can leave this undefined.
  */
 type JudgeMetadata = {
-    /** Model identifier, e.g. ``"google/gemini-2.5-flash-lite"``. */
+    /** Model identifier, e.g. ``"deepseek/deepseek-v4.1-flash"``. */
     model?: string;
     /**
      * Which response contract elicited this judgment (#163): reasoning-first
@@ -357,6 +360,15 @@ type JudgeMetadata = {
      * content-hashed evidence manifest (replay audit backbone).
      */
     session?: AgentJudgeSession;
+    /**
+     * Trace span id of this judgment itself (issue #288): the span the trace
+     * context opens around the judge call / agent session. Lets any surface
+     * holding this metadata deep-link into the judge's span in the trace view
+     * (`/traces/{trace_run_id}?observation={span_id}`) instead of re-finding
+     * it by name. Absent when the run is untraced or the span id can't be
+     * trusted (the noop trace context's sentinel).
+     */
+    span_id?: string;
     /**
      * Second-grader evidence: a typed-decision model (Jev via OpenRouter's
      * `/alpha/decisions`) graded the same deliverable + instruction alongside
@@ -508,6 +520,16 @@ type EvaluationItemResult = {
     id: string;
     pass: boolean;
     reasoning: string;
+    /**
+     * The outcome category, rolled up from this check's assertions. Set only
+     * when the check failed and *every* failing assertion failed for lack of
+     * a verdict: ``"error"`` — the judge never answered (unreachable, HTTP
+     * error, empty reply), so the check's quality is unknown, not failed;
+     * ``"unsupported"`` — the trace projection lacked the evidence. Absent
+     * when the check passed or genuinely failed on evidence, so a real FAIL
+     * is never masked by an incidental judge error alongside it (issue #323).
+     */
+    outcome?: AssertionOutcome;
     /** The rubric instruction from the task definition ("PASS if …"). */
     instruction?: string;
     /** Name of the deliverable this check was evaluated against. */
@@ -782,7 +804,7 @@ type ToolCallOptions = {
  * `judge` layer of its own; see {@link resolveJudgeConfig} for precedence.
  */
 type JudgeConfig = {
-    /** Model id, e.g. ``"google/gemini-2.5-flash-lite"``. */
+    /** Model id, e.g. ``"deepseek/deepseek-v4.1-flash"``. */
     model: string;
     /** Override the OpenRouter/OpenAI-compatible base URL. */
     baseURL?: string;
@@ -1404,11 +1426,19 @@ interface RejudgeOptions {
     /** Progress sink for human-facing operators (CLI prints to stderr). */
     onProgress?: (message: string) => void;
 }
-/** Per-check pass counts across samples — "fails 2 of 5", not "failed". */
+/**
+ * Per-check pass counts across samples — "fails 2 of 5", not "failed".
+ *
+ * Samples where the check errored (judge produced no verdict) are tracked
+ * in {@link errored} and count as neither pass nor fail: an unreachable
+ * judge says nothing about stability, so it must not read as a flaky fail.
+ */
 interface RejudgeCheckStability {
     check_id: string;
     passes: number;
     samples: number;
+    /** Samples that produced no verdict (judge error) — issue #323. */
+    errored?: number;
 }
 interface RejudgeOutcome {
     runId: string;