@apo-ai/sdk 0.14.0 → 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

package/dist/agent-task/public.js

@@ -16,13 +16,13 @@ import {
 import "../chunk-HRYXLRXJ.js";
 import {
   createApoOpenAI
-} from "../chunk-A3UYQ7M4.js";
+} from "../chunk-NSMKSPBU.js";
 import {
   createApoAnthropic
-} from "../chunk-X73JGMK3.js";
+} from "../chunk-4BQSCD6T.js";
 import {
   servingHostFromBaseURL
-} from "../chunk-JTWQD27M.js";
+} from "../chunk-JN7RHBTS.js";
 
 // src/agent-task/run/steer-scheduler.ts
 var TASK_STEER_SPAN_NAME = "task.steer";
@@ -691,7 +691,7 @@ var RepetitionGuard = class {
     this.tail = "";
     this.lastChar = "";
     this.charRun = 0;
-    this.periodRuns = new Array(MAX_UNIT_CHARS + 1).fill(0);
+    this.periodRuns = Array.from({ length: MAX_UNIT_CHARS + 1 }, () => 0);
     // Low-diversity state: the current raw line, and the folded window built
     // from finished lines that are not tables.
     this.line = "";
@@ -872,9 +872,19 @@ var openAIProvider = {
   apiRootExample: "https://api.openai.com/v1",
   reserved: "The OpenAI Decisions API provider is not implemented yet: the endpoint is limited-preview and its schema is unpublished. " + OPENROUTER_RESERVED_HINT + "."
 };
+var anthropicProvider = {
+  decisionsURL: (chatBaseURL) => `${chatBaseURL.replace(/\/v1\/?$/, "")}/decisions`,
+  requestBody: () => {
+    throw new Error("unreachable \u2014 reserved guard returns before the call");
+  },
+  parseResponse: () => ({ error: "unreachable \u2014 reserved guard returns before the call" }),
+  apiRootExample: "https://api.anthropic.com",
+  reserved: "Anthropic-direct serves no decisions API \u2014 the second judge cannot run there. " + OPENROUTER_RESERVED_HINT + "."
+};
 var PROVIDERS = {
   openrouter: openRouterProvider,
-  openai: openAIProvider
+  openai: openAIProvider,
+  anthropic: anthropicProvider
 };
 function resolveSecondJudgeModel() {
   const value = process.env.APO_SECOND_JUDGE_MODEL?.trim();
@@ -900,7 +910,10 @@ function resolveSecondJudgeProvider(baseURL) {
     }
     return provider;
   }
-  return /(^|\.)api\.openai\.com$/.test(new URL(baseURL).hostname) ? openAIProvider : openRouterProvider;
+  const hostname = new URL(baseURL).hostname;
+  if (/(^|\.)api\.openai\.com$/.test(hostname)) return openAIProvider;
+  if (/(^|\.)api\.anthropic\.com$/.test(hostname)) return anthropicProvider;
+  return openRouterProvider;
 }
 var INPUT_TOO_LARGE_MARKERS = [
   "max_tokens_exceeded",
@@ -1049,10 +1062,12 @@ ${formatValue(v)}`).join("\n\n");
 }
 function parseJudgeUsage(usage) {
   if (!usage) return void 0;
+  const anthropicShape = usage.cache_creation_input_tokens !== void 0 || usage.cache_read_input_tokens !== void 0;
   const cacheCreation = usage.cache_creation_input_tokens ?? usage.prompt_tokens_details?.cache_write_tokens;
   const cacheRead = usage.cache_read_input_tokens ?? usage.prompt_tokens_details?.cached_tokens;
+  const prompt = usage.prompt_tokens ?? 0;
   const tokens = {
-    input: usage.prompt_tokens ?? 0,
+    input: anthropicShape ? prompt + (cacheCreation ?? 0) + (cacheRead ?? 0) : prompt,
     output: usage.completion_tokens ?? 0
   };
   if (typeof cacheCreation === "number") tokens.cache_creation = cacheCreation;
@@ -1089,6 +1104,7 @@ function judgeVerdict(raw) {
   };
 }
 var JUDGE_IDLE_TIMEOUT_MS = 9e4;
+var ANTHROPIC_JUDGE_MAX_TOKENS = 8192;
 function envMs(name, fallback) {
   const raw = process.env[name]?.trim();
   if (!raw) return fallback;
@@ -1121,12 +1137,14 @@ async function readCompletion(response, onData, guards) {
     return {
       text: data.choices?.[0]?.message?.content ?? "",
       usage: data.usage,
-      finishReason: data.choices?.[0]?.finish_reason ?? void 0
+      finishReason: data.choices?.[0]?.finish_reason ?? void 0,
+      servedModel: typeof data.model === "string" && data.model ? data.model : void 0
     };
   }
   let text = "";
   let usage;
   let finishReason;
+  let servedModel;
   let eventType = "message";
   let dataLines = [];
   let finished = false;
@@ -1162,6 +1180,7 @@ async function readCompletion(response, onData, guards) {
     }
     if (choice?.finish_reason) finishReason = choice.finish_reason;
     if (chunk.usage) usage = chunk.usage;
+    if (typeof chunk.model === "string" && chunk.model) servedModel = chunk.model;
   };
   const handleLine = (line) => {
     if (line === "") return dispatch();
@@ -1174,7 +1193,7 @@ async function readCompletion(response, onData, guards) {
     else if (field === "event") eventType = value;
   };
   const reader = response.body?.getReader();
-  if (!reader) return { text, usage, finishReason };
+  if (!reader) return { text, usage, finishReason, servedModel };
   const decoder = new TextDecoder();
   let buffer = "";
   for (; ; ) {
@@ -1185,13 +1204,29 @@ async function readCompletion(response, onData, guards) {
     buffer = lines.pop() ?? "";
     for (const line of lines) {
       handleLine(line);
-      if (finished) return { text, usage, finishReason };
+      if (finished) return { text, usage, finishReason, servedModel };
     }
   }
   buffer += decoder.decode();
   for (const line of buffer.split(/\r\n|\r|\n/)) if (line) handleLine(line);
   dispatch();
-  return { text, usage, finishReason };
+  return { text, usage, finishReason, servedModel };
+}
+async function readAnthropicCompletion(response) {
+  const data = await response.json();
+  const text = (data.content ?? []).filter((block) => block.type === "text" && typeof block.text === "string").map((block) => block.text).join("");
+  const finishReason = data.stop_reason === "max_tokens" ? "length" : data.stop_reason === "refusal" ? "content_filter" : data.stop_reason;
+  return {
+    text,
+    finishReason,
+    servedModel: typeof data.model === "string" && data.model ? data.model : void 0,
+    usage: data.usage ? {
+      prompt_tokens: data.usage.input_tokens,
+      completion_tokens: data.usage.output_tokens,
+      ...data.usage.cache_creation_input_tokens !== void 0 ? { cache_creation_input_tokens: data.usage.cache_creation_input_tokens } : {},
+      ...data.usage.cache_read_input_tokens !== void 0 ? { cache_read_input_tokens: data.usage.cache_read_input_tokens } : {}
+    } : void 0
+  };
 }
 var prefixLanes = /* @__PURE__ */ new Map();
 function judgeConcurrency() {
@@ -1262,26 +1297,58 @@ function runWithSharedPrefix(key, task2) {
   };
   return lane.warmed.then(acquire).then(() => task2().finally(release));
 }
-function judgeTemperature(configured) {
+function judgeTemperature(configured, wire = "openai-compatible") {
+  const max = wire === "anthropic" ? 1 : 2;
   const raw = process.env.APO_JUDGE_TEMPERATURE?.trim();
   const fromEnv = raw ? Number(raw) : void 0;
   const value = configured ?? fromEnv ?? 0;
-  if (!Number.isFinite(value) || value < 0 || value > 2) {
+  if (!Number.isFinite(value) || value < 0 || value > max) {
     const source = configured !== void 0 ? "judge.temperature" : "APO_JUDGE_TEMPERATURE";
-    throw new Error(`${source} must be a number between 0 and 2, got ${configured ?? raw}`);
+    throw new Error(`${source} must be a number between 0 and ${max}, got ${configured ?? raw}`);
   }
   return value;
 }
-function defaultJudgeBaseURL(baseURL) {
+function isAnthropicHost(baseURL) {
+  try {
+    return /(^|\.)api\.anthropic\.com$/.test(new URL(baseURL).hostname);
+  } catch {
+    return false;
+  }
+}
+function resolveJudgeWire(opts) {
+  if (opts.provider === "anthropic") return "anthropic";
+  if (opts.baseURL !== void 0 && isAnthropicHost(opts.baseURL)) return "anthropic";
+  if (opts.baseURL === void 0 && /^claude-[a-z0-9]/.test(opts.model)) return "anthropic";
+  return "openai-compatible";
+}
+function anthropicAuth(apiKey) {
+  if (apiKey) return { apiKey, asBearer: false };
+  if (process.env.ANTHROPIC_API_KEY) return { apiKey: process.env.ANTHROPIC_API_KEY, asBearer: false };
+  if (process.env.ANTHROPIC_AUTH_TOKEN) return { apiKey: process.env.ANTHROPIC_AUTH_TOKEN, asBearer: true };
+  return { apiKey: void 0, asBearer: false };
+}
+function defaultJudgeBaseURL(baseURL, wire = "openai-compatible") {
+  if (wire === "anthropic") {
+    return baseURL ?? process.env.ANTHROPIC_BASE_URL ?? "https://api.anthropic.com";
+  }
   return baseURL ?? process.env.OPENROUTER_BASE_URL ?? "https://openrouter.ai/api/v1";
 }
-function defaultJudgeAPIKey(apiKey) {
+function defaultJudgeAPIKey(apiKey, wire = "openai-compatible") {
+  if (wire === "anthropic") {
+    return anthropicAuth(apiKey).apiKey;
+  }
   return apiKey ?? process.env.OPENROUTER_API_KEY ?? process.env.OPENAI_API_KEY;
 }
 async function callJudge(args) {
-  const baseURL = defaultJudgeBaseURL(args.baseURL);
-  const apiKey = defaultJudgeAPIKey(args.apiKey);
-  const temperature = judgeTemperature(args.temperature);
+  const wire = resolveJudgeWire({
+    baseURL: args.baseURL,
+    model: args.model,
+    provider: args.provider
+  });
+  const baseURL = defaultJudgeBaseURL(args.baseURL, wire);
+  const auth = wire === "anthropic" ? anthropicAuth(args.apiKey) : { apiKey: void 0, asBearer: false };
+  const apiKey = wire === "anthropic" ? auth.apiKey : defaultJudgeAPIKey(args.apiKey);
+  const temperature = judgeTemperature(args.temperature, wire);
   const { briefingText, instructionText, systemPromptText, deliverableText, secondJudgeState, secondJudgeProjected } = buildJudgePromptParts({
     values: args.values,
     instruction: args.instruction,
@@ -1298,6 +1365,21 @@ async function callJudge(args) {
     ...secondJudgeProjected ? { projected: true } : {}
   }) : void 0;
   const cacheKey = `${args.model}\0${briefingText}\0${deliverableText}`;
+  const anthropicRequestBody = () => JSON.stringify({
+    model: args.model,
+    max_tokens: ANTHROPIC_JUDGE_MAX_TOKENS,
+    temperature,
+    system: [
+      { type: "text", text: briefingText },
+      { type: "text", text: deliverableText, cache_control: { type: "ephemeral" } }
+    ],
+    messages: [{ role: "user", content: instructionText }]
+  });
+  const anthropicHeaders = {
+    "Content-Type": "application/json",
+    "anthropic-version": "2023-06-01",
+    ...apiKey !== void 0 ? auth.asBearer ? { Authorization: `Bearer ${apiKey}` } : { "x-api-key": apiKey } : {}
+  };
   const requestBody = (stream, schema) => JSON.stringify({
     model: args.model,
     prompt_cache_key: promptCacheKey(cacheKey),
@@ -1348,7 +1430,16 @@ async function callJudge(args) {
     try {
       let response;
       try {
-        response = await fetchWithPromptCacheKey(`${baseURL}/chat/completions`, {
+        response = wire === "anthropic" ? (
+          // Plain fetch: the Anthropic body never carries prompt_cache_key,
+          // so the resend wrapper would be a pass-through.
+          await fetch(`${baseURL}/v1/messages`, {
+            method: "POST",
+            signal: controller.signal,
+            headers: anthropicHeaders,
+            body: anthropicRequestBody()
+          })
+        ) : await fetchWithPromptCacheKey(`${baseURL}/chat/completions`, {
           method: "POST",
           signal: controller.signal,
           headers: {
@@ -1378,7 +1469,10 @@ async function callJudge(args) {
         content: new RepetitionGuard({ diversity: false })
       } : void 0;
       try {
-        return { kind: "completion", completion: await readCompletion(response, onData, guards) };
+        return {
+          kind: "completion",
+          completion: wire === "anthropic" ? await readAnthropicCompletion(response) : await readCompletion(response, onData, guards)
+        };
       } catch (error) {
         if (error instanceof JudgeReasoningLoopError) {
           return { kind: "unavailable", error, retryable: true, loop: true };
@@ -1438,6 +1532,9 @@ async function callJudge(args) {
     }
     const judgeMetadata = async (completion) => ({
       model: args.model,
+      // The model that actually judged, when the gateway fell back to a
+      // different one than requested.
+      ...completion?.servedModel && completion.servedModel !== args.model ? { served_model: completion.servedModel } : {},
       temperature,
       contract: judgeContractInUse(),
       prompt: { system: systemPromptText, user: instructionText },
@@ -1555,11 +1652,16 @@ async function cascadePreflight(args) {
     ...args.context ? { context: args.context } : {},
     ...args.secondJudgeValue !== void 0 ? { secondJudgeValue: args.secondJudgeValue } : {}
   });
+  const wire = resolveJudgeWire({
+    baseURL: args.effective.baseURL,
+    model: args.effective.model,
+    provider: args.effective.provider
+  });
   const evidence = await callSecondJudge({
     state: parts.secondJudgeState,
     model,
-    baseURL: resolveSecondJudgeBaseURL(defaultJudgeBaseURL(args.effective.baseURL)),
-    apiKey: resolveSecondJudgeAPIKey(defaultJudgeAPIKey(args.effective.apiKey)),
+    baseURL: resolveSecondJudgeBaseURL(defaultJudgeBaseURL(args.effective.baseURL, wire)),
+    apiKey: resolveSecondJudgeAPIKey(defaultJudgeAPIKey(args.effective.apiKey, wire)),
     ...parts.secondJudgeProjected ? { projected: true } : {}
   });
   return { kind: "ran", evidence, parts };
@@ -1962,6 +2064,7 @@ var DEFAULT_BUDGET = {
   timeoutMs: 3e5,
   maxReadBytes: 2 * 1024 * 1024
 };
+var ANTHROPIC_AGENT_MAX_TOKENS = 8192;
 var MAX_READ_WINDOW = 12e3;
 var SEARCH_HITS = 8;
 var SEARCH_CONTEXT = 400;
@@ -2297,29 +2400,50 @@ async function runAgentSession(spec) {
   const budget = { ...DEFAULT_BUDGET, ...spec.budget };
   const started = Date.now();
   const ai = await import("ai");
-  const { createOpenAICompatible } = await import("@ai-sdk/openai-compatible");
   const { z } = await import("zod");
   const { generateText } = ai;
   const stepCountIs = (n) => ai.stepCountIs(n);
   const tool = (def) => ai.tool(def);
-  const provider = createOpenAICompatible({
-    name: "openrouter",
-    baseURL: spec.baseURL ?? "https://openrouter.ai/api/v1",
-    fetch: fetchWithPromptCacheKey,
-    apiKey: spec.apiKey ?? process.env.OPENROUTER_API_KEY ?? process.env.OPENAI_API_KEY,
-    // Mark the briefing prefix cacheable — repeated samples (rejudge) re-bill
-    // nothing for the shared turn-0 context — and keep every step of one
-    // session on the replica holding its growing prefix. The key is per
-    // session (briefing + criterion): sessions run concurrently, unlike
-    // t.judge calls, and one key per task would pile them onto one replica.
-    transformRequestBody: (body) => ({
-      ...body,
-      prompt_cache_key: promptCacheKey(sessionPrefix(body)),
-      messages: body.messages?.map(
-        (m) => m.role === "system" && typeof m.content === "string" ? { ...m, content: [{ type: "text", text: m.content, cache_control: { type: "ephemeral" } }] } : m
-      )
-    })
+  const wire = resolveJudgeWire({
+    baseURL: spec.baseURL,
+    model: spec.model,
+    provider: spec.provider
   });
+  let model;
+  let wireOptions;
+  if (wire === "anthropic") {
+    const { createAnthropic } = await import("@ai-sdk/anthropic");
+    const auth = anthropicAuth(spec.apiKey);
+    const base = (spec.baseURL ?? process.env.ANTHROPIC_BASE_URL)?.replace(/\/+$/, "");
+    const anthropicEngine = createAnthropic({
+      ...base ? { baseURL: /\/v\d+$/.test(base) ? base : `${base}/v1` } : {},
+      ...auth.apiKey ? auth.asBearer ? { authToken: auth.apiKey } : { apiKey: auth.apiKey } : {}
+    });
+    model = anthropicEngine(spec.model);
+    wireOptions = { maxOutputTokens: ANTHROPIC_AGENT_MAX_TOKENS };
+  } else {
+    const { createOpenAICompatible } = await import("@ai-sdk/openai-compatible");
+    const provider = createOpenAICompatible({
+      name: "openrouter",
+      baseURL: spec.baseURL ?? "https://openrouter.ai/api/v1",
+      fetch: fetchWithPromptCacheKey,
+      apiKey: spec.apiKey ?? process.env.OPENROUTER_API_KEY ?? process.env.OPENAI_API_KEY,
+      // Mark the briefing prefix cacheable — repeated samples (rejudge) re-bill
+      // nothing for the shared turn-0 context — and keep every step of one
+      // session on the replica holding its growing prefix. The key is per
+      // session (briefing + criterion): sessions run concurrently, unlike
+      // t.judge calls, and one key per task would pile them onto one replica.
+      transformRequestBody: (body) => ({
+        ...body,
+        prompt_cache_key: promptCacheKey(sessionPrefix(body)),
+        messages: body.messages?.map(
+          (m) => m.role === "system" && typeof m.content === "string" ? { ...m, content: [{ type: "text", text: m.content, cache_control: { type: "ephemeral" } }] } : m
+        )
+      })
+    });
+    model = provider(spec.model);
+    wireOptions = {};
+  }
   const ledger = { toolCallsUsed: 0, readBytesUsed: 0, manifest: [] };
   const budgetGuard = () => {
     ledger.toolCallsUsed += 1;
@@ -2382,8 +2506,17 @@ async function runAgentSession(spec) {
   try {
     try {
       const result = await generateText({
-        model: provider(spec.model),
-        system,
+        model,
+        // On the Anthropic wire the briefing's cache breakpoint is a per-part
+        // provider option on the system message (the request-body rewrite the
+        // OpenAI-compatible path uses has no equivalent there).
+        system: wire === "anthropic" ? [{
+          role: "system",
+          content: system,
+          // Message-level providerOptions: the Anthropic engine reads
+          // the cache breakpoint from here, not from content parts.
+          providerOptions: { anthropic: { cacheControl: { type: "ephemeral" } } }
+        }] : system,
         prompt: spec.instruction + exhibitsBlock,
         temperature: 0,
         toolChoice: "required",
@@ -2391,7 +2524,8 @@ async function runAgentSession(spec) {
         tools,
         abortSignal: AbortSignal.timeout(budget.timeoutMs),
         maxRetries: 2,
-        onStepFinish: onStep
+        onStepFinish: onStep,
+        ...wireOptions
       });
       engineUsage = result.usage;
     } catch (error) {
@@ -2449,7 +2583,7 @@ function createAgentMethod(rec, judgeConfig, judgeScope, evidence, judgeTracer,
       rec.record(
         label,
         false,
-        "No judge model configured for t.agent. Set one of:\n\u2022 OPENROUTER_MODEL + OPENROUTER_API_KEY (OpenRouter \u2014 needs a tool-calling-capable model)\n\u2022 OPENAI_MODEL + OPENAI_API_KEY (OpenAI direct)\nOr pass { judge } to runTask() programmatically.",
+        "No judge model configured for t.agent. Set one of:\n\u2022 OPENROUTER_MODEL + OPENROUTER_API_KEY (OpenRouter \u2014 needs a tool-calling-capable model)\n\u2022 OPENAI_MODEL + OPENAI_API_KEY (OpenAI direct)\n\u2022 ANTHROPIC_MODEL + ANTHROPIC_API_KEY (Anthropic direct \u2014 needs a tool-calling-capable model; or ANTHROPIC_AUTH_TOKEN for plan credits)\nOr pass { judge } to runTask() programmatically.",
         { evaluator_type: "agent", location, outcome: "error" }
       );
       return;
@@ -2459,6 +2593,7 @@ function createAgentMethod(rec, judgeConfig, judgeScope, evidence, judgeTracer,
       model: effective.model,
       baseURL: effective.baseURL,
       apiKey: effective.apiKey,
+      provider: effective.provider,
       evidence: evidence ?? { deliverables: {} },
       scope: judgeScope,
       exhibits,
@@ -2552,6 +2687,7 @@ function resolveJudgeConfig(judgeConfig, override) {
     model,
     baseURL: override?.baseURL ?? judgeConfig?.baseURL,
     apiKey: override?.apiKey ?? judgeConfig?.apiKey,
+    provider: override?.provider ?? judgeConfig?.provider,
     prompt: override?.prompt ?? judgeConfig?.prompt,
     mode: override?.mode ?? judgeConfig?.mode,
     temperature: override?.temperature ?? judgeConfig?.temperature
@@ -2657,14 +2793,21 @@ function createJudgeMethod(rec, judgeConfig, judgeScope, judgeTracer) {
       rec.record(
         label,
         false,
-        "No judge model configured. Set one of:\n\u2022 OPENROUTER_MODEL + OPENROUTER_API_KEY (OpenRouter \u2014 works with 200+ models, one account)\n\u2022 OPENAI_MODEL + OPENAI_API_KEY (OpenAI direct)\nOr pass { judge } to runTask() programmatically.",
+        "No judge model configured. Set one of:\n\u2022 OPENROUTER_MODEL + OPENROUTER_API_KEY (OpenRouter \u2014 works with 200+ models, one account)\n\u2022 OPENAI_MODEL + OPENAI_API_KEY (OpenAI direct)\n\u2022 ANTHROPIC_MODEL + ANTHROPIC_API_KEY (Anthropic direct \u2014 or ANTHROPIC_AUTH_TOKEN for plan credits)\nOr pass { judge } to runTask() programmatically.",
         // No judge means no verdict — the same bucket as an unreachable judge.
         { evaluator_type: "llm", location, outcome: "error" }
       );
       return;
     }
     try {
-      judgeTemperature(effective.temperature);
+      judgeTemperature(
+        effective.temperature,
+        resolveJudgeWire({
+          baseURL: effective.baseURL,
+          model: effective.model,
+          provider: effective.provider
+        })
+      );
     } catch (error) {
       rec.record(label, false, `judge failed: ${error instanceof Error ? error.message : String(error)}`, {
         evaluator_type: "llm",
@@ -2730,6 +2873,7 @@ function createJudgeMethod(rec, judgeConfig, judgeScope, judgeTracer) {
       model: effective.model,
       baseURL: effective.baseURL,
       apiKey: effective.apiKey,
+      provider: effective.provider,
       prompt: effective.prompt,
       ...effective.temperature !== void 0 ? { temperature: effective.temperature } : {},
       ...context2 ? { context: context2 } : {},
@@ -2761,8 +2905,21 @@ function createJudgeMethod(rec, judgeConfig, judgeScope, judgeTracer) {
           };
         },
         usage: (r) => {
-          const tokens = r?.judge?.tokens;
-          return tokens ? { prompt_tokens: tokens.input, completion_tokens: tokens.output } : void 0;
+          const judgeMeta = r?.judge;
+          const tokens = judgeMeta?.tokens;
+          return {
+            ...tokens ? {
+              prompt_tokens: tokens.input,
+              completion_tokens: tokens.output,
+              // Without these the backend prices every cached
+              // prefix read at the full input rate.
+              ...tokens.cache_read !== void 0 ? { cache_read_tokens: tokens.cache_read } : {},
+              ...tokens.cache_creation !== void 0 ? { cache_creation_tokens: tokens.cache_creation } : {}
+            } : {},
+            // The model that actually judged — a gateway fallback
+            // diverges from the requested judge model.
+            ...judgeMeta?.served_model ? { served_model: judgeMeta.served_model } : {}
+          };
         }
       },
       async (spanId) => {
@@ -6619,6 +6776,18 @@ function createOtelAgentTaskTraceClient(config) {
     if (params.completion_tokens !== void 0) {
       active.span.setAttribute("gen_ai.usage.output_tokens", params.completion_tokens);
     }
+    if (params.cache_read_tokens !== void 0 || params.cache_creation_tokens !== void 0) {
+      active.span.setAttribute("apo.usage.input_includes_cache", true);
+      if (params.cache_read_tokens !== void 0) {
+        active.span.setAttribute("gen_ai.usage.cache_read.input_tokens", params.cache_read_tokens);
+      }
+      if (params.cache_creation_tokens !== void 0) {
+        active.span.setAttribute(
+          "gen_ai.usage.cache_creation.input_tokens",
+          params.cache_creation_tokens
+        );
+      }
+    }
     if (params.reasoning_tokens !== void 0) {
       active.span.setAttribute(
         "gen_ai.usage.reasoning.output_tokens",
@@ -6628,6 +6797,12 @@ function createOtelAgentTaskTraceClient(config) {
     if (params.served_model) {
       active.span.setAttribute("gen_ai.response.model", params.served_model);
     }
+    if (params.provider) {
+      active.span.setAttribute("gen_ai.provider.name", params.provider);
+    }
+    if (params.route) {
+      active.span.setAttribute("apo.llm.route", params.route);
+    }
     if (params.status_message) {
       active.span.setAttribute("apo.status_message", params.status_message);
     }
@@ -6961,12 +7136,23 @@ async function loadTaskRuntime(_taskDir) {
 }
 function resolveJudgeFromEnv() {
   const openRouterModel = process.env.OPENROUTER_MODEL;
-  const model = openRouterModel ?? process.env.OPENAI_MODEL;
+  const openAIModel = process.env.OPENAI_MODEL;
+  const anthropicModel = process.env.ANTHROPIC_MODEL;
+  const model = openRouterModel ?? openAIModel ?? anthropicModel;
   if (!model) return void 0;
   return {
     model,
-    baseURL: openRouterModel ? process.env.OPENROUTER_BASE_URL : process.env.OPENAI_BASE_URL ?? "https://api.openai.com/v1",
-    apiKey: openRouterModel ? process.env.OPENROUTER_API_KEY : process.env.OPENAI_API_KEY,
+    baseURL: openRouterModel ? process.env.OPENROUTER_BASE_URL : openAIModel ? process.env.OPENAI_BASE_URL ?? "https://api.openai.com/v1" : process.env.ANTHROPIC_BASE_URL ?? "https://api.anthropic.com",
+    // ANTHROPIC_AUTH_TOKEN deliberately stays out of apiKey: an explicit
+    // config key is always sent as x-api-key, which 401s for the OAuth-style
+    // token. Leaving apiKey unset lets anthropicAuth pick the token up from
+    // the env and send it as Authorization: Bearer.
+    apiKey: openRouterModel ? process.env.OPENROUTER_API_KEY : openAIModel ? process.env.OPENAI_API_KEY : process.env.ANTHROPIC_API_KEY,
+    // Anthropic is last in the precedence chain (OPENROUTER > OPENAI >
+    // ANTHROPIC): a workspace that already configures one of the others
+    // keeps its behavior, so the arm can be reached only deliberately.
+    // Reaching this return with neither set means ANTHROPIC_MODEL won.
+    ...openRouterModel || openAIModel ? {} : { provider: "anthropic" },
     // "cascade" turns on second-judge-first judging (a confident decision
     // model verdict stands without calling the primary judge); any other
     // value is ignored so a typo degrades to the default dual mode, never