@tangle-network/agent-eval 0.117.1 → 0.118.0

package/dist/storyboard/index.d.ts

@@ -1,4 +1,89 @@
-import { S as Span } from '../schema-B3Q3l9Z_.js';
+type SpanKind = 'agent' | 'llm' | 'tool' | 'retrieval' | 'judge' | 'sandbox' | 'custom';
+type SpanStatus = 'ok' | 'error';
+interface SpanBase {
+    spanId: string;
+    parentSpanId?: string;
+    runId: string;
+    kind: SpanKind;
+    name: string;
+    startedAt: number;
+    endedAt?: number;
+    status?: SpanStatus;
+    error?: string;
+    /** Anything not covered by typed fields. Kept deliberately free-form. */
+    attributes?: Record<string, unknown>;
+}
+interface Message {
+    role: 'system' | 'user' | 'assistant' | 'tool';
+    content: string;
+    tokens?: number;
+    /** Multi-modal content descriptors; blobs themselves live in Artifacts. */
+    images?: Array<{
+        artifactId?: string;
+        url?: string;
+        mime?: string;
+    }>;
+}
+interface LlmSpan extends SpanBase {
+    kind: 'llm';
+    model: string;
+    messages: Message[];
+    output?: string;
+    inputTokens?: number;
+    /** All generated tokens, including the reasoning subset when present. */
+    outputTokens?: number;
+    cachedTokens?: number;
+    cacheWriteTokens?: number;
+    /** Reasoning-token subset of `outputTokens`. */
+    reasoningTokens?: number;
+    costUsd?: number;
+    finishReason?: string;
+}
+interface ToolSpan extends SpanBase {
+    kind: 'tool';
+    toolName: string;
+    args: unknown;
+    /** False when the source observed the call but did not capture its arguments. */
+    argsCaptured?: boolean;
+    result?: unknown;
+    latencyMs?: number;
+}
+interface RetrievalSpan extends SpanBase {
+    kind: 'retrieval';
+    query: string;
+    hits: Array<{
+        docId: string;
+        score: number;
+        content?: string;
+    }>;
+}
+interface JudgeSpan extends SpanBase {
+    kind: 'judge';
+    judgeId: string;
+    /** Span this judgment applies to. */
+    targetSpanId: string;
+    dimension: string;
+    /** Numeric score (free-range; interpretation up to the judge). */
+    score: number;
+    rationale?: string;
+    evidence?: string;
+}
+interface SandboxSpan extends SpanBase {
+    kind: 'sandbox';
+    image?: string;
+    command?: string;
+    exitCode?: number;
+    testsTotal?: number;
+    testsPassed?: number;
+    stdoutHash?: string;
+    stderrHash?: string;
+    /** Duration in ms; the harness fills this explicitly (endedAt - startedAt may miss setup). */
+    wallMs?: number;
+}
+interface GenericSpan extends SpanBase {
+    kind: 'agent' | 'custom';
+}
+type Span = LlmSpan | ToolSpan | RetrievalSpan | JudgeSpan | SandboxSpan | GenericSpan;
 
 /**
  * Code-edit extraction + the "watch the agent write code" animation.

package/dist/trace-attributes.d.ts

@@ -0,0 +1,16 @@
+/** Canonical OpenInference-over-OTLP attribute names used at trace boundaries. */
+declare const OPENINFERENCE_SPAN_KIND = "openinference.span.kind";
+declare const LLM_MODEL_NAME = "llm.model_name";
+declare const LLM_INPUT_TOKENS = "llm.token_count.prompt";
+declare const LLM_OUTPUT_TOKENS = "llm.token_count.completion";
+declare const LLM_REASONING_TOKENS = "llm.token_count.reasoning";
+declare const LLM_CACHED_TOKENS = "llm.token_count.prompt_cache_hit";
+declare const LLM_CACHE_WRITE_TOKENS = "llm.token_count.prompt_cache_write";
+declare const LLM_COST_USD = "llm.cost_usd";
+declare const TOOL_NAME = "tool.name";
+declare const TOOL_ARGS_CAPTURED = "tool.args_captured";
+declare const TOOL_LATENCY_MS = "tool.latency_ms";
+declare const INPUT_VALUE = "input.value";
+declare const OUTPUT_VALUE = "output.value";
+
+export { INPUT_VALUE, LLM_CACHED_TOKENS, LLM_CACHE_WRITE_TOKENS, LLM_COST_USD, LLM_INPUT_TOKENS, LLM_MODEL_NAME, LLM_OUTPUT_TOKENS, LLM_REASONING_TOKENS, OPENINFERENCE_SPAN_KIND, OUTPUT_VALUE, TOOL_ARGS_CAPTURED, TOOL_LATENCY_MS, TOOL_NAME };

package/dist/trace-attributes.js

@@ -0,0 +1,32 @@
+import {
+  INPUT_VALUE,
+  LLM_CACHED_TOKENS,
+  LLM_CACHE_WRITE_TOKENS,
+  LLM_COST_USD,
+  LLM_INPUT_TOKENS,
+  LLM_MODEL_NAME,
+  LLM_OUTPUT_TOKENS,
+  LLM_REASONING_TOKENS,
+  OPENINFERENCE_SPAN_KIND,
+  OUTPUT_VALUE,
+  TOOL_ARGS_CAPTURED,
+  TOOL_LATENCY_MS,
+  TOOL_NAME
+} from "./chunk-WDHBCA3M.js";
+import "./chunk-PZ5AY32C.js";
+export {
+  INPUT_VALUE,
+  LLM_CACHED_TOKENS,
+  LLM_CACHE_WRITE_TOKENS,
+  LLM_COST_USD,
+  LLM_INPUT_TOKENS,
+  LLM_MODEL_NAME,
+  LLM_OUTPUT_TOKENS,
+  LLM_REASONING_TOKENS,
+  OPENINFERENCE_SPAN_KIND,
+  OUTPUT_VALUE,
+  TOOL_ARGS_CAPTURED,
+  TOOL_LATENCY_MS,
+  TOOL_NAME
+};
+//# sourceMappingURL=trace-attributes.js.map

package/dist/trace-attributes.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}