@alis-build/harness-eval 0.1.2 → 0.1.3

package/dist/{projections-BcX7w-f6.js → reporter-Biy-5-9M.js}

@@ -1,506 +1,394 @@
-import { i as buildJudgeArgs } from "./claude-code-DZ4Vkgp6.js";
-import { n as createLimit } from "./suite-Dlzl-HI0.js";
+import { i as buildJudgeArgs } from "./claude-code-C_7hxC8z.js";
+import { n as createLimit, t as runSuite, u as getAdapter } from "./suite-BcP64nlb.js";
+import { s as buildJudgeArgs$1 } from "./codex-0cHO2te9.js";
+import { i as loadGradingConfig, l as ConfigError, o as loadSuiteDocument, s as DEFAULT_PIPELINE_OUTPUTS, t as loadSuite } from "./loader-DnQ6Jt0i.js";
 import { spawn } from "node:child_process";
-import { readFile } from "node:fs/promises";
-import { parse } from "yaml";
+import { readFile, stat, writeFile } from "node:fs/promises";
+import { basename, dirname, join, resolve } from "node:path";
 import { createHash, randomUUID } from "node:crypto";
+import { parse } from "yaml";
+import { fileURLToPath } from "node:url";
 //#region src/types/eval-record.ts
 /** Schema version for {@link EvalRunEnvelope} JSON documents. */
 const EVAL_RUN_SCHEMA_VERSION = "1.0";
 /** Schema version embedded in each {@link TrajectoryView} at export time. */
 const TRAJECTORY_SCHEMA_VERSION = "1.0";
 //#endregion
-//#region src/otel/attributes.ts
-/** Build a string-typed OTLP attribute. */
-function strAttr(key, value) {
-	return {
-		key,
-		value: { stringValue: value }
-	};
+//#region src/grader/prompt.ts
+/**
+* Build the full grader prompt including eval prompt, transcript, and schema.
+*
+* When `systemInstruction` is set it is prepended as a judge-specific prefix.
+*/
+function buildGraderPrompt(input) {
+	const expectationList = input.expectations.map((e, i) => `${i + 1}. ${e}`).join("\n");
+	return `${input.systemInstruction ? `${input.systemInstruction.trim()}\n\n` : ""}You are an automated evaluation grader (not the agent under test). Your only job is to score expectations against the transcript below.
+
+Your job is to evaluate each expectation against the transcript and final response.
+PASS only when there is clear evidence in the transcript or final response.
+When uncertain, FAIL — burden of proof is on PASS.
+
+Also critique the expectations themselves if any are trivially satisfied or miss important outcomes.
+
+## Eval prompt
+
+${input.prompt}
+
+## Execution transcript
+
+${input.transcript}
+
+## Expectations to grade
+
+${expectationList}
+
+## Output format
+
+Respond with ONLY a single JSON object (no markdown fences, no commentary) matching this schema:
+
+{
+  "expectations": [
+    { "text": "<original expectation>", "passed": true|false, "evidence": "<quote or description>" }
+  ],
+  "summary": { "passed": <int>, "failed": <int>, "total": <int>, "pass_rate": <0.0-1.0> },
+  "eval_feedback": {
+    "suggestions": [{ "assertion": "<optional>", "reason": "<string>" }],
+    "overall": "<brief assessment>"
+  }
 }
-/** Build an integer-typed OTLP attribute (stored as decimal string). */
-function intAttr(key, value) {
-	return {
-		key,
-		value: { intValue: String(value) }
-	};
+
+Include every expectation in the same order. summary must match the expectations array.`;
 }
-/** Build a boolean-typed OTLP attribute. */
-function boolAttr(key, value) {
-	return {
-		key,
-		value: { boolValue: value }
-	};
+//#endregion
+//#region src/grader/parse.ts
+/**
+* Extract assistant text from Claude stdout.
+*
+* Handles plain text, single JSON result envelopes, stream-json arrays, and
+* assistant message objects — the judge subprocess may emit any of these
+* depending on Claude Code version and flags.
+*/
+function extractClaudeResponseText(stdout) {
+	const trimmed = stdout.trim();
+	if (!trimmed) return "";
+	try {
+		const data = JSON.parse(trimmed);
+		if (Array.isArray(data)) return extractFromEventArray(data) ?? trimmed;
+		if (typeof data === "object" && data !== null) {
+			const event = data;
+			if (event.type === "result" && typeof event.result === "string") return event.result;
+			if (event.type === "assistant" && event.message) {
+				const text = textFromAssistantMessage(event.message);
+				if (text) return text;
+			}
+		}
+	} catch {}
+	return trimmed;
 }
-/** Build a JSON-serialized string attribute (common for message arrays). */
-function jsonAttr(key, value) {
-	return {
-		key,
-		value: { stringValue: JSON.stringify(value) }
-	};
+/**
+* Extract assistant text from Codex judge stdout.
+*
+* Handles plain text and JSONL streams from accidental `--json` usage.
+*/
+function extractCodexResponseText(stdout) {
+	const trimmed = stdout.trim();
+	if (!trimmed) return "";
+	const lines = trimmed.split("\n").filter((line) => line.trim().length > 0);
+	if (lines.length > 1) for (let i = lines.length - 1; i >= 0; i--) try {
+		const event = JSON.parse(lines[i]);
+		if (event.type === "item.completed" && (event.item?.type === "assistant_message" || event.item?.item_type === "assistant_message") && event.item.text) return event.item.text;
+	} catch {
+		continue;
+	}
+	return trimmed;
+}
+/** Walk a stream-json event array and return the final assistant or result text. */
+function extractFromEventArray(events) {
+	const result = events.find((e) => typeof e === "object" && e !== null && e.type === "result");
+	if (result?.result) return result.result;
+	const assistantTexts = [];
+	for (const event of events) if (typeof event === "object" && event !== null && event.type === "assistant") {
+		const text = textFromAssistantMessage(event.message);
+		if (text) assistantTexts.push(text);
+	}
+	if (assistantTexts.length > 0) return assistantTexts[assistantTexts.length - 1];
+	return null;
+}
+/** Concatenate text blocks from an Anthropic-style assistant message object. */
+function textFromAssistantMessage(message) {
+	if (!message || typeof message !== "object") return null;
+	const content = message.content;
+	if (typeof content === "string") return content;
+	if (!Array.isArray(content)) return null;
+	const texts = [];
+	for (const block of content) if (typeof block === "object" && block !== null && block.type === "text" && typeof block.text === "string") texts.push(block.text);
+	return texts.length > 0 ? texts.join("\n") : null;
 }
-//#endregion
-//#region src/otel/messages.ts
 /**
-* Map harness stop reasons to GenAI semconv finish_reason values.
+* Parse grader JSON from response text.
 *
-* Unknown reasons pass through unchanged for forward compatibility.
+* Tries the raw string first, then fenced code blocks and brace-delimited
+* substrings. Returns null when no valid expectations array is found.
 */
-function mapStopReason(reason) {
-	if (!reason) return void 0;
-	switch (reason) {
-		case "end_turn": return "stop";
-		case "tool_use": return "tool_calls";
-		case "max_tokens": return "length";
-		case "stop_sequence": return "stop";
-		default: return reason;
+function parseGraderJson(text) {
+	const candidates = [text.trim(), extractJsonBlock(text)];
+	for (const candidate of candidates) {
+		if (!candidate) continue;
+		try {
+			const normalized = normalizeGraderJson(JSON.parse(candidate));
+			if (normalized.expectations.length > 0) return normalized;
+		} catch {
+			continue;
+		}
 	}
+	return null;
 }
-/** Build a tool_call part from a {@link ToolCall}. */
-function toolCallPart(call) {
-	return {
-		type: "tool_call",
-		id: call.callId,
-		name: call.name,
-		arguments: call.args ?? {}
-	};
+/** Extract JSON from markdown fences or the outermost `{...}` substring. */
+function extractJsonBlock(text) {
+	const fence = text.match(/```(?:json)?\s*([\s\S]*?)```/);
+	if (fence?.[1]) return fence[1].trim();
+	const start = text.indexOf("{");
+	const end = text.lastIndexOf("}");
+	if (start >= 0 && end > start) return text.slice(start, end + 1);
+	return null;
 }
-/** Build a tool_call_response part from a {@link ToolCall} result. */
-function toolResponsePart(call) {
-	return {
-		type: "tool_call_response",
-		id: call.callId,
-		result: call.result
+/** Map raw grader JSON to runtime {@link GraderOutput} with computed summary. */
+function normalizeGraderJson(raw) {
+	const expectations = (raw.expectations ?? []).map((e) => ({
+		text: e.text ?? "",
+		passed: Boolean(e.passed),
+		evidence: e.evidence ?? ""
+	}));
+	const passed = expectations.filter((e) => e.passed).length;
+	const failed = expectations.length - passed;
+	const total = expectations.length;
+	const passRate = raw.summary?.pass_rate ?? raw.summary?.passRate ?? (total === 0 ? 0 : passed / total);
+	const summary = {
+		passed: raw.summary?.passed ?? passed,
+		failed: raw.summary?.failed ?? failed,
+		total: raw.summary?.total ?? total,
+		passRate
 	};
-}
-/** Convert one assistant turn to a GenAI semconv assistant message. */
-function assistantMessageFromTurn(turn) {
-	const parts = [];
-	if (turn.text) parts.push({
-		type: "text",
-		content: turn.text
-	});
-	for (const call of turn.toolCalls) parts.push(toolCallPart(call));
-	const finish = mapStopReason(turn.stopReason);
-	return {
-		role: "assistant",
-		parts,
-		...finish ? { finish_reason: finish } : {}
+	let evalFeedback;
+	if (raw.eval_feedback) evalFeedback = {
+		suggestions: (raw.eval_feedback.suggestions ?? []).map((s) => ({
+			assertion: s.assertion,
+			reason: s.reason ?? ""
+		})),
+		overall: raw.eval_feedback.overall ?? ""
 	};
-}
-/** Aggregate tool results from a turn into a single tool-role message, if any. */
-function toolResultsMessage(calls) {
-	const parts = calls.filter((c) => c.result !== null).map((c) => toolResponsePart(c));
-	if (parts.length === 0) return null;
 	return {
-		role: "tool",
-		parts
+		expectations,
+		summary,
+		evalFeedback
 	};
 }
+//#endregion
+//#region src/grader/spawn-judge.ts
 /**
-* Input history before the assistant turn at `turnIndex`.
+* Shared subprocess utilities for judge graders (Claude + Codex).
+*
+* Owns detached spawn, process-group teardown, and SIGTERM → SIGKILL
+* escalation so both graders share one implementation.
 */
-function inputMessagesBeforeTurn(view, turnIndex, prompt) {
-	const messages = [];
-	if (prompt) messages.push({
-		role: "user",
-		parts: [{
-			type: "text",
-			content: prompt
-		}]
-	});
-	for (let i = 0; i < turnIndex; i++) {
-		const turn = view.turns[i];
-		if (!turn) continue;
-		messages.push(assistantMessageFromTurn(turn));
-		const toolMsg = toolResultsMessage(turn.toolCalls);
-		if (toolMsg) messages.push(toolMsg);
+const KILL_GRACE_MS = 5e3;
+/** Kill the detached process group (fallback to single process if group kill fails). */
+function killTree(child, signal) {
+	if (child.pid === void 0) return;
+	try {
+		process.kill(-child.pid, signal);
+	} catch {
+		try {
+			child.kill(signal);
+		} catch {}
 	}
-	return messages;
 }
-//#endregion
-//#region src/otel/types.ts
-/** OTLP span kinds (enum integers). */
-const SpanKind = {
-	INTERNAL: 1,
-	CLIENT: 2
-};
-/** OTLP status codes. */
-const StatusCode = {
-	UNSET: 0,
-	OK: 1,
-	ERROR: 2
-};
-//#endregion
-//#region src/otel/emitter.ts
-/**
-* TrajectoryView → OTLP JSON export using OpenTelemetry GenAI semantic conventions.
-*
-* Produces an `ExportTraceServiceRequest` suitable for OTLP/HTTP JSON ingestion.
-* Assertions continue to use {@link TrajectoryView} directly; this is export-only.
-*/
-const INSTRUMENTATION_VERSION = "0.1.0";
 /**
-* Map a {@link TrajectoryView} to OTLP trace JSON.
+* Spawn a judge subprocess with process-group teardown and collect stdout.
 *
-* Span tree (siblings under `invoke_agent`, not nested):
-* ```
-* invoke_agent
-* ├── chat {model}
-* ├── execute_tool {name}
-* ├── chat {model}
-* └── ...
-* ```
+* Non-zero exit with empty stdout is treated as failure; partial stdout on
+* non-zero exit is retained (judges sometimes exit non-zero after emitting JSON).
 */
-function trajectoryToOtlp(view, options = {}) {
-	const agentName = options.agentName ?? "claude-code";
-	const providerName = options.providerName ?? "anthropic";
-	const serviceName = options.serviceName ?? "harness-eval";
-	const scopeName = options.instrumentationScope ?? "@alis-build/harness-eval";
-	const traceId = traceIdFromSession(view.meta.sessionId);
-	const rootSpanId = spanIdFromKey(traceId, "invoke_agent");
-	const durationMs = Math.max(view.usage.durationMs, 1);
-	const endMs = options.endTimeMs ?? Date.now();
-	const startMs = endMs - durationMs;
-	const rootStartNs = msToNs(startMs);
-	const rootEndNs = msToNs(endMs);
-	const spans = [];
-	const timings = buildSpanTimings(view, startMs, endMs);
-	spans.push({
-		traceId,
-		spanId: rootSpanId,
-		name: "invoke_agent",
-		kind: SpanKind.INTERNAL,
-		startTimeUnixNano: rootStartNs,
-		endTimeUnixNano: rootEndNs,
-		attributes: [
-			strAttr("gen_ai.operation.name", "invoke_agent"),
-			strAttr("gen_ai.agent.name", agentName),
-			strAttr("gen_ai.provider.name", providerName),
-			strAttr("gen_ai.conversation.id", view.meta.sessionId),
-			strAttr("gen_ai.request.model", view.meta.model),
-			strAttr("gen_ai.response.model", view.meta.model),
-			intAttr("gen_ai.usage.input_tokens", view.usage.inputTokens),
-			intAttr("gen_ai.usage.output_tokens", view.usage.outputTokens),
-			boolAttr("harness_eval.success", view.success)
-		],
-		status: viewStatus(view)
-	});
-	let opIndex = 0;
-	for (const turn of view.turns) {
-		const chatTiming = timings[opIndex++];
-		const chatSpanId = spanIdFromKey(traceId, `chat:${turn.turnIndex}`);
-		const inputMessages = inputMessagesBeforeTurn(view, turn.turnIndex, options.prompt);
-		const outputMessages = [assistantMessageFromTurn(turn)];
-		spans.push({
-			traceId,
-			spanId: chatSpanId,
-			parentSpanId: rootSpanId,
-			name: `chat ${view.meta.model}`,
-			kind: SpanKind.CLIENT,
-			startTimeUnixNano: chatTiming.startNs,
-			endTimeUnixNano: chatTiming.endNs,
-			attributes: [
-				strAttr("gen_ai.operation.name", "chat"),
-				strAttr("gen_ai.provider.name", providerName),
-				strAttr("gen_ai.request.model", view.meta.model),
-				strAttr("gen_ai.response.model", view.meta.model),
-				...inputMessages.length > 0 ? [jsonAttr("gen_ai.input.messages", inputMessages)] : [],
-				jsonAttr("gen_ai.output.messages", outputMessages),
-				...turn.stopReason ? [jsonAttr("gen_ai.response.finish_reasons", [mapStopReason(turn.stopReason) ?? turn.stopReason])] : []
+function spawnCollectStdout(options) {
+	const { binary, args, timeoutMs, env, cwd } = options;
+	return new Promise((resolve, reject) => {
+		const child = spawn(binary, args, {
+			env: env ?? process.env,
+			cwd,
+			stdio: [
+				"ignore",
+				"pipe",
+				"pipe"
 			],
-			status: { code: StatusCode.OK }
+			detached: true
 		});
-		if (turn.toolCalls.length === 0) continue;
-		const toolTiming = timings[opIndex++];
-		for (const call of turn.toolCalls) {
-			const toolSpanId = spanIdFromKey(traceId, `tool:${call.callId}`);
-			spans.push({
-				traceId,
-				spanId: toolSpanId,
-				parentSpanId: rootSpanId,
-				name: `execute_tool ${call.name}`,
-				kind: SpanKind.INTERNAL,
-				startTimeUnixNano: toolTiming.startNs,
-				endTimeUnixNano: toolTiming.endNs,
-				attributes: [
-					strAttr("gen_ai.operation.name", "execute_tool"),
-					strAttr("gen_ai.provider.name", providerName),
-					strAttr("gen_ai.tool.name", call.name),
-					strAttr("gen_ai.tool.call.id", call.callId),
-					jsonAttr("gen_ai.tool.call.arguments", call.args ?? {}),
-					...call.result !== null ? [jsonAttr("gen_ai.tool.call.result", call.result)] : [],
-					...call.namespace ? [strAttr("harness_eval.tool.namespace", call.namespace)] : [],
-					boolAttr("harness_eval.tool.is_error", call.isError)
-				],
-				status: call.isError ? {
-					code: StatusCode.ERROR,
-					message: "tool reported error"
-				} : { code: StatusCode.OK }
-			});
-		}
-	}
-	return { resourceSpans: [{
-		resource: { attributes: [strAttr("service.name", serviceName), strAttr("gen_ai.agent.name", agentName)] },
-		scopeSpans: [{
-			scope: {
-				name: scopeName,
-				version: INSTRUMENTATION_VERSION
-			},
-			spans
-		}]
-	}] };
+		const chunks = [];
+		child.stdout?.setEncoding("utf8");
+		child.stdout?.on("data", (c) => chunks.push(c));
+		const stderrChunks = [];
+		child.stderr?.setEncoding("utf8");
+		child.stderr?.on("data", (c) => stderrChunks.push(c));
+		let killEscalation = null;
+		const timer = setTimeout(() => {
+			killTree(child, "SIGTERM");
+			killEscalation = setTimeout(() => killTree(child, "SIGKILL"), KILL_GRACE_MS);
+			const stderrHint = stderrChunks.join("").trim().slice(0, 400);
+			reject(/* @__PURE__ */ new Error(`grader timed out after ${timeoutMs}ms` + (stderrHint ? ` (stderr: ${stderrHint})` : "")));
+		}, timeoutMs);
+		const finalize = (err) => {
+			clearTimeout(timer);
+			if (killEscalation) clearTimeout(killEscalation);
+			if (err) reject(err);
+			else resolve(chunks.join(""));
+		};
+		child.on("error", (err) => finalize(err));
+		child.on("close", (code) => {
+			if (code !== 0 && chunks.length === 0) finalize(/* @__PURE__ */ new Error(`grader exited ${code}: ${stderrChunks.join("").slice(0, 500)}`));
+			else finalize();
+		});
+	});
 }
-/** Alias for {@link trajectoryToOtlp} — matches implementation plan naming. */
-const emitOtel = trajectoryToOtlp;
-/** Map view success flag to OTLP span status on the root invoke_agent span. */
-function viewStatus(view) {
-	if (view.success) return { code: StatusCode.OK };
+//#endregion
+//#region src/grader/claude-grader.ts
+/**
+* Grade expectations by spawning Claude as judge (skill-creator grader pattern).
+*/
+const DEFAULT_TIMEOUT_MS$1 = 3e5;
+/**
+* Judge subprocess defaults — grading is a single-shot JSON response, not an agent session.
+* Without these, Claude Code may load plugins/MCP and loop on tools until timeout.
+*/
+const JUDGE_CLAUDE_DEFAULTS = {
+	maxTurns: 1,
+	bare: true,
+	disableSlashCommands: true,
+	noSessionPersistence: true
+};
+/** Merge user-supplied Claude Code options over judge-safe defaults. */
+function mergeJudgeClaudeOptions(claudeCode) {
 	return {
-		code: StatusCode.ERROR,
-		message: "harness run did not complete successfully"
+		...JUDGE_CLAUDE_DEFAULTS,
+		...claudeCode
 	};
 }
+/** Factory returning a {@link GraderFn} bound to subprocess options. */
+function createClaudeGrader(options = {}) {
+	return (input) => runClaudeGrader(input, options);
+}
 /**
-* Assign synthetic timestamps to chat and tool spans.
+* Spawn Claude as judge, parse JSON response, align with input expectations.
 *
-* Stream-json does not carry per-turn wall times, so we divide the session
-* duration evenly across chat/tool slots for OTLP consumers that require
-* start/end times on every span.
+* Unparseable output fails all expectations and sets {@link GraderOutput.error}.
 */
-function buildSpanTimings(view, startMs, endMs) {
-	const slots = [];
-	for (const turn of view.turns) {
-		slots.push("chat");
-		if (turn.toolCalls.length > 0) slots.push("tools");
-	}
-	if (slots.length === 0) return [];
-	const slotMs = Math.max(endMs - startMs, 1) / slots.length;
-	const timings = [];
-	let offset = startMs;
-	for (const slot of slots) {
-		const slotStart = offset;
-		const slotEnd = offset + slotMs;
-		timings.push({
-			startNs: msToNs(slotStart),
-			endNs: msToNs(slotEnd)
-		});
-		offset = slotEnd;
-	}
-	return timings;
+async function runClaudeGrader(input, options = {}) {
+	const binary = options.binary ?? options.claudeCode?.binary ?? "claude";
+	const timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS$1;
+	const prompt = buildGraderPrompt(input);
+	const model = options.model ?? options.claudeCode?.model;
+	const responseText = extractClaudeResponseText(await spawnCollectStdout({
+		binary,
+		args: buildJudgeArgs(prompt, {
+			...mergeJudgeClaudeOptions(options.claudeCode),
+			model
+		}),
+		timeoutMs,
+		env: buildChildEnv(options.env),
+		cwd: options.cwd
+	}));
+	const parsed = parseGraderJson(responseText);
+	if (!parsed) return {
+		expectations: input.expectations.map((text) => ({
+			text,
+			passed: false,
+			evidence: "Grader returned unparseable output"
+		})),
+		summary: {
+			passed: 0,
+			failed: input.expectations.length,
+			total: input.expectations.length,
+			passRate: 0
+		},
+		error: `failed to parse grader JSON from response: ${responseText.slice(0, 200)}`
+	};
+	const expectations = input.expectations.map((text, i) => {
+		const graded = parsed.expectations[i];
+		return {
+			text,
+			passed: graded?.passed ?? false,
+			evidence: graded?.evidence ?? "No evidence returned"
+		};
+	});
+	const passed = expectations.filter((e) => e.passed).length;
+	const total = expectations.length;
+	return {
+		expectations,
+		summary: {
+			passed,
+			failed: total - passed,
+			total,
+			passRate: total === 0 ? 0 : passed / total
+		},
+		evalFeedback: parsed.evalFeedback
+	};
 }
 /**
-* Derive a deterministic 128-bit trace id from the harness session id.
-*
-* Uses SHA-256 truncation so the same session always maps to the same trace.
+* Build subprocess env, stripping CLAUDECODE to avoid nested-session guards.
 */
-function traceIdFromSession(sessionId) {
-	return createHash("sha256").update(`harness-eval:trace:${sessionId}`).digest("hex").slice(0, 32).toUpperCase();
+function buildChildEnv(extraEnv) {
+	const env = {
+		...process.env,
+		...extraEnv
+	};
+	delete env.CLAUDECODE;
+	return env;
 }
+//#endregion
+//#region src/grader/codex-grader.ts
 /**
-* Derive a deterministic 64-bit span id from trace id and a logical span key.
-*/
-function spanIdFromKey(traceId, key) {
-	return createHash("sha256").update(`${traceId}:span:${key}`).digest("hex").slice(0, 16).toUpperCase();
-}
-/** Convert milliseconds since epoch to OTLP nanosecond timestamp string. */
-function msToNs(ms) {
-	return String(Math.round(ms * 1e6));
-}
-//#endregion
-//#region src/grader/prompt.ts
-/**
-* Build the full grader prompt including eval prompt, transcript, and schema.
-*
-* When `systemInstruction` is set it is prepended as a judge-specific prefix.
-*/
-function buildGraderPrompt(input) {
-	const expectationList = input.expectations.map((e, i) => `${i + 1}. ${e}`).join("\n");
-	return `${input.systemInstruction ? `${input.systemInstruction.trim()}\n\n` : ""}You are an automated evaluation grader (not the agent under test). Your only job is to score expectations against the transcript below.
-
-Your job is to evaluate each expectation against the transcript and final response.
-PASS only when there is clear evidence in the transcript or final response.
-When uncertain, FAIL — burden of proof is on PASS.
-
-Also critique the expectations themselves if any are trivially satisfied or miss important outcomes.
-
-## Eval prompt
-
-${input.prompt}
-
-## Execution transcript
-
-${input.transcript}
-
-## Expectations to grade
-
-${expectationList}
-
-## Output format
-
-Respond with ONLY a single JSON object (no markdown fences, no commentary) matching this schema:
-
-{
-  "expectations": [
-    { "text": "<original expectation>", "passed": true|false, "evidence": "<quote or description>" }
-  ],
-  "summary": { "passed": <int>, "failed": <int>, "total": <int>, "pass_rate": <0.0-1.0> },
-  "eval_feedback": {
-    "suggestions": [{ "assertion": "<optional>", "reason": "<string>" }],
-    "overall": "<brief assessment>"
-  }
-}
-
-Include every expectation in the same order. summary must match the expectations array.`;
-}
-//#endregion
-//#region src/grader/parse.ts
-/**
-* Extract assistant text from Claude stdout.
-*
-* Handles plain text, single JSON result envelopes, stream-json arrays, and
-* assistant message objects — the judge subprocess may emit any of these
-* depending on Claude Code version and flags.
-*/
-function extractClaudeResponseText(stdout) {
-	const trimmed = stdout.trim();
-	if (!trimmed) return "";
-	try {
-		const data = JSON.parse(trimmed);
-		if (Array.isArray(data)) return extractFromEventArray(data) ?? trimmed;
-		if (typeof data === "object" && data !== null) {
-			const event = data;
-			if (event.type === "result" && typeof event.result === "string") return event.result;
-			if (event.type === "assistant" && event.message) {
-				const text = textFromAssistantMessage(event.message);
-				if (text) return text;
-			}
-		}
-	} catch {}
-	return trimmed;
-}
-/** Walk a stream-json event array and return the final assistant or result text. */
-function extractFromEventArray(events) {
-	const result = events.find((e) => typeof e === "object" && e !== null && e.type === "result");
-	if (result?.result) return result.result;
-	const assistantTexts = [];
-	for (const event of events) if (typeof event === "object" && event !== null && event.type === "assistant") {
-		const text = textFromAssistantMessage(event.message);
-		if (text) assistantTexts.push(text);
-	}
-	if (assistantTexts.length > 0) return assistantTexts[assistantTexts.length - 1];
-	return null;
-}
-/** Concatenate text blocks from an Anthropic-style assistant message object. */
-function textFromAssistantMessage(message) {
-	if (!message || typeof message !== "object") return null;
-	const content = message.content;
-	if (typeof content === "string") return content;
-	if (!Array.isArray(content)) return null;
-	const texts = [];
-	for (const block of content) if (typeof block === "object" && block !== null && block.type === "text" && typeof block.text === "string") texts.push(block.text);
-	return texts.length > 0 ? texts.join("\n") : null;
-}
-/**
-* Parse grader JSON from response text.
-*
-* Tries the raw string first, then fenced code blocks and brace-delimited
-* substrings. Returns null when no valid expectations array is found.
-*/
-function parseGraderJson(text) {
-	const candidates = [text.trim(), extractJsonBlock(text)];
-	for (const candidate of candidates) {
-		if (!candidate) continue;
-		try {
-			const normalized = normalizeGraderJson(JSON.parse(candidate));
-			if (normalized.expectations.length > 0) return normalized;
-		} catch {
-			continue;
-		}
-	}
-	return null;
-}
-/** Extract JSON from markdown fences or the outermost `{...}` substring. */
-function extractJsonBlock(text) {
-	const fence = text.match(/```(?:json)?\s*([\s\S]*?)```/);
-	if (fence?.[1]) return fence[1].trim();
-	const start = text.indexOf("{");
-	const end = text.lastIndexOf("}");
-	if (start >= 0 && end > start) return text.slice(start, end + 1);
-	return null;
-}
-/** Map raw grader JSON to runtime {@link GraderOutput} with computed summary. */
-function normalizeGraderJson(raw) {
-	const expectations = (raw.expectations ?? []).map((e) => ({
-		text: e.text ?? "",
-		passed: Boolean(e.passed),
-		evidence: e.evidence ?? ""
-	}));
-	const passed = expectations.filter((e) => e.passed).length;
-	const failed = expectations.length - passed;
-	const total = expectations.length;
-	const passRate = raw.summary?.pass_rate ?? raw.summary?.passRate ?? (total === 0 ? 0 : passed / total);
-	const summary = {
-		passed: raw.summary?.passed ?? passed,
-		failed: raw.summary?.failed ?? failed,
-		total: raw.summary?.total ?? total,
-		passRate
-	};
-	let evalFeedback;
-	if (raw.eval_feedback) evalFeedback = {
-		suggestions: (raw.eval_feedback.suggestions ?? []).map((s) => ({
-			assertion: s.assertion,
-			reason: s.reason ?? ""
-		})),
-		overall: raw.eval_feedback.overall ?? ""
-	};
-	return {
-		expectations,
-		summary,
-		evalFeedback
-	};
-}
-//#endregion
-//#region src/grader/claude-grader.ts
-/**
-* Grade expectations by spawning Claude as judge (skill-creator grader pattern).
+* Grade expectations by spawning Codex as judge.
 */
 const DEFAULT_TIMEOUT_MS = 3e5;
-/**
-* Judge subprocess defaults — grading is a single-shot JSON response, not an agent session.
-* Without these, Claude Code may load plugins/MCP and loop on tools until timeout.
-*/
-const JUDGE_CLAUDE_DEFAULTS = {
-	maxTurns: 1,
-	bare: true,
-	disableSlashCommands: true,
-	noSessionPersistence: true
+/** Judge subprocess defaults — single-shot grading without persistent sessions. */
+const JUDGE_CODEX_DEFAULTS = {
+	ephemeral: true,
+	ignoreUserConfig: true,
+	skipGitRepoCheck: true
 };
-/** Merge user-supplied Claude Code options over judge-safe defaults. */
-function mergeJudgeClaudeOptions(claudeCode) {
+/** Merge user-supplied Codex options over judge-safe defaults. */
+function mergeJudgeCodexOptions(codex) {
 	return {
-		...JUDGE_CLAUDE_DEFAULTS,
-		...claudeCode
+		...JUDGE_CODEX_DEFAULTS,
+		...codex
 	};
 }
 /** Factory returning a {@link GraderFn} bound to subprocess options. */
-function createClaudeGrader(options = {}) {
-	return (input) => runClaudeGrader(input, options);
+function createCodexGrader(options = {}) {
+	return (input) => runCodexGrader(input, options);
 }
 /**
-* Spawn Claude as judge, parse JSON response, align with input expectations.
+* Spawn Codex as judge, parse JSON response, align with input expectations.
 *
 * Unparseable output fails all expectations and sets {@link GraderOutput.error}.
 */
-async function runClaudeGrader(input, options = {}) {
-	const binary = options.binary ?? options.claudeCode?.binary ?? "claude";
+async function runCodexGrader(input, options = {}) {
+	const binary = options.binary ?? options.codex?.binary ?? "codex";
 	const timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS;
 	const prompt = buildGraderPrompt(input);
-	const model = options.model ?? options.claudeCode?.model;
-	const responseText = extractClaudeResponseText(await spawnCollectStdout(binary, buildJudgeArgs(prompt, {
-		...mergeJudgeClaudeOptions(options.claudeCode),
-		model
-	}), timeoutMs, options.env, options.cwd));
+	const model = options.model ?? options.codex?.model;
+	const responseText = extractCodexResponseText(await spawnCollectStdout({
+		binary,
+		args: buildJudgeArgs$1(prompt, {
+			...mergeJudgeCodexOptions(options.codex),
+			model,
+			cwd: options.cwd
+		}),
+		timeoutMs,
+		env: {
+			...process.env,
+			...options.env
+		},
+		cwd: options.cwd
+	}));
 	const parsed = parseGraderJson(responseText);
 	if (!parsed) return {
 		expectations: input.expectations.map((text) => ({
@@ -537,57 +425,6 @@ async function runClaudeGrader(input, options = {}) {
 		evalFeedback: parsed.evalFeedback
 	};
 }
-/**
-* Spawn a child process and collect stdout until exit or timeout.
-*
-* Non-zero exit with empty stdout is treated as failure; partial stdout on
-* non-zero exit is retained (Claude sometimes exits non-zero after emitting JSON).
-*/
-function spawnCollectStdout(binary, args, timeoutMs, extraEnv, cwd) {
-	return new Promise((resolve, reject) => {
-		const child = spawn(binary, args, {
-			env: buildChildEnv(extraEnv),
-			cwd,
-			stdio: [
-				"ignore",
-				"pipe",
-				"pipe"
-			]
-		});
-		const chunks = [];
-		child.stdout?.setEncoding("utf8");
-		child.stdout?.on("data", (c) => chunks.push(c));
-		const stderrChunks = [];
-		child.stderr?.setEncoding("utf8");
-		child.stderr?.on("data", (c) => stderrChunks.push(c));
-		const timer = setTimeout(() => {
-			child.kill("SIGTERM");
-			const stderrHint = stderrChunks.join("").trim().slice(0, 400);
-			reject(/* @__PURE__ */ new Error(`grader timed out after ${timeoutMs}ms` + (stderrHint ? ` (stderr: ${stderrHint})` : "")));
-		}, timeoutMs);
-		const finalize = (err) => {
-			clearTimeout(timer);
-			if (err) reject(err);
-			else resolve(chunks.join(""));
-		};
-		child.on("error", (err) => finalize(err));
-		child.on("close", (code) => {
-			if (code !== 0 && chunks.length === 0) finalize(/* @__PURE__ */ new Error(`grader exited ${code}: ${stderrChunks.join("").slice(0, 500)}`));
-			else finalize();
-		});
-	});
-}
-/**
-* Build subprocess env, stripping CLAUDECODE to avoid nested-session guards.
-*/
-function buildChildEnv(extraEnv) {
-	const env = {
-		...process.env,
-		...extraEnv
-	};
-	delete env.CLAUDECODE;
-	return env;
-}
 //#endregion
 //#region src/grader/expectations.ts
 /**
@@ -663,6 +500,32 @@ function truncate(text) {
 	return `${text.slice(0, MAX_RESULT_CHARS)}… (truncated)`;
 }
 //#endregion
+//#region src/eval-record/judge-metadata.ts
+/** Map harness grading adapter id to a stable judge identifier. */
+function judgeIdForAdapter(adapter) {
+	switch (adapter) {
+		case "codex": return "harness-eval/codex-grader";
+		case "claude-code": return "harness-eval/claude-grader";
+		default: return adapter ? `harness-eval/${adapter}-grader` : "harness-eval/claude-grader";
+	}
+}
+/** Build {@link JudgeInfo} from grading adapter and optional model override. */
+function resolveJudgeInfo(options) {
+	const adapter = options.adapter ?? "claude-code";
+	return {
+		id: options.id ?? judgeIdForAdapter(adapter),
+		model: options.model,
+		adapter
+	};
+}
+/** Derive judge metadata from a parsed grading YAML config. */
+function judgeInfoFromGradingConfig(config) {
+	return resolveJudgeInfo({
+		adapter: config.judge.adapter ?? "claude-code",
+		model: config.judge.model ?? config.judge.codex?.model ?? config.judge.claudeCode?.model
+	});
+}
+//#endregion
 //#region src/grader/grade-report.ts
 /**
 * Grade a harness-eval SuiteReport with outcome expectations (LLM judge).
@@ -675,14 +538,21 @@ function truncate(text) {
 */
 async function gradeReport(report, options = {}) {
 	const expectationsMap = options.expectationsPath ? await loadExpectationsMap(options.expectationsPath) : {};
-	const gradeFn = options.gradeFn ?? createClaudeGrader({
+	const gradeFn = options.gradeFn ?? (options.judgeAdapter === "codex" ? createCodexGrader({
+		binary: options.binary,
+		model: options.model,
+		timeoutMs: options.timeoutMs,
+		env: options.env,
+		cwd: options.cwd,
+		codex: options.codex
+	}) : createClaudeGrader({
 		binary: options.binary,
 		model: options.model,
 		timeoutMs: options.timeoutMs,
 		env: options.env,
 		cwd: options.cwd,
 		claudeCode: options.claudeCode
-	});
+	}));
 	const limit = createLimit(options.maxConcurrent ?? 2);
 	const tasks = [];
 	for (const cell of report.cells) {
@@ -787,6 +657,10 @@ async function gradeReport(report, options = {}) {
 		gradedAt: (/* @__PURE__ */ new Date()).toISOString(),
 		sourceReport: options.sourceReport ?? "",
 		gradingConfigPath: options.gradingConfigPath,
+		judge: resolveJudgeInfo({
+			adapter: options.judgeAdapter ?? "claude-code",
+			model: options.model
+		}),
 		results,
 		summary: {
 			passed: passedExpectations,
@@ -809,11 +683,12 @@ async function loadSuiteReport(path) {
 function resolveGradeOptions(fileConfig, cli = {}, configPath) {
 	const judge = fileConfig?.judge;
 	const adapter = judge?.adapter ?? "claude-code";
-	if (adapter !== "claude-code") throw new Error(`unsupported grading adapter "${adapter}" (only claude-code today)`);
 	const claudeCode = judge?.claudeCode ?? {};
-	const binary = cli.binary ?? claudeCode.binary;
-	const model = cli.model ?? judge?.model ?? claudeCode.model;
-	return {
+	const codex = judge?.codex ?? {};
+	const adapterBlock = adapter === "codex" ? codex : claudeCode;
+	const binary = cli.binary ?? adapterBlock.binary;
+	const model = cli.model ?? judge?.model ?? adapterBlock.model;
+	if (adapter === "codex") return {
 		sourceReport: cli.sourceReport,
 		expectationsPath: cli.expectationsPath,
 		model,
@@ -823,17 +698,37 @@ function resolveGradeOptions(fileConfig, cli = {}, configPath) {
 		systemInstruction: judge?.system_instruction,
 		env: judge?.env,
 		cwd: judge?.cwd,
-		claudeCode: {
-			...claudeCode,
+		judgeAdapter: "codex",
+		codex: {
+			...codex,
 			binary: void 0,
 			model: void 0
 		},
 		gradingConfigPath: configPath
 	};
-}
-//#endregion
-//#region src/grader/format-console.ts
-const RESET$1 = "\x1B[0m";
+	if (adapter !== "claude-code") throw new Error(`unsupported grading adapter "${adapter}" (supported: claude-code, codex)`);
+	return {
+		sourceReport: cli.sourceReport,
+		expectationsPath: cli.expectationsPath,
+		model,
+		binary,
+		timeoutMs: cli.timeoutMs ?? judge?.timeoutMs,
+		maxConcurrent: cli.maxConcurrent ?? judge?.maxConcurrent,
+		systemInstruction: judge?.system_instruction,
+		env: judge?.env,
+		cwd: judge?.cwd,
+		judgeAdapter: "claude-code",
+		claudeCode: {
+			...claudeCode,
+			binary: void 0,
+			model: void 0
+		},
+		gradingConfigPath: configPath
+	};
+}
+//#endregion
+//#region src/grader/format-console.ts
+const RESET$1 = "\x1B[0m";
 const GREEN$1 = "\x1B[32m";
 const RED$1 = "\x1B[31m";
 const DIM = "\x1B[2m";
@@ -870,168 +765,6 @@ function gradingReportPassed(report) {
 	return report.results.every((r) => !r.graderError && r.summary.failed === 0 && r.summary.total > 0);
 }
 //#endregion
-//#region src/reporter/format-console.ts
-const RESET = "\x1B[0m";
-const GREEN = "\x1B[32m";
-const RED = "\x1B[31m";
-const YELLOW = "\x1B[33m";
-/**
-* Render renderable rows as ANSI-colored console output.
-*
-* @param color When false, emit plain text without escape codes.
-*/
-function formatConsole(rows, color = true) {
-	const lines = [];
-	for (const row of rows) {
-		const status = row.passed ? color ? `${GREEN}PASS${RESET}` : "PASS" : color ? `${RED}FAIL${RESET}` : "FAIL";
-		const crashNote = row.adapterErrors > 0 ? ` ${color ? YELLOW : ""}[${row.adapterErrors} adapter errors]${color ? RESET : ""}` : "";
-		lines.push(`${row.caseId} @ ${row.cellLabel}  ${status}${crashNote}`);
-		if (row.category) lines.push(`  category: ${row.category}`);
-		for (const stat of row.stats) {
-			const marker = stat.meetsThreshold ? color ? `${GREEN}✓${RESET}` : "✓" : color ? `${RED}✗${RESET}` : "✗";
-			const rateStr = formatRate$1(stat);
-			const thresholdPct = (stat.threshold * 100).toFixed(0);
-			let line = `  ├─ ${stat.description}: ${rateStr} [threshold ${thresholdPct}%] ${marker}`;
-			if (stat.delta !== void 0 && stat.baselinePassRate !== void 0) {
-				const arrow = stat.delta >= 0 ? "↑" : "↓";
-				const basePct = (stat.baselinePassRate * 100).toFixed(0);
-				const curPct = (stat.passRate * 100).toFixed(0);
-				const deltaPct = (stat.delta * 100).toFixed(0);
-				line += `  (${basePct}% → ${curPct}% (${arrow}${deltaPct}%))`;
-			}
-			lines.push(line);
-		}
-		lines.push("");
-	}
-	return lines.join("\n").trimEnd();
-}
-/** Format pass rate for display, noting when all reps crashed. */
-function formatRate$1(stat) {
-	if (stat.evaluatedCount === 0) return `0/${stat.totalReps} (all reps crashed)`;
-	const pct = (stat.passRate * 100).toFixed(0);
-	return `${stat.passedCount}/${stat.evaluatedCount} (${pct}%)`;
-}
-//#endregion
-//#region src/reporter/format-json.ts
-/**
-* Serialize a suite report as indented JSON (no transformation).
-*
-* Used by `--format json` and `--output` persistence.
-*/
-function formatJson(report) {
-	return JSON.stringify(report, null, 2);
-}
-//#endregion
-//#region src/reporter/format-markdown.ts
-/** Render renderable rows as a GitHub-flavored markdown report. */
-function formatMarkdown(rows) {
-	const lines = ["# Harness Eval Report", ""];
-	for (const row of rows) {
-		const status = row.passed ? "PASS" : "FAIL";
-		const crashNote = row.adapterErrors > 0 ? ` (${row.adapterErrors} adapter errors)` : "";
-		lines.push(`## ${row.caseId} @ ${row.cellLabel} — ${status}${crashNote}`);
-		if (row.category) lines.push(`**Category:** ${row.category}`);
-		if (row.notes) lines.push("<details><summary>Notes</summary>", row.notes, "</details>");
-		lines.push("");
-		lines.push("| Assertion | Result | Threshold | Status |");
-		lines.push("| --- | --- | --- | --- |");
-		for (const stat of row.stats) {
-			const rateStr = formatRate(stat);
-			const threshold = `${(stat.threshold * 100).toFixed(0)}%`;
-			const statusCell = stat.meetsThreshold ? "✓" : "✗";
-			let result = rateStr;
-			if (stat.delta !== void 0 && stat.baselinePassRate !== void 0) {
-				const base = (stat.baselinePassRate * 100).toFixed(0);
-				const cur = (stat.passRate * 100).toFixed(0);
-				const d = (stat.delta * 100).toFixed(0);
-				const sign = stat.delta >= 0 ? "+" : "";
-				result += ` (${base}% → ${cur}%, ${sign}${d}%)`;
-			}
-			lines.push(`| ${stat.description} | ${result} | ${threshold} | ${statusCell} |`);
-		}
-		lines.push("");
-	}
-	return lines.join("\n").trimEnd();
-}
-/** Format pass rate for markdown tables, noting when all reps crashed. */
-function formatRate(stat) {
-	if (stat.evaluatedCount === 0) return `0/${stat.totalReps} (all reps crashed)`;
-	const pct = (stat.passRate * 100).toFixed(0);
-	return `${stat.passedCount}/${stat.evaluatedCount} (${pct}%)`;
-}
-//#endregion
-//#region src/reporter/renderable.ts
-/** Map a suite report to formatter-ready rows (one per cell). */
-function toRenderableRows(report) {
-	return report.cells.map((cell) => cellToRow(cell));
-}
-/**
-* Attach baseline pass-rate deltas to matching rows.
-*
-* Rows without a matching baseline cell are returned unchanged.
-*/
-function applyBaseline(rows, baseline) {
-	const baselineMap = new Map(baseline.cells.map((c) => [`${c.caseId}::${c.cell.label}`, c]));
-	return rows.map((row) => {
-		const baseCell = baselineMap.get(`${row.caseId}::${row.cellLabel}`);
-		if (!baseCell) return row;
-		const stats = row.stats.map((stat, i) => {
-			const baseStat = baseCell.assertionStats[i];
-			if (!baseStat) return stat;
-			const delta = stat.passRate - baseStat.passRate;
-			return {
-				...stat,
-				baselinePassRate: baseStat.passRate,
-				delta
-			};
-		});
-		return {
-			...row,
-			stats
-		};
-	});
-}
-/** Convert one {@link CellReport} to a {@link RenderableRow}. */
-function cellToRow(cell) {
-	const totalReps = cell.repetitions.length;
-	const stats = cell.assertionStats.map((s) => ({
-		description: s.description,
-		threshold: s.threshold,
-		passedCount: s.passedCount,
-		evaluatedCount: s.evaluatedCount,
-		totalReps,
-		adapterErrors: cell.adapterErrors,
-		passRate: s.passRate,
-		meetsThreshold: s.meetsThreshold
-	}));
-	return {
-		caseId: cell.caseId,
-		category: cell.category,
-		notes: cell.notes,
-		cellLabel: cell.cell.label,
-		passed: cell.passed,
-		adapterErrors: cell.adapterErrors,
-		totalReps,
-		stats
-	};
-}
-//#endregion
-//#region src/reporter/index.ts
-/**
-* Format a {@link SuiteReport} for console, markdown, or JSON output.
-*
-* JSON format bypasses the renderable intermediate model and serializes the
-* report directly. Console and markdown apply optional baseline deltas.
-*/
-function formatReport(report, options) {
-	if (options.format === "json") return formatJson(report);
-	let rows = toRenderableRows(report);
-	if (options.baseline) rows = applyBaseline(rows, options.baseline);
-	const useColor = options.color ?? options.format === "console";
-	if (options.format === "markdown") return formatMarkdown(rows);
-	return formatConsole(rows, useColor);
-}
-//#endregion
 //#region src/eval-interchange/normalize.ts
 /**
 * Serialize tool arguments to the Vertex wire string format.
@@ -1431,6 +1164,36 @@ function outcomePassForCell(_caseId, _cellLabel, repetitions) {
 	if (graded.length === 0) return void 0;
 	return graded.every((r) => r.outcomeGrades.error === void 0 && r.outcomeGrades.summary.failed === 0);
 }
+/** Resolve judge metadata for envelope export (explicit options win). */
+async function resolveEnvelopeJudge(options) {
+	if (options.grading?.judge) return options.grading.judge;
+	if (options.gradingConfigPath) try {
+		return judgeInfoFromGradingConfig(await loadGradingConfig(resolve(options.gradingConfigPath)));
+	} catch {}
+	return resolveJudgeInfo({ adapter: "claude-code" });
+}
+/** Path to pass to {@link loadSuite} (directory layout uses the suite folder). */
+async function resolveSuiteLoadPath(suitePath) {
+	const abs = resolve(suitePath);
+	if (basename(abs) === "suite.yaml") return dirname(abs);
+	try {
+		if ((await stat(abs)).isDirectory()) return abs;
+	} catch {}
+	return abs;
+}
+/** Read suite YAML bytes for content hashing. */
+async function readSuiteYamlContent(suitePath) {
+	const loadPath = await resolveSuiteLoadPath(suitePath);
+	return readFile(basename(resolve(suitePath)) === "suite.yaml" ? resolve(suitePath) : join(loadPath, "suite.yaml"), "utf8");
+}
+async function resolveEnvelopeHarnessAdapter(options) {
+	if (options.harnessAdapter) return options.harnessAdapter;
+	if (options.suitePath) try {
+		const suite = await loadSuite(await resolveSuiteLoadPath(options.suitePath));
+		if (suite.adapter) return suite.adapter;
+	} catch {}
+	return "claude-code";
+}
 /**
 * Convert a {@link SuiteReport} (and optional grading) into a versioned
 * {@link EvalRunEnvelope} for storage or API handoff.
@@ -1442,7 +1205,7 @@ function outcomePassForCell(_caseId, _cellLabel, repetitions) {
 function buildEvalRunEnvelope(report, options = {}) {
 	const includeTranscript = options.includeTranscript !== false;
 	const includeRaw = options.includeRawStreamEvents === true;
-	const judge = options.grading?.judge ?? { id: "harness-eval/claude-grader" };
+	const judge = options.grading?.judge ?? resolveJudgeInfo({ adapter: "claude-code" });
 	const cells = report.cells.map((cell) => {
 		const prompt = cell.prompt ?? "";
 		const referenceTrajectoryConfig = cell.reference_trajectory;
@@ -1529,111 +1292,925 @@ function buildEvalRunEnvelope(report, options = {}) {
 	};
 }
 /**
-* Build an envelope from on-disk runner and grader JSON artifacts.
-*
-* Reads `reportPath` as a {@link SuiteReport}. When `gradingPath` is set, merges
-* outcome grades from a {@link SuiteGradingReport}. When `suitePath` is set,
-* attaches suite URI and SHA-256 content hash for reproducibility.
+* Build an envelope from on-disk runner and grader JSON artifacts.
+*
+* Reads `reportPath` as a {@link SuiteReport}. When `gradingPath` is set, merges
+* outcome grades from a {@link SuiteGradingReport}. When `suitePath` is set,
+* attaches suite URI and SHA-256 content hash for reproducibility.
+*
+* @param reportPath - Path to the suite run report JSON from `harness-eval run`.
+* @param options - Same build options as {@link buildEvalRunEnvelope}, plus file paths.
+*/
+async function buildEvalRunEnvelopeFromFiles(reportPath, options = {}) {
+	const reportText = await readFile(reportPath, "utf8");
+	const report = JSON.parse(reportText);
+	const harnessAdapter = await resolveEnvelopeHarnessAdapter({
+		harnessAdapter: options.harness?.adapter,
+		suitePath: options.suitePath
+	});
+	let grading = options.grading;
+	if (options.gradingPath) {
+		const gradingText = await readFile(options.gradingPath, "utf8");
+		const parsed = JSON.parse(gradingText);
+		const judge = parsed.judge ?? await resolveEnvelopeJudge({ gradingConfigPath: parsed.gradingConfigPath });
+		grading = {
+			gradedAt: parsed.gradedAt,
+			sourceReport: parsed.sourceReport,
+			results: parsed.results,
+			judge
+		};
+	}
+	let suite = options.suite;
+	if (options.suitePath) {
+		const content = await readSuiteYamlContent(options.suitePath);
+		suite = {
+			...suite,
+			uri: options.suitePath,
+			contentHash: createHash("sha256").update(content).digest("hex")
+		};
+	}
+	return buildEvalRunEnvelope(report, {
+		...options,
+		suite,
+		grading,
+		harness: {
+			...options.harness,
+			adapter: harnessAdapter
+		}
+	});
+}
+//#endregion
+//#region src/eval-interchange/projections.ts
+/** Trajectory instance keys emitted in stable order for JSONL export. */
+const TRAJECTORY_INSTANCE_KEYS = [
+	"exactMatch",
+	"inOrderMatch",
+	"anyOrderMatch",
+	"precision",
+	"recall",
+	"singleToolUse"
+];
+/**
+* Flatten one repetition into a trajectory dataset row.
+*
+* Pulls prompt from the cell, response from evaluationInstance, and falls
+* back to duration-based latency when enrich did not set latencySeconds.
+*/
+function repetitionToDatasetRow(cell, repetition) {
+	return {
+		caseId: cell.caseId,
+		repetitionIndex: repetition.repetitionIndex,
+		prompt: cell.prompt,
+		response: repetition.evaluationInstance?.response?.text,
+		evaluationInstance: repetition.evaluationInstance,
+		latencySeconds: repetition.latencySeconds ?? repetition.durationMs / 1e3,
+		failure: repetition.failure ?? (repetition.trajectory?.success ? 0 : 1),
+		humanRatings: cell.humanRatings
+	};
+}
+/**
+* Expand one repetition into type-tagged instance rows for EvaluateInstances.
+*
+* Returns an empty array when the repetition has no reference trajectory
+* (and therefore no trajectoryInstances block).
+*/
+function repetitionToInstanceRows(cell, repetition) {
+	if (!repetition.trajectoryInstances) return [];
+	const rows = [];
+	for (const key of TRAJECTORY_INSTANCE_KEYS) {
+		const instance = repetition.trajectoryInstances[key];
+		if (!instance) continue;
+		rows.push({
+			messageType: trajectoryInstanceMessageType(key),
+			caseId: cell.caseId,
+			repetitionIndex: repetition.repetitionIndex,
+			instance
+		});
+	}
+	return rows;
+}
+/**
+* Trajectory projection — all repetitions in the envelope as dataset rows.
+*/
+function toTrajectory(envelope) {
+	const rows = [];
+	for (const cell of envelope.cells) for (const repetition of cell.repetitions) rows.push(repetitionToDatasetRow(cell, repetition));
+	return rows;
+}
+/**
+* Instances projection — all trajectory metric instances as JSONL rows.
+*/
+function toInstancesJsonl(envelope) {
+	const rows = [];
+	for (const cell of envelope.cells) for (const repetition of cell.repetitions) rows.push(...repetitionToInstanceRows(cell, repetition));
+	return rows;
+}
+//#endregion
+//#region src/pipeline/resolve-inputs.ts
+/**
+* Resolve pipeline step inputs and outputs with precedence rules.
+*
+* Precedence: CLI override > explicit YAML > prior step in this run > default path on disk > error.
+*/
+/** Resolve absolute paths for enabled pipeline steps. */
+async function resolvePipelineInputs(options) {
+	const { suitePath, suiteDir, pipeline, steps, overrides } = options;
+	const executed = options.executed ?? {};
+	const stepSet = new Set(steps);
+	const resolved = { suitePath: resolve(suitePath) };
+	const defaultRunOutput = resolve(suiteDir, pipeline.run?.output ?? DEFAULT_PIPELINE_OUTPUTS.run);
+	const defaultGradeOutput = resolve(suiteDir, pipeline.grade?.output ?? DEFAULT_PIPELINE_OUTPUTS.grade);
+	if (stepSet.has("run") && pipeline.run) resolved.run = {
+		output: resolve(suiteDir, overrides?.run?.output ?? pipeline.run.output),
+		maxConcurrent: overrides?.run?.maxConcurrent ?? pipeline.run.maxConcurrent
+	};
+	if (stepSet.has("grade") && pipeline.grade) resolved.grade = {
+		input: await resolveReportPath({
+			explicit: overrides?.grade?.input ?? pipeline.grade.input,
+			executedOutput: executed.run?.output,
+			defaultPath: defaultRunOutput,
+			label: "grade input (report)"
+		}),
+		output: resolve(suiteDir, overrides?.grade?.output ?? pipeline.grade.output),
+		maxConcurrent: overrides?.grade?.maxConcurrent ?? pipeline.grade.maxConcurrent
+	};
+	if (stepSet.has("envelope") && pipeline.envelope) resolved.envelope = {
+		report: await resolveReportPath({
+			explicit: overrides?.envelope?.report ?? pipeline.envelope.report,
+			executedOutput: executed.run?.output,
+			defaultPath: defaultRunOutput,
+			label: "envelope report"
+		}),
+		grading: await resolveOptionalGradingPath({
+			explicit: overrides?.envelope?.grading ?? pipeline.envelope.grading,
+			executedOutput: executed.grade?.output,
+			defaultPath: defaultGradeOutput
+		}),
+		output: resolve(suiteDir, overrides?.envelope?.output ?? pipeline.envelope.output),
+		projection: overrides?.envelope?.projection ?? pipeline.envelope.projection ?? "envelope",
+		includeRawStreamEvents: pipeline.envelope.includeRawStreamEvents ?? false,
+		noTranscript: pipeline.envelope.noTranscript ?? false
+	};
+	return resolved;
+}
+/**
+* Resolve a required report path: explicit override → prior step output → default on disk.
+* Throws when none of the above exist.
+*/
+async function resolveReportPath(options) {
+	if (options.explicit) return resolve(options.explicit);
+	if (options.executedOutput) return resolve(options.executedOutput);
+	if (await pathExists(options.defaultPath)) return options.defaultPath;
+	throw new ConfigError(`pipeline: could not resolve ${options.label}; specify an explicit path or run the run step first`, options.defaultPath);
+}
+/** Resolve optional grading path; returns undefined when grading was not run and file is absent. */
+async function resolveOptionalGradingPath(options) {
+	if (options.explicit) return resolve(options.explicit);
+	if (options.executedOutput) return resolve(options.executedOutput);
+	if (await pathExists(options.defaultPath)) return options.defaultPath;
+}
+async function pathExists(filePath) {
+	try {
+		await stat(filePath);
+		return true;
+	} catch {
+		return false;
+	}
+}
+/**
+* Resolve a grading artifact path from a unified suite's `pipeline:` block.
+*
+* Used by `harness-eval envelope --suite` when `--grading` is omitted (spec C-7).
+* Checks `pipeline.envelope.grading` then default `pipeline.grade.output` on disk.
+*/
+async function resolveGradingArtifactFromSuite(suitePath) {
+	let doc;
+	try {
+		doc = await loadSuiteDocument(suitePath);
+	} catch {
+		return;
+	}
+	if (!doc.pipeline) return void 0;
+	const explicit = doc.pipeline.envelope?.grading;
+	if (explicit && await pathExists(explicit)) return explicit;
+	const defaultGrade = doc.pipeline.grade?.output;
+	if (defaultGrade && await pathExists(defaultGrade)) return defaultGrade;
+}
+/** Parse `--steps run,grade,envelope` against configured pipeline keys. */
+function parsePipelineSteps(pipeline, stepsArg) {
+	const configured = [];
+	if (pipeline.run !== void 0) configured.push("run");
+	if (pipeline.grade !== void 0) configured.push("grade");
+	if (pipeline.envelope !== void 0) configured.push("envelope");
+	if (configured.length === 0) throw new ConfigError("pipeline block has no steps configured");
+	if (!stepsArg) return configured;
+	const validStepNames = /* @__PURE__ */ new Set([
+		"run",
+		"grade",
+		"envelope"
+	]);
+	const requested = stepsArg.split(",").map((s) => s.trim()).filter(Boolean);
+	for (const step of requested) {
+		if (!validStepNames.has(step)) throw new ConfigError(`unknown pipeline step "${step}"; valid steps are: run, grade, envelope`);
+		if (!configured.includes(step)) throw new ConfigError(`pipeline step "${step}" is not configured in suite.yaml`);
+	}
+	const requestedSet = new Set(requested);
+	return configured.filter((step) => requestedSet.has(step));
+}
+/** Parent directory of suite.yaml. */
+function suiteDirectoryFromPath(suitePath) {
+	return dirname(resolve(suitePath));
+}
+//#endregion
+//#region src/cli/args.ts
+/** Parse process argv into command, positional args, and options. */
+function parseArgs(argv) {
+	const positional = [];
+	const options = {};
+	let command;
+	const args = [...argv];
+	if (args.length > 0 && !args[0].startsWith("-")) command = args.shift();
+	for (let i = 0; i < args.length; i++) {
+		const arg = args[i];
+		if (arg === "--") {
+			positional.push(...args.slice(i + 1));
+			break;
+		}
+		if (arg.startsWith("--")) {
+			const key = arg.slice(2);
+			const next = args[i + 1];
+			if (next && !next.startsWith("-")) {
+				options[key] = next;
+				i++;
+			} else options[key] = true;
+		} else if (arg.startsWith("-") && arg.length === 2) {
+			const key = arg.slice(1);
+			const next = args[i + 1];
+			if (next && !next.startsWith("-")) {
+				options[key] = next;
+				i++;
+			} else options[key] = true;
+		} else positional.push(arg);
+	}
+	return {
+		command,
+		positional,
+		options
+	};
+}
+/** Return a string option value, or undefined when absent or boolean. */
+function getOption(options, name) {
+	const v = options[name];
+	return typeof v === "string" ? v : void 0;
+}
+/** Parse an integer option with fallback when absent or non-numeric. */
+function getOptionInt(options, name, defaultValue) {
+	const v = getOption(options, name);
+	if (v === void 0) return defaultValue;
+	const n = Number.parseInt(v, 10);
+	if (!Number.isFinite(n)) return defaultValue;
+	return n;
+}
+/** True when a boolean flag is set or explicitly `"true"`. */
+function hasOption(options, name) {
+	const v = options[name];
+	return v === true || typeof v === "string" && v === "true";
+}
+//#endregion
+//#region src/cli/commands/envelope.ts
+/**
+* `harness-eval envelope` — build EvalRunEnvelope and interchange projections.
+*
+* Reads a suite run report (and optional grading JSON), builds a versioned
+* {@link EvalRunEnvelope}, and serializes one of three projections:
+*
+*   - `envelope` — full nested JSON document (default)
+*   - `trajectory` — JSONL of {@link EvalDatasetRow} per repetition
+*   - `instances` — JSONL of {@link InstancesJsonlRow} for Vertex batch upload
+*
+* Exit code 0 when behavioral pass, 1 when any cell failed assertions.
+*/
+const PROJECTIONS = /* @__PURE__ */ new Set([
+	"envelope",
+	"trajectory",
+	"instances"
+]);
+/**
+* Parse and validate `--projection` CLI flag.
+*
+* @returns `"envelope"` when omitted; `undefined` when value is invalid.
+*/
+function parseEnvelopeProjection(value) {
+	if (value === void 0) return "envelope";
+	if (PROJECTIONS.has(value)) return value;
+}
+/**
+* Serialize an envelope to stdout/file string for the chosen projection.
+*
+* Trajectory and instances projections emit NDJSON (one JSON object per line).
+*/
+function serializeEnvelopeProjection(envelope, projection) {
+	switch (projection) {
+		case "trajectory": return `${toTrajectory(envelope).map((row) => JSON.stringify(row)).join("\n")}\n`;
+		case "instances": return `${toInstancesJsonl(envelope).map((row) => JSON.stringify(row)).join("\n")}\n`;
+		default: return `${JSON.stringify(envelope, null, 2)}\n`;
+	}
+}
+/** Read harness-eval package version for envelope harness.frameworkVersion. */
+async function readFrameworkVersion() {
+	try {
+		const text = await readFile(join(dirname(fileURLToPath(import.meta.url)), "../../../package.json"), "utf8");
+		return JSON.parse(text).version;
+	} catch {
+		return;
+	}
+}
+/**
+* CLI entry point for the `envelope` subcommand.
+*
+* @returns Process exit code: 0 on behavioral pass, 1 on failure, 2 on usage/error.
+*/
+async function envelopeCommand(args) {
+	const reportPath = args.positional[0];
+	if (!reportPath) {
+		console.error("usage: harness-eval envelope <report.json> [--output path] [--grading path] [--suite path] [--projection envelope|trajectory|instances] [--include-raw-stream-events] [--no-transcript]");
+		return 2;
+	}
+	const outputPath = getOption(args.options, "output");
+	const suitePath = getOption(args.options, "suite");
+	let gradingPath = getOption(args.options, "grading");
+	if (!gradingPath && suitePath) gradingPath = await resolveGradingArtifactFromSuite(suitePath);
+	const projection = parseEnvelopeProjection(getOption(args.options, "projection"));
+	if (!projection) {
+		console.error("invalid --projection; expected envelope, trajectory, or instances");
+		return 2;
+	}
+	let envelope;
+	try {
+		const frameworkVersion = await readFrameworkVersion();
+		envelope = await buildEvalRunEnvelopeFromFiles(reportPath, {
+			gradingPath,
+			suitePath,
+			includeTranscript: !hasOption(args.options, "no-transcript"),
+			includeRawStreamEvents: hasOption(args.options, "include-raw-stream-events"),
+			harness: { frameworkVersion }
+		});
+	} catch (err) {
+		console.error(err instanceof Error ? err.message : String(err));
+		return 2;
+	}
+	const serialized = serializeEnvelopeProjection(envelope, projection);
+	if (outputPath) await writeFile(outputPath, serialized, "utf8");
+	else process.stdout.write(serialized);
+	return envelope.summary.behavioralPass ? 0 : 1;
+}
+//#endregion
+//#region src/pipeline/run-pipeline.ts
+/**
+* Orchestrate run → grade → envelope pipeline steps.
+*/
+/** Execute configured pipeline steps in order; stop on first failure. */
+async function runPipeline(doc, options = {}) {
+	if (!doc.pipeline) throw new ConfigError("suite document has no pipeline block", doc.suitePath);
+	const steps = parsePipelineSteps(doc.pipeline, options.steps);
+	const suiteDir = suiteDirectoryFromPath(doc.suitePath);
+	const executed = {};
+	let runReport;
+	let exitCode = 0;
+	for (const step of steps) {
+		const resolved = await resolvePipelineInputs({
+			suitePath: doc.suitePath,
+			suiteDir,
+			pipeline: doc.pipeline,
+			steps: [step],
+			executed,
+			overrides: options.overrides
+		});
+		if (step === "run" && resolved.run) {
+			const adapter = getAdapter(doc.suite.adapter ?? "claude-code");
+			runReport = await runSuite(doc.suite, {
+				adapter,
+				maxConcurrent: resolved.run.maxConcurrent ?? options.maxConcurrent ?? 4,
+				onProgress: options.onRunProgress
+			});
+			await writeFile(resolved.run.output, JSON.stringify(runReport, null, 2), "utf8");
+			executed.run = { output: resolved.run.output };
+			if (!runReport.cells.every((cell) => cell.passed)) return {
+				exitCode: 1,
+				stepsRun: steps.slice(0, steps.indexOf(step) + 1),
+				runReport
+			};
+			continue;
+		}
+		if (step === "grade" && resolved.grade) {
+			if (!doc.judge) throw new ConfigError("grade step requires inline judge: block in suite.yaml", doc.suitePath);
+			const gradeOptions = resolveGradeOptions({ judge: doc.judge }, {
+				sourceReport: resolved.grade.input,
+				maxConcurrent: resolved.grade.maxConcurrent
+			}, doc.suitePath);
+			const grading = await gradeReport(await loadSuiteReport(resolved.grade.input), {
+				...gradeOptions,
+				onProgress: options.onGradeProgress
+			});
+			await writeFile(resolved.grade.output, JSON.stringify(grading, null, 2), "utf8");
+			executed.grade = {
+				input: resolved.grade.input,
+				output: resolved.grade.output
+			};
+			if (!gradingReportPassed(grading)) return {
+				exitCode: 1,
+				stepsRun: steps.slice(0, steps.indexOf(step) + 1),
+				runReport
+			};
+			continue;
+		}
+		if (step === "envelope" && resolved.envelope) {
+			const envelope = await buildEvalRunEnvelopeFromFiles(resolved.envelope.report, {
+				gradingPath: resolved.envelope.grading,
+				suitePath: doc.suitePath,
+				includeTranscript: !resolved.envelope.noTranscript,
+				includeRawStreamEvents: resolved.envelope.includeRawStreamEvents,
+				harness: { frameworkVersion: options.frameworkVersion }
+			});
+			const serialized = serializeEnvelopeProjection(envelope, resolved.envelope.projection);
+			await writeFile(resolved.envelope.output, serialized, "utf8");
+			const behavioralFail = !envelope.summary.behavioralPass;
+			const outcomeFail = envelope.summary.outcomePass !== void 0 && !envelope.summary.outcomePass;
+			if (behavioralFail || outcomeFail) return {
+				exitCode: 1,
+				stepsRun: steps.slice(0, steps.indexOf(step) + 1),
+				runReport
+			};
+			continue;
+		}
+	}
+	return {
+		exitCode,
+		stepsRun: steps,
+		runReport
+	};
+}
+//#endregion
+//#region src/otel/attributes.ts
+/** Build a string-typed OTLP attribute. */
+function strAttr(key, value) {
+	return {
+		key,
+		value: { stringValue: value }
+	};
+}
+/** Build an integer-typed OTLP attribute (stored as decimal string). */
+function intAttr(key, value) {
+	return {
+		key,
+		value: { intValue: String(value) }
+	};
+}
+/** Build a boolean-typed OTLP attribute. */
+function boolAttr(key, value) {
+	return {
+		key,
+		value: { boolValue: value }
+	};
+}
+/** Build a JSON-serialized string attribute (common for message arrays). */
+function jsonAttr(key, value) {
+	return {
+		key,
+		value: { stringValue: JSON.stringify(value) }
+	};
+}
+//#endregion
+//#region src/otel/messages.ts
+/**
+* Map harness stop reasons to GenAI semconv finish_reason values.
+*
+* Unknown reasons pass through unchanged for forward compatibility.
+*/
+function mapStopReason(reason) {
+	if (!reason) return void 0;
+	switch (reason) {
+		case "end_turn": return "stop";
+		case "tool_use": return "tool_calls";
+		case "max_tokens": return "length";
+		case "stop_sequence": return "stop";
+		default: return reason;
+	}
+}
+/** Build a tool_call part from a {@link ToolCall}. */
+function toolCallPart(call) {
+	return {
+		type: "tool_call",
+		id: call.callId,
+		name: call.name,
+		arguments: call.args ?? {}
+	};
+}
+/** Build a tool_call_response part from a {@link ToolCall} result. */
+function toolResponsePart(call) {
+	return {
+		type: "tool_call_response",
+		id: call.callId,
+		result: call.result
+	};
+}
+/** Convert one assistant turn to a GenAI semconv assistant message. */
+function assistantMessageFromTurn(turn) {
+	const parts = [];
+	if (turn.text) parts.push({
+		type: "text",
+		content: turn.text
+	});
+	for (const call of turn.toolCalls) parts.push(toolCallPart(call));
+	const finish = mapStopReason(turn.stopReason);
+	return {
+		role: "assistant",
+		parts,
+		...finish ? { finish_reason: finish } : {}
+	};
+}
+/** Aggregate tool results from a turn into a single tool-role message, if any. */
+function toolResultsMessage(calls) {
+	const parts = calls.filter((c) => c.result !== null).map((c) => toolResponsePart(c));
+	if (parts.length === 0) return null;
+	return {
+		role: "tool",
+		parts
+	};
+}
+/**
+* Input history before the assistant turn at `turnIndex`.
+*/
+function inputMessagesBeforeTurn(view, turnIndex, prompt) {
+	const messages = [];
+	if (prompt) messages.push({
+		role: "user",
+		parts: [{
+			type: "text",
+			content: prompt
+		}]
+	});
+	for (let i = 0; i < turnIndex; i++) {
+		const turn = view.turns[i];
+		if (!turn) continue;
+		messages.push(assistantMessageFromTurn(turn));
+		const toolMsg = toolResultsMessage(turn.toolCalls);
+		if (toolMsg) messages.push(toolMsg);
+	}
+	return messages;
+}
+//#endregion
+//#region src/otel/types.ts
+/** OTLP span kinds (enum integers). */
+const SpanKind = {
+	INTERNAL: 1,
+	CLIENT: 2
+};
+/** OTLP status codes. */
+const StatusCode = {
+	UNSET: 0,
+	OK: 1,
+	ERROR: 2
+};
+//#endregion
+//#region src/otel/emitter.ts
+/**
+* TrajectoryView → OTLP JSON export using OpenTelemetry GenAI semantic conventions.
+*
+* Produces an `ExportTraceServiceRequest` suitable for OTLP/HTTP JSON ingestion.
+* Assertions continue to use {@link TrajectoryView} directly; this is export-only.
+*/
+const INSTRUMENTATION_VERSION = "0.1.0";
+/**
+* Map a {@link TrajectoryView} to OTLP trace JSON.
+*
+* Span tree (siblings under `invoke_agent`, not nested):
+* ```
+* invoke_agent
+* ├── chat {model}
+* ├── execute_tool {name}
+* ├── chat {model}
+* └── ...
+* ```
+*/
+function trajectoryToOtlp(view, options = {}) {
+	const agentName = options.agentName ?? "claude-code";
+	const providerName = options.providerName ?? "anthropic";
+	const serviceName = options.serviceName ?? "harness-eval";
+	const scopeName = options.instrumentationScope ?? "@alis-build/harness-eval";
+	const traceId = traceIdFromSession(view.meta.sessionId);
+	const rootSpanId = spanIdFromKey(traceId, "invoke_agent");
+	const durationMs = Math.max(view.usage.durationMs, 1);
+	const endMs = options.endTimeMs ?? Date.now();
+	const startMs = endMs - durationMs;
+	const rootStartNs = msToNs(startMs);
+	const rootEndNs = msToNs(endMs);
+	const spans = [];
+	const timings = buildSpanTimings(view, startMs, endMs);
+	spans.push({
+		traceId,
+		spanId: rootSpanId,
+		name: "invoke_agent",
+		kind: SpanKind.INTERNAL,
+		startTimeUnixNano: rootStartNs,
+		endTimeUnixNano: rootEndNs,
+		attributes: [
+			strAttr("gen_ai.operation.name", "invoke_agent"),
+			strAttr("gen_ai.agent.name", agentName),
+			strAttr("gen_ai.provider.name", providerName),
+			strAttr("gen_ai.conversation.id", view.meta.sessionId),
+			strAttr("gen_ai.request.model", view.meta.model),
+			strAttr("gen_ai.response.model", view.meta.model),
+			intAttr("gen_ai.usage.input_tokens", view.usage.inputTokens),
+			intAttr("gen_ai.usage.output_tokens", view.usage.outputTokens),
+			boolAttr("harness_eval.success", view.success)
+		],
+		status: viewStatus(view)
+	});
+	let opIndex = 0;
+	for (const turn of view.turns) {
+		const chatTiming = timings[opIndex++];
+		const chatSpanId = spanIdFromKey(traceId, `chat:${turn.turnIndex}`);
+		const inputMessages = inputMessagesBeforeTurn(view, turn.turnIndex, options.prompt);
+		const outputMessages = [assistantMessageFromTurn(turn)];
+		spans.push({
+			traceId,
+			spanId: chatSpanId,
+			parentSpanId: rootSpanId,
+			name: `chat ${view.meta.model}`,
+			kind: SpanKind.CLIENT,
+			startTimeUnixNano: chatTiming.startNs,
+			endTimeUnixNano: chatTiming.endNs,
+			attributes: [
+				strAttr("gen_ai.operation.name", "chat"),
+				strAttr("gen_ai.provider.name", providerName),
+				strAttr("gen_ai.request.model", view.meta.model),
+				strAttr("gen_ai.response.model", view.meta.model),
+				...inputMessages.length > 0 ? [jsonAttr("gen_ai.input.messages", inputMessages)] : [],
+				jsonAttr("gen_ai.output.messages", outputMessages),
+				...turn.stopReason ? [jsonAttr("gen_ai.response.finish_reasons", [mapStopReason(turn.stopReason) ?? turn.stopReason])] : []
+			],
+			status: { code: StatusCode.OK }
+		});
+		if (turn.toolCalls.length === 0) continue;
+		const toolTiming = timings[opIndex++];
+		for (const call of turn.toolCalls) {
+			const toolSpanId = spanIdFromKey(traceId, `tool:${call.callId}`);
+			spans.push({
+				traceId,
+				spanId: toolSpanId,
+				parentSpanId: rootSpanId,
+				name: `execute_tool ${call.name}`,
+				kind: SpanKind.INTERNAL,
+				startTimeUnixNano: toolTiming.startNs,
+				endTimeUnixNano: toolTiming.endNs,
+				attributes: [
+					strAttr("gen_ai.operation.name", "execute_tool"),
+					strAttr("gen_ai.provider.name", providerName),
+					strAttr("gen_ai.tool.name", call.name),
+					strAttr("gen_ai.tool.call.id", call.callId),
+					jsonAttr("gen_ai.tool.call.arguments", call.args ?? {}),
+					...call.result !== null ? [jsonAttr("gen_ai.tool.call.result", call.result)] : [],
+					...call.namespace ? [strAttr("harness_eval.tool.namespace", call.namespace)] : [],
+					boolAttr("harness_eval.tool.is_error", call.isError)
+				],
+				status: call.isError ? {
+					code: StatusCode.ERROR,
+					message: "tool reported error"
+				} : { code: StatusCode.OK }
+			});
+		}
+	}
+	return { resourceSpans: [{
+		resource: { attributes: [strAttr("service.name", serviceName), strAttr("gen_ai.agent.name", agentName)] },
+		scopeSpans: [{
+			scope: {
+				name: scopeName,
+				version: INSTRUMENTATION_VERSION
+			},
+			spans
+		}]
+	}] };
+}
+/** Alias for {@link trajectoryToOtlp} — matches implementation plan naming. */
+const emitOtel = trajectoryToOtlp;
+/** Map view success flag to OTLP span status on the root invoke_agent span. */
+function viewStatus(view) {
+	if (view.success) return { code: StatusCode.OK };
+	return {
+		code: StatusCode.ERROR,
+		message: "harness run did not complete successfully"
+	};
+}
+/**
+* Assign synthetic timestamps to chat and tool spans.
+*
+* Stream-json does not carry per-turn wall times, so we divide the session
+* duration evenly across chat/tool slots for OTLP consumers that require
+* start/end times on every span.
+*/
+function buildSpanTimings(view, startMs, endMs) {
+	const slots = [];
+	for (const turn of view.turns) {
+		slots.push("chat");
+		if (turn.toolCalls.length > 0) slots.push("tools");
+	}
+	if (slots.length === 0) return [];
+	const slotMs = Math.max(endMs - startMs, 1) / slots.length;
+	const timings = [];
+	let offset = startMs;
+	for (const slot of slots) {
+		const slotStart = offset;
+		const slotEnd = offset + slotMs;
+		timings.push({
+			startNs: msToNs(slotStart),
+			endNs: msToNs(slotEnd)
+		});
+		offset = slotEnd;
+	}
+	return timings;
+}
+/**
+* Derive a deterministic 128-bit trace id from the harness session id.
 *
-* @param reportPath - Path to the suite run report JSON from `harness-eval run`.
-* @param options - Same build options as {@link buildEvalRunEnvelope}, plus file paths.
+* Uses SHA-256 truncation so the same session always maps to the same trace.
 */
-async function buildEvalRunEnvelopeFromFiles(reportPath, options = {}) {
-	const reportText = await readFile(reportPath, "utf8");
-	const report = JSON.parse(reportText);
-	let grading = options.grading;
-	if (options.gradingPath) {
-		const gradingText = await readFile(options.gradingPath, "utf8");
-		const parsed = JSON.parse(gradingText);
-		grading = {
-			gradedAt: parsed.gradedAt,
-			sourceReport: parsed.sourceReport,
-			results: parsed.results,
-			judge: options.grading?.judge ?? { id: "harness-eval/claude-grader" }
-		};
-	}
-	let suite = options.suite;
-	if (options.suitePath) {
-		const content = await readFile(options.suitePath, "utf8");
-		suite = {
-			...suite,
-			uri: options.suitePath,
-			contentHash: createHash("sha256").update(content).digest("hex")
-		};
-	}
-	return buildEvalRunEnvelope(report, {
-		...options,
-		suite,
-		grading
-	});
+function traceIdFromSession(sessionId) {
+	return createHash("sha256").update(`harness-eval:trace:${sessionId}`).digest("hex").slice(0, 32).toUpperCase();
+}
+/**
+* Derive a deterministic 64-bit span id from trace id and a logical span key.
+*/
+function spanIdFromKey(traceId, key) {
+	return createHash("sha256").update(`${traceId}:span:${key}`).digest("hex").slice(0, 16).toUpperCase();
+}
+/** Convert milliseconds since epoch to OTLP nanosecond timestamp string. */
+function msToNs(ms) {
+	return String(Math.round(ms * 1e6));
 }
 //#endregion
-//#region src/eval-interchange/projections.ts
-/** Trajectory instance keys emitted in stable order for JSONL export. */
-const TRAJECTORY_INSTANCE_KEYS = [
-	"exactMatch",
-	"inOrderMatch",
-	"anyOrderMatch",
-	"precision",
-	"recall",
-	"singleToolUse"
-];
+//#region src/reporter/format-console.ts
+const RESET = "\x1B[0m";
+const GREEN = "\x1B[32m";
+const RED = "\x1B[31m";
+const YELLOW = "\x1B[33m";
 /**
-* Flatten one repetition into a trajectory dataset row.
+* Render renderable rows as ANSI-colored console output.
 *
-* Pulls prompt from the cell, response from evaluationInstance, and falls
-* back to duration-based latency when enrich did not set latencySeconds.
+* @param color When false, emit plain text without escape codes.
 */
-function repetitionToDatasetRow(cell, repetition) {
-	return {
-		caseId: cell.caseId,
-		repetitionIndex: repetition.repetitionIndex,
-		prompt: cell.prompt,
-		response: repetition.evaluationInstance?.response?.text,
-		evaluationInstance: repetition.evaluationInstance,
-		latencySeconds: repetition.latencySeconds ?? repetition.durationMs / 1e3,
-		failure: repetition.failure ?? (repetition.trajectory?.success ? 0 : 1),
-		humanRatings: cell.humanRatings
-	};
+function formatConsole(rows, color = true) {
+	const lines = [];
+	for (const row of rows) {
+		const status = row.passed ? color ? `${GREEN}PASS${RESET}` : "PASS" : color ? `${RED}FAIL${RESET}` : "FAIL";
+		const crashNote = row.adapterErrors > 0 ? ` ${color ? YELLOW : ""}[${row.adapterErrors} adapter errors]${color ? RESET : ""}` : "";
+		lines.push(`${row.caseId} @ ${row.cellLabel}  ${status}${crashNote}`);
+		if (row.category) lines.push(`  category: ${row.category}`);
+		for (const stat of row.stats) {
+			const marker = stat.meetsThreshold ? color ? `${GREEN}✓${RESET}` : "✓" : color ? `${RED}✗${RESET}` : "✗";
+			const rateStr = formatRate$1(stat);
+			const thresholdPct = (stat.threshold * 100).toFixed(0);
+			let line = `  ├─ ${stat.description}: ${rateStr} [threshold ${thresholdPct}%] ${marker}`;
+			if (stat.delta !== void 0 && stat.baselinePassRate !== void 0) {
+				const arrow = stat.delta >= 0 ? "↑" : "↓";
+				const basePct = (stat.baselinePassRate * 100).toFixed(0);
+				const curPct = (stat.passRate * 100).toFixed(0);
+				const deltaPct = (stat.delta * 100).toFixed(0);
+				line += `  (${basePct}% → ${curPct}% (${arrow}${deltaPct}%))`;
+			}
+			lines.push(line);
+		}
+		lines.push("");
+	}
+	return lines.join("\n").trimEnd();
+}
+/** Format pass rate for display, noting when all reps crashed. */
+function formatRate$1(stat) {
+	if (stat.evaluatedCount === 0) return `0/${stat.totalReps} (all reps crashed)`;
+	const pct = (stat.passRate * 100).toFixed(0);
+	return `${stat.passedCount}/${stat.evaluatedCount} (${pct}%)`;
 }
+//#endregion
+//#region src/reporter/format-json.ts
 /**
-* Expand one repetition into type-tagged instance rows for EvaluateInstances.
+* Serialize a suite report as indented JSON (no transformation).
 *
-* Returns an empty array when the repetition has no reference trajectory
-* (and therefore no trajectoryInstances block).
+* Used by `--format json` and `--output` persistence.
 */
-function repetitionToInstanceRows(cell, repetition) {
-	if (!repetition.trajectoryInstances) return [];
-	const rows = [];
-	for (const key of TRAJECTORY_INSTANCE_KEYS) {
-		const instance = repetition.trajectoryInstances[key];
-		if (!instance) continue;
-		rows.push({
-			messageType: trajectoryInstanceMessageType(key),
-			caseId: cell.caseId,
-			repetitionIndex: repetition.repetitionIndex,
-			instance
-		});
+function formatJson(report) {
+	return JSON.stringify(report, null, 2);
+}
+//#endregion
+//#region src/reporter/format-markdown.ts
+/** Render renderable rows as a GitHub-flavored markdown report. */
+function formatMarkdown(rows) {
+	const lines = ["# Harness Eval Report", ""];
+	for (const row of rows) {
+		const status = row.passed ? "PASS" : "FAIL";
+		const crashNote = row.adapterErrors > 0 ? ` (${row.adapterErrors} adapter errors)` : "";
+		lines.push(`## ${row.caseId} @ ${row.cellLabel} — ${status}${crashNote}`);
+		if (row.category) lines.push(`**Category:** ${row.category}`);
+		if (row.notes) lines.push("<details><summary>Notes</summary>", row.notes, "</details>");
+		lines.push("");
+		lines.push("| Assertion | Result | Threshold | Status |");
+		lines.push("| --- | --- | --- | --- |");
+		for (const stat of row.stats) {
+			const rateStr = formatRate(stat);
+			const threshold = `${(stat.threshold * 100).toFixed(0)}%`;
+			const statusCell = stat.meetsThreshold ? "✓" : "✗";
+			let result = rateStr;
+			if (stat.delta !== void 0 && stat.baselinePassRate !== void 0) {
+				const base = (stat.baselinePassRate * 100).toFixed(0);
+				const cur = (stat.passRate * 100).toFixed(0);
+				const d = (stat.delta * 100).toFixed(0);
+				const sign = stat.delta >= 0 ? "+" : "";
+				result += ` (${base}% → ${cur}%, ${sign}${d}%)`;
+			}
+			lines.push(`| ${stat.description} | ${result} | ${threshold} | ${statusCell} |`);
+		}
+		lines.push("");
 	}
-	return rows;
+	return lines.join("\n").trimEnd();
+}
+/** Format pass rate for markdown tables, noting when all reps crashed. */
+function formatRate(stat) {
+	if (stat.evaluatedCount === 0) return `0/${stat.totalReps} (all reps crashed)`;
+	const pct = (stat.passRate * 100).toFixed(0);
+	return `${stat.passedCount}/${stat.evaluatedCount} (${pct}%)`;
+}
+//#endregion
+//#region src/reporter/renderable.ts
+/** Map a suite report to formatter-ready rows (one per cell). */
+function toRenderableRows(report) {
+	return report.cells.map((cell) => cellToRow(cell));
 }
 /**
-* Trajectory projection — all repetitions in the envelope as dataset rows.
+* Attach baseline pass-rate deltas to matching rows.
+*
+* Rows without a matching baseline cell are returned unchanged.
 */
-function toTrajectory(envelope) {
-	const rows = [];
-	for (const cell of envelope.cells) for (const repetition of cell.repetitions) rows.push(repetitionToDatasetRow(cell, repetition));
-	return rows;
+function applyBaseline(rows, baseline) {
+	const baselineMap = new Map(baseline.cells.map((c) => [`${c.caseId}::${c.cell.label}`, c]));
+	return rows.map((row) => {
+		const baseCell = baselineMap.get(`${row.caseId}::${row.cellLabel}`);
+		if (!baseCell) return row;
+		const stats = row.stats.map((stat, i) => {
+			const baseStat = baseCell.assertionStats[i];
+			if (!baseStat) return stat;
+			const delta = stat.passRate - baseStat.passRate;
+			return {
+				...stat,
+				baselinePassRate: baseStat.passRate,
+				delta
+			};
+		});
+		return {
+			...row,
+			stats
+		};
+	});
+}
+/** Convert one {@link CellReport} to a {@link RenderableRow}. */
+function cellToRow(cell) {
+	const totalReps = cell.repetitions.length;
+	const stats = cell.assertionStats.map((s) => ({
+		description: s.description,
+		threshold: s.threshold,
+		passedCount: s.passedCount,
+		evaluatedCount: s.evaluatedCount,
+		totalReps,
+		adapterErrors: cell.adapterErrors,
+		passRate: s.passRate,
+		meetsThreshold: s.meetsThreshold
+	}));
+	return {
+		caseId: cell.caseId,
+		category: cell.category,
+		notes: cell.notes,
+		cellLabel: cell.cell.label,
+		passed: cell.passed,
+		adapterErrors: cell.adapterErrors,
+		totalReps,
+		stats
+	};
 }
+//#endregion
+//#region src/reporter/index.ts
 /**
-* Instances projection — all trajectory metric instances as JSONL rows.
+* Format a {@link SuiteReport} for console, markdown, or JSON output.
+*
+* JSON format bypasses the renderable intermediate model and serializes the
+* report directly. Console and markdown apply optional baseline deltas.
 */
-function toInstancesJsonl(envelope) {
-	const rows = [];
-	for (const cell of envelope.cells) for (const repetition of cell.repetitions) rows.push(...repetitionToInstanceRows(cell, repetition));
-	return rows;
+function formatReport(report, options) {
+	if (options.format === "json") return formatJson(report);
+	let rows = toRenderableRows(report);
+	if (options.baseline) rows = applyBaseline(rows, options.baseline);
+	const useColor = options.color ?? options.format === "console";
+	if (options.format === "markdown") return formatMarkdown(rows);
+	return formatConsole(rows, useColor);
 }
 //#endregion
-export { loadSuiteReport as C, trajectoryToOtlp as D, emitOtel as E, EVAL_RUN_SCHEMA_VERSION as O, gradeReport as S, createClaudeGrader as T, serializeToolInput as _, enrichRepetitionWithProtojson as a, gradingReportPassed as b, parseToolInput as c, trajectoryInOrderMatch as d, trajectoryPrecision as f, toTrajectoryInstances as g, toEvaluationInstance as h, buildEvalRunEnvelopeFromFiles as i, TRAJECTORY_SCHEMA_VERSION as k, trajectoryAnyOrderMatch as l, trajectorySingleToolUse as m, toTrajectory as n, toHarnessMetrics as o, trajectoryRecall as p, buildEvalRunEnvelope as r, computeTrajectoryMetrics as s, toInstancesJsonl as t, trajectoryExactMatch as u, formatReport as v, trajectoryToTranscript as w, resolveGradeOptions as x, formatGradingConsole as y };
+export { serializeToolInput as A, TRAJECTORY_SCHEMA_VERSION as B, trajectoryExactMatch as C, trajectorySingleToolUse as D, trajectoryRecall as E, loadSuiteReport as F, trajectoryToTranscript as I, createCodexGrader as L, gradingReportPassed as M, resolveGradeOptions as N, toEvaluationInstance as O, gradeReport as P, createClaudeGrader as R, trajectoryAnyOrderMatch as S, trajectoryPrecision as T, buildEvalRunEnvelopeFromFiles as _, envelopeCommand as a, computeTrajectoryMetrics as b, getOptionInt as c, resolveGradingArtifactFromSuite as d, resolvePipelineInputs as f, buildEvalRunEnvelope as g, toTrajectory as h, runPipeline as i, formatGradingConsole as j, toTrajectoryInstances as k, hasOption as l, toInstancesJsonl as m, emitOtel as n, parseEnvelopeProjection as o, suiteDirectoryFromPath as p, trajectoryToOtlp as r, getOption as s, formatReport as t, parseArgs as u, enrichRepetitionWithProtojson as v, trajectoryInOrderMatch as w, parseToolInput as x, toHarnessMetrics as y, EVAL_RUN_SCHEMA_VERSION as z };
 
-//# sourceMappingURL=projections-BcX7w-f6.js.map
+//# sourceMappingURL=reporter-Biy-5-9M.js.map