@apo-ai/sdk 0.14.0 → 0.16.0

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
@@ -658,6 +665,22 @@ type JudgeMetadata = {
      * trusted (the noop trace context's sentinel).
      */
     span_id?: string;
+    /**
+     * What the judge itself concluded about the deliverable — the raw verdict,
+     * distinct from the check's `pass` (which is agreement with the author's
+     * ground truth under `expect` polarity). Identical to the check outcome
+     * when no `expect` was set; deliberately opposite on a correctly-judged
+     * `expect: "fail"` check. Recorded so downstream consumers — the history
+     * plane above all — can state what prior judges verdicted without
+     * parsing reasoning prose.
+     */
+    verdict_pass?: boolean;
+    /**
+     * The author's ground-truth polarity for this check (`expect`), when one
+     * was pinned. Together with `verdict_pass` it makes the check outcome
+     * reconstructible: outcome = verdict_pass matches expected_verdict.
+     */
+    expected_verdict?: "pass" | "fail";
     /**
      * Second-grader evidence: a typed-decision model (Jev via OpenRouter's
      * `/alpha/decisions`) graded the same deliverable + instruction alongside
@@ -1004,6 +1027,8 @@ 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}.
  */
 
 /**
@@ -1055,7 +1080,16 @@ type HistoryRunDetail = {
     status: string;
     checks: {
         id: string;
+        /** Check outcome — agreement with the author's ground truth. */
         pass: boolean;
+        /**
+         * What the judge itself concluded. Equal to `pass` without `expect`;
+         * deliberately opposite on a correctly-judged `expect: "fail"` check.
+         * Absent on pre-polarity recordings and non-judged (code) checks.
+         */
+        verdict_pass?: boolean;
+        /** The author's ground truth (`expect`), when one was pinned. */
+        expected_verdict?: "pass" | "fail";
         evaluator: string;
         reasoning?: string;
         assertions?: {
@@ -1294,6 +1328,13 @@ 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
@@ -1771,8 +1812,15 @@ type SecondJudgeFacts = {
  * types the CLI and dashboard read from the backend are all assignable.
  */
 type SecondJudgeFactsInput = {
+    /**
+     * The check outcome. Under `expect` polarity this is agreement with the
+     * author's ground truth, not the judge's verdict — the verdict comparison
+     * must read `judge.verdict_pass` instead.
+     */
     pass: boolean;
     judge?: {
+        /** What the judge itself concluded; absent on runs recorded before it. */
+        verdict_pass?: boolean;
         secondJudge?: {
             choice?: "pass" | "fail";
             confidence?: number;
@@ -1785,6 +1833,44 @@ declare function secondJudgeFacts(check: SecondJudgeFactsInput): SecondJudgeFact
 /** One honest sentence about the relation; null when corroborated. */
 declare function secondJudgeTakeaway(check: SecondJudgeFactsInput): string | null;
 
+/**
+ * The judge's compute tool: an arithmetic expression in, an exact value out.
+ *
+ * This is an expression evaluator, not a code-execution sandbox — the
+ * distinction is the security model. The grammar below has numbers, four
+ * arithmetic operations, remainder, exponentiation, parentheses, a fixed
+ * allowlist of math functions, and two constants. A model cannot express
+ * anything else: no strings, no variables, no property access, no loops, no
+ * function definitions. There is no I/O to forbid and nothing to time out —
+ * evaluation is a single linear pass whose only recursion is bounded by the
+ * expression-length cap, with parenthesization nesting tracked explicitly.
+ *
+ * Failures are honest return values (an `{ error }` with the allowlist, the
+ * cap, or the offending operation in the message) so the judge can rephrase,
+ * never exceptions past the tool boundary.
+ */
+/** A successful evaluation: the raw IEEE-754 double and a display form. */
+type ComputeSuccess = {
+    value: number;
+    text: string;
+};
+/** Every way an expression fails to evaluate. Never thrown. */
+type ComputeResult = ComputeSuccess | {
+    error: string;
+};
+/** Input-length cap; re-checked here although the tool schema also enforces it. */
+declare const COMPUTE_MAX_EXPR = 512;
+/** Maximum parenthesization / function-argument nesting depth. */
+declare const COMPUTE_MAX_DEPTH = 64;
+/**
+ * Evaluate one arithmetic expression. Pure: no I/O, no ambient state, never
+ * throws. `text` normalizes the double to at most 12 significant digits so a
+ * judge comparing against human-stated figures ("3.5%") is not eyeballing
+ * IEEE remainder digits (`3.4013605442176873` → `"3.40136054422"`,
+ * `0.30000000000000004` → `"0.3"`); `value` keeps the exact double.
+ */
+declare function evaluateComputeExpression(expression: string): ComputeResult;
+
 /**
  * Check submission compaction (issue #175).
  *
@@ -2613,4 +2699,4 @@ type TranscriptCaptureResult = {
 };
 declare function replayAdapterTranscript(capture: AdapterTranscriptCapture, options: ReplayAdapterTranscriptOptions): Promise<TranscriptCaptureResult>;
 
-export { type AISDKResult, type AdapterDefinition, type AdapterRuntimeState, type AdapterSession, type AdapterTranscriptCapture, type AgentBudget, type AgentEvidence, type AgentJudgeOptions, type AgentJudgeOutcome, type AgentJudgeSession, type AgentJudgeStep, type AgentProgressEvent, type AgentTaskCliOptions, type AgentTaskRunConfiguration, AgentTaskRunError, type AgentTaskRunSummary, type AgentTaskRuntime, AgentTaskTraceContext, AgentTaskTraceOptions, type AgentTurnResult, type AnthropicMessage, type ApoRunContext, ApoSpanProcessor, type ArtifactUploadConfig, type CheckCompaction, type CheckContext, type CheckGroup, type CleanupContext, type CollectDeliverablesContext, type CollectedDeliverables, type DeliverableDefinition, type DeliverableSummary, type DeliverableValidationResult, type DescribeRegistration, type DurationBudgetOptions, type EvidenceFingerprint, type ExportOtlpTracesOptions, type FileArtifact, type FileArtifactOptions, type FileEntry, type Flow, type FlowEvent, FlowView, type InitializeContext, JUDGE_SEGMENT_LIMIT, type JudgeConfig, type JudgeContext, type JudgePromptBuilder, type JudgeScope, type JudgeToolsConfig, type LoadedTask, type ManifestFile, type ManifestFileInput, type Matcher, type McpHttpTransport, type McpServerConfig, type McpStdioTransport, type ModeClass, NO_CHECKS_REGISTERED_MESSAGE, type NameMatcher, type OpenAIMessage, type OtlpAttribute, type OtlpAttributeValue, type OtlpSpan, type OtlpTracesPayload, type ParsedTranscriptSession, type PreparedTaskResult, RECEIVED_VALUE_LIMIT, type RegisterApoTracingOptions, type RejudgeCheckStability, type RejudgeEndpoints, RejudgeError, type RejudgeOptions, type RejudgeOutcome, type ReplayAdapterTranscriptOptions, type RunTaskOptions, type SecondJudgeFacts, type SecondJudgeFactsInput, type StartSessionContext, type SteerContext, type SteerDeliveryBoundary, type SteerRecord, type SteerResult, type SteerSpec, type SteerTrigger, TEST_METHOD_NAMES, type TaskConfig, type TaskDefinition, type TaskEvaluationResult, TaskFiles, type TaskRevisionManifestV1, type TaskRunResult, type TaskScope, type TaskTranscript, type TaskTranscriptTurn, type TestContext, type TestRegistration, type TokenBudgetOptions, type ToolCallOptions, type ToolCallStatus, type TranscriptCaptureResult, TranscriptReplayError, type TranscriptReplayOptions, type TranscriptSource, type TranscriptToolCall, type TranscriptTurn, type TranscriptUsage, type TurnContext, type TurnFn, type TurnRecord, type TypedAdapterDefinition, type ValueMatcher, buildApoAuthHeaders, buildManifest, canonicalManifestJson, compactChecksForSubmission, connectMcpServers, contentSha256, defineAdapter, defineTask, describe, detectTranscriptSource, discoverAgentTaskDirs, equals, expandSecretPlaceholders, exportOtlpTraces, fileArtifact, filePaths, fromAISDK, fromAnthropicMessages, fromOpenAIMessages, getActiveApoRun, getTaskSteers, getTaskTurn, includes, isFileArtifact, loadTask, loadTaskRuntime, matchValue, matches, normalizeManifestPath, parseAgentTaskCliArgs, parseClaudeCodeTranscript, parseCodexTranscript, persistFileArtifacts, registerApoTracing, rejudgeTaskRun, replayAdapterTranscript, resetApoTracing, resetFlowChecks, resetTaskRegistry, resetTaskSteers, resetTaskTurn, resolveMcpServerPaths, resolveOtlpTracesUrl, resolveTurn, runAgentTaskCli, runTask, runTaskDir, satisfies, secondJudgeFacts, secondJudgeTakeaway, sha256Hex, similarity, steer, task, defineCheck as test, transcriptSessionToOtlp, turn, withApoRun, withApoRunSync };
+export { type AISDKResult, type AdapterDefinition, type AdapterRuntimeState, type AdapterSession, type AdapterTranscriptCapture, type AgentBudget, type AgentEvidence, type AgentJudgeOptions, type AgentJudgeOutcome, type AgentJudgeSession, type AgentJudgeStep, type AgentProgressEvent, type AgentTaskCliOptions, type AgentTaskRunConfiguration, AgentTaskRunError, type AgentTaskRunSummary, type AgentTaskRuntime, AgentTaskTraceContext, AgentTaskTraceOptions, type AgentTurnResult, type AnthropicMessage, type ApoRunContext, ApoSpanProcessor, type ArtifactUploadConfig, COMPUTE_MAX_DEPTH, COMPUTE_MAX_EXPR, type CheckCompaction, type CheckContext, type CheckGroup, type CleanupContext, type CollectDeliverablesContext, type CollectedDeliverables, type ComputeResult, type DeliverableDefinition, type DeliverableSummary, type DeliverableValidationResult, type DescribeRegistration, type DurationBudgetOptions, type EvidenceFingerprint, type ExportOtlpTracesOptions, type FileArtifact, type FileArtifactOptions, type FileEntry, type Flow, type FlowEvent, FlowView, type InitializeContext, JUDGE_SEGMENT_LIMIT, type JudgeConfig, type JudgeContext, type JudgePromptBuilder, type JudgeScope, type JudgeToolsConfig, type LoadedTask, type ManifestFile, type ManifestFileInput, type Matcher, type McpHttpTransport, type McpServerConfig, type McpStdioTransport, type ModeClass, NO_CHECKS_REGISTERED_MESSAGE, type NameMatcher, type OpenAIMessage, type OtlpAttribute, type OtlpAttributeValue, type OtlpSpan, type OtlpTracesPayload, type ParsedTranscriptSession, type PreparedTaskResult, RECEIVED_VALUE_LIMIT, type RegisterApoTracingOptions, type RejudgeCheckStability, type RejudgeEndpoints, RejudgeError, type RejudgeOptions, type RejudgeOutcome, type ReplayAdapterTranscriptOptions, type RunTaskOptions, type SecondJudgeFacts, type SecondJudgeFactsInput, type StartSessionContext, type SteerContext, type SteerDeliveryBoundary, type SteerRecord, type SteerResult, type SteerSpec, type SteerTrigger, TEST_METHOD_NAMES, type TaskConfig, type TaskDefinition, type TaskEvaluationResult, TaskFiles, type TaskRevisionManifestV1, type TaskRunResult, type TaskScope, type TaskTranscript, type TaskTranscriptTurn, type TestContext, type TestRegistration, type TokenBudgetOptions, type ToolCallOptions, type ToolCallStatus, type TranscriptCaptureResult, TranscriptReplayError, type TranscriptReplayOptions, type TranscriptSource, type TranscriptToolCall, type TranscriptTurn, type TranscriptUsage, type TurnContext, type TurnFn, type TurnRecord, type TypedAdapterDefinition, type ValueMatcher, buildApoAuthHeaders, buildManifest, canonicalManifestJson, compactChecksForSubmission, connectMcpServers, contentSha256, defineAdapter, defineTask, describe, detectTranscriptSource, discoverAgentTaskDirs, equals, evaluateComputeExpression, expandSecretPlaceholders, exportOtlpTraces, fileArtifact, filePaths, fromAISDK, fromAnthropicMessages, fromOpenAIMessages, getActiveApoRun, getTaskSteers, getTaskTurn, includes, isFileArtifact, loadTask, loadTaskRuntime, matchValue, matches, normalizeManifestPath, parseAgentTaskCliArgs, parseClaudeCodeTranscript, parseCodexTranscript, persistFileArtifacts, registerApoTracing, rejudgeTaskRun, replayAdapterTranscript, resetApoTracing, resetFlowChecks, resetTaskRegistry, resetTaskSteers, resetTaskTurn, resolveMcpServerPaths, resolveOtlpTracesUrl, resolveTurn, runAgentTaskCli, runTask, runTaskDir, satisfies, secondJudgeFacts, secondJudgeTakeaway, sha256Hex, similarity, steer, task, defineCheck as test, transcriptSessionToOtlp, turn, withApoRun, withApoRunSync };