@apo-ai/sdk 0.13.0 → 0.15.0

package/README.md

@@ -250,8 +250,11 @@ For apo tasks the flow is built automatically from the run's trace
 projection — no extra wiring.
 
 **`t.judge(value, instruction)` — LLM-backed assertions.** Configure the judge
-with `runTask(dir, { judge: { model, apiKey?, baseURL? } })`, or set
-`OPENROUTER_MODEL`/`OPENAI_MODEL` for the CLI runtime. Judge verdicts use the
+with `runTask(dir, { judge: { model, apiKey?, baseURL?, provider? } })`, or set
+`OPENROUTER_MODEL`/`OPENAI_MODEL`/`ANTHROPIC_MODEL` for the CLI runtime. A bare
+`claude-*` model id (or an Anthropic base URL) switches the call to the
+Anthropic Messages wire, authenticated with `ANTHROPIC_API_KEY` or
+`ANTHROPIC_AUTH_TOKEN`. Judge verdicts use the
 same recorder and result format as code assertions, including model, prompt,
 response, token usage, and latency metadata.
 

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

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

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

@@ -1,4 +1,4 @@
-import { A as AgentTaskTraceContext } from '../../tracing-FqFObrmt.js';
+import { A as AgentTaskTraceContext } from '../../tracing-CEucJqQI.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-X73JGMK3.js";
-import "../../chunk-JTWQD27M.js";
+} from "../../chunk-4BQSCD6T.js";
+import "../../chunk-JN7RHBTS.js";
 export {
   createApoAnthropic
 };

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

@@ -1,4 +1,4 @@
-import { A as AgentTaskTraceContext } from '../../tracing-FqFObrmt.js';
+import { A as AgentTaskTraceContext } from '../../tracing-CEucJqQI.js';
 
 /**
  * OpenAI JS SDK tracing integration for apo.
@@ -56,6 +56,9 @@ interface OpenAIClientLike {
     chat: {
         completions: {
             create: (...args: any[]) => Promise<any>;
+            withRawResponse?: {
+                create: (...args: any[]) => Promise<any>;
+            };
         };
     };
     [key: string]: any;

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-A3UYQ7M4.js";
-import "../../chunk-JTWQD27M.js";
+} from "../../chunk-NSMKSPBU.js";
+import "../../chunk-JN7RHBTS.js";
 export {
   createApoOpenAI
 };

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

@@ -1,4 +1,4 @@
-import { A as AgentTaskTraceContext, a as AgentTaskTraceOptions } from '../tracing-FqFObrmt.js';
+import { A as AgentTaskTraceContext, a as AgentTaskTraceOptions } from '../tracing-CEucJqQI.js';
 export { CreateApoTracerOptions, createApoTracer } from './integrations/ai-sdk.js';
 export { CreateApoOpenAIOptions, createApoOpenAI } from './integrations/openai.js';
 export { CreateApoAnthropicOptions, createApoAnthropic } from './integrations/anthropic.js';
@@ -83,6 +83,7 @@ interface TraceProjectionObservation {
     errorMessage?: string;
     input?: unknown;
     output?: unknown;
+    /** The model the caller requested (often a gateway alias). */
     model?: string;
     toolName?: string;
     toolParameters?: unknown;
@@ -615,6 +616,12 @@ type AssertionOutcome = "pass" | "fail" | "unsupported" | "error";
 type JudgeMetadata = {
     /** Model identifier, e.g. ``"deepseek/deepseek-v4.1-flash"``. */
     model?: string;
+    /**
+     * The model the provider reports as having served the judge call (the
+     * response's `model` field). Diverges from `model` when the judge's
+     * gateway fell back — judge spend and attribution key on this.
+     */
+    served_model?: string;
     /**
      * Which response contract elicited this judgment (#163): reasoning-first
      * (the default) or the legacy verdict-first (``pass`` before
@@ -1004,6 +1011,8 @@ declare function similarity(expected: string, threshold?: number): Matcher<unkno
  * LLM-as-judge call. Used by `t.judge(values, instruction)` to evaluate
  * deliverables against a natural-language rubric. Calls an OpenAI-compatible
  * endpoint (OpenRouter, OpenAI, etc.) via fetch and parses the verdict.
+ * Anthropic-direct (`api.anthropic.com` Messages API) speaks its own wire —
+ * see {@link resolveJudgeWire}.
  */
 
 /**
@@ -1294,6 +1303,13 @@ type JudgeConfig = {
     baseURL?: string;
     /** Override the API key (defaults to env). */
     apiKey?: string;
+    /**
+     * Force the Anthropic Messages wire (`api.anthropic.com/v1/messages`) for a
+     * base URL whose hostname does not identify it — an Anthropic-protocol
+     * gateway. Unset = auto-detection (Anthropic host, or a bare `claude-*`
+     * model id); see `resolveJudgeWire` in judge.ts.
+     */
+    provider?: "anthropic";
     /**
      * Build a custom judge briefing from the threaded context (#161). The SDK
      * keeps ownership of the JSON response contract — see
@@ -1307,6 +1323,13 @@ type JudgeConfig = {
      * dual behavior (second judge as parallel, evidence-only shadow).
      */
     mode?: "cascade";
+    /**
+     * Sampling temperature for `t.judge` requests (not `t.agent`, whose session
+     * stays at 0). Unset falls back to `APO_JUDGE_TEMPERATURE`, then `0`. Some
+     * reasoning models recommend a non-zero value (DeepSeek: 0.6) to avoid
+     * repetition at greedy decoding. Recorded on the check's judge metadata.
+     */
+    temperature?: number;
 };
 /**
  * What a check knows about its scope when it invokes the judge — the