@tangle-network/agent-eval 0.128.2 → 0.130.0

package/dist/off-policy-mskQw8Mb.d.ts

@@ -0,0 +1,153 @@
+//#region src/rl/off-policy.d.ts
+/**
+ * Off-policy evaluation primitives.
+ *
+ * Standard inverse-probability-weighted (IPS), self-normalized
+ * importance-weighted (SNIPS), and doubly-robust (DR) estimators for the
+ * value of a *target* policy given trajectories collected under a
+ * *behavior* policy. This is the canonical RL eval task: "we have last
+ * week's runs, we changed the policy — how would the new one do without
+ * re-running?"
+ *
+ * The math here is textbook (Dudík, Langford, Li 2011 for DR; Swaminathan
+ * & Joachims 2015 for SNIPS) but the *application* to LLM-agent
+ * evaluation needs care:
+ *
+ *   - The "policy" is the (prompt, tool config, model snapshot) triple.
+ *     Two policies have the same probability over an action *iff* their
+ *     LLM call would emit the same token with the same probability —
+ *     which is generally unknowable without the model log-probs.
+ *   - For LLM agents, propensity scores must be supplied by the caller
+ *     (logged in the trace, recovered from token log-probs, or estimated
+ *     via a learned propensity model). We do NOT estimate propensity here.
+ *   - Doubly-robust requires two outputs from a Q-function: its prediction
+ *     for the logged action and its expectation under the target policy.
+ *     Consumers compute these with a tabular estimate, regression fit, or
+ *     learned reward model before constructing the trajectories.
+ *
+ * Bias / variance tradeoffs:
+ *   - IPS: unbiased; high variance for small overlap, infinite variance
+ *     when target has support outside behavior.
+ *   - SNIPS: lower variance, slight bias; usually preferred in practice.
+ *   - DR: doubly-robust — unbiased if either propensity OR Q-function is
+ *     correct. Lowest practical variance when Q is decent. Use this.
+ *
+ * Caveat the panel will land: on the LLM-agent setting, propensity scores
+ * recovered from token log-probs are noisy, the action space is enormous,
+ * and overlap is often poor. These estimators are useful but not magic;
+ * complement with `replayCampaign` (exact replay where the request hashes
+ * match) for high-confidence answers and OPE for the gap.
+ */
+interface OffPolicyTrajectory {
+  /** Stable id, for traceability through the dataset. */
+  runId: string;
+  /** Reward observed under the behavior policy (the realized outcome). */
+  reward: number;
+  /**
+   * Behavior-policy probability of the action that was taken. For LLM
+   * agents this is typically `exp(sum(token_log_probs))` over the chosen
+   * trajectory. Must be in (0, 1].
+   */
+  behaviorProb: number;
+  /**
+   * Target-policy probability of the same action. For replay-style
+   * counterfactual evaluation this is what the *new* policy would have
+   * assigned to the *old* trajectory. Must be in [0, 1].
+   */
+  targetProb: number;
+  /**
+   * Model-based reward prediction for the action selected by the behavior
+   * policy: `Q_hat(context, loggedAction)`. Supply this together with
+   * `vHatTarget` for contextual-bandit doubly-robust estimation.
+   */
+  qHatChosen?: number | null;
+  /**
+   * Expected model-based reward under the target policy:
+   * `sum_action targetPolicy(action | context) * Q_hat(context, action)`.
+   * Supply this together with `qHatChosen`. For an honest evaluation, both
+   * values must come from a model cross-fitted or trained outside this row.
+   */
+  vHatTarget?: number | null;
+}
+interface OffPolicyContributionCounts {
+  /** Contributions using the contextual-bandit doubly-robust formula. */
+  dr: number;
+  /** Contributions using exact IPS because no reward-model estimate was supplied. */
+  ipsFallback: number;
+}
+interface OffPolicyEstimate {
+  /** Estimated value of the target policy. */
+  value: number;
+  /** Standard error of the estimate. */
+  standardError: number;
+  /** Effective sample size (Kong 1992). Lower = more reliance on a few high-weight samples. */
+  effectiveSampleSize: number;
+  /** Number of trajectories used. */
+  n: number;
+  /**
+   * Diagnostic: maximum importance weight observed. Large values (>>10x
+   * mean) are a red flag — variance is dominated by a few outliers.
+   */
+  maxImportanceWeight: number;
+  /** Populated by `doublyRobust` to expose which formula each row used. */
+  contributionCounts?: OffPolicyContributionCounts;
+}
+interface OffPolicyOptions {
+  /**
+   * Cap importance weights at this value (Ionides 2008 truncated IS) to
+   * trade unbiasedness for variance reduction. Default `Infinity` (no cap).
+   * Set e.g. `10` for stable estimates when the policies are close.
+   */
+  weightCap?: number;
+  /** Reward clipping range. Default `[0, 1]`. */
+  rewardClip?: {
+    low: number;
+    high: number;
+  };
+}
+/**
+ * Inverse Probability Weighting (Horvitz-Thompson). Unbiased estimator
+ * of E[reward under target policy]. Variance scales with the spread of
+ * target/behavior ratios.
+ */
+declare function inverseProbabilityWeighting(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): OffPolicyEstimate;
+/**
+ * Self-Normalized Importance Sampling. Lower variance than vanilla IPS at
+ * the cost of small bias (vanishing as N grows). The right default for
+ * LLM-agent evaluation where overlap is often poor.
+ */
+declare function selfNormalizedImportanceWeighting(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): OffPolicyEstimate;
+/**
+ * Doubly-robust off-policy estimator (Dudík, Langford, Li 2011).
+ *
+ *     V_DR = (1/N) * sum_i [ v_hat_target_i
+ *              + (target_prob_i / behavior_prob_i) * (r_i - q_hat_chosen_i) ]
+ *
+ * Unbiased if EITHER:
+ *   - the importance ratios are correct (IPS-style validity), OR
+ *   - the Q-hat function is correct (model-based validity).
+ *
+ * In practice both are imperfect, but the residual bias is the *product*
+ * of both errors — much smaller than either alone. This is why DR is the
+ * default in production OPE pipelines.
+ *
+ * `qHatChosen` and `vHatTarget` must be supplied together. Rows with neither
+ * use the exact IPS contribution. `contributionCounts` makes the mix explicit
+ * in the result.
+ * Callers must cross-fit the Q-function or train it on independent rows;
+ * fitting and evaluating Q on the same outcomes leaks the answer.
+ */
+declare function doublyRobust(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): OffPolicyEstimate;
+/**
+ * Convenience: run all three estimators and return them side-by-side.
+ * The recommended diagnostic — agreement across estimators is a much
+ * stronger signal than any single one.
+ */
+declare function offPolicyEstimateAll(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): {
+  ips: OffPolicyEstimate;
+  snips: OffPolicyEstimate;
+  dr: OffPolicyEstimate;
+};
+//#endregion
+export { doublyRobust as a, selfNormalizedImportanceWeighting as c, OffPolicyTrajectory as i, OffPolicyEstimate as n, inverseProbabilityWeighting as o, OffPolicyOptions as r, offPolicyEstimateAll as s, OffPolicyContributionCounts as t };
+//# sourceMappingURL=off-policy-mskQw8Mb.d.ts.map

package/dist/off-policy-mskQw8Mb.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"off-policy-mskQw8Mb.d.ts","names":[],"sources":["../src/rl/off-policy.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UA0CiB;;EAEf;;EAEA;;;;;;EAMA;;;;;;EAMA;;;;;;EAMA;;;;;;;EAOA;;UAGe;;EAEf;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;EAKA;;EAEA,qBAAqB;;UAGN;;;;;;EAMf;;EAEA;IAAe;IAAa;;;;;;;;iBAQd,4BACd,cAAc,uBACd,OAAM,mBACL;;;;;;iBA4Ca,kCACd,cAAc,uBACd,OAAM,mBACL;;;;;;;;;;;;;;;;;;;;;iBAwDa,aACd,cAAc,uBACd,OAAM,mBACL;;;;;;iBAuEa,qBACd,cAAc,uBACd,OAAM;EACH,KAAK;EAAmB,OAAO;EAAmB,IAAI"}

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.128.2",
+    "version": "0.130.0",
     "description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.0.0. Bumps on breaking changes to request/response schemas.",
     "contact": {
       "name": "Tangle Network",

package/dist/opencode-sqlite-BGrHeDu3.js

@@ -0,0 +1,318 @@
+import { createRequire } from "node:module";
+import { join } from "node:path";
+import { readFile, readdir } from "node:fs/promises";
+import { homedir } from "node:os";
+//#region src/rollout/readers/claude-jsonl.ts
+/**
+* Backfill reader over Claude Code project transcripts
+* (~/.claude/projects/<cwd-slug>/<sessionId>.jsonl) → canonical
+* chat-with-tools messages plus per-session token usage.
+*
+* Transcript lines consumed: type:"user" (string content or content blocks —
+* text + tool_result) and type:"assistant" (content blocks — thinking, text,
+* tool_use; message.usage carries tokens). Sidechain lines (isSidechain=true,
+* subagent threads) are separate invocations and are excluded from the main
+* transcript. Everything else (queue-operation, attachment, last-prompt…) is
+* transport metadata, not conversation.
+*/
+const DEFAULT_CLAUDE_PROJECTS_DIR = join(homedir(), ".claude", "projects");
+/** Claude Code's project-directory slug for a working directory. */
+function claudeProjectSlug(cwd) {
+	return cwd.replace(/[^a-zA-Z0-9-]/g, "-");
+}
+/** Transcript files recorded for sessions launched from `cwd`. */
+async function findClaudeTranscripts(cwd, projectsDir = DEFAULT_CLAUDE_PROJECTS_DIR) {
+	const dir = join(projectsDir, claudeProjectSlug(cwd));
+	return (await readdir(dir).catch(() => [])).filter((n) => n.endsWith(".jsonl")).sort().map((n) => ({
+		sessionId: n.replace(/\.jsonl$/, ""),
+		path: join(dir, n)
+	}));
+}
+const isRecord$1 = (v) => typeof v === "object" && v !== null && !Array.isArray(v);
+/** Parse transcript jsonl text into conversation lines. Non-conversation lines are dropped. */
+function parseClaudeEntries(raw) {
+	const out = [];
+	for (const line of raw.split("\n")) {
+		if (!line.trim()) continue;
+		let entry;
+		try {
+			const parsed = JSON.parse(line);
+			if (!isRecord$1(parsed)) continue;
+			entry = parsed;
+		} catch {
+			continue;
+		}
+		if (entry.type !== "user" && entry.type !== "assistant") continue;
+		const message = entry.message;
+		if (!isRecord$1(message)) continue;
+		out.push({
+			type: entry.type,
+			timestamp: typeof entry.timestamp === "string" ? entry.timestamp : null,
+			message,
+			toolUseResult: entry.toolUseResult,
+			isSidechain: entry.isSidechain === true,
+			agentId: typeof entry.agentId === "string" ? entry.agentId : null
+		});
+	}
+	return out;
+}
+function blockText(content) {
+	if (typeof content === "string") return content;
+	if (!Array.isArray(content)) return "";
+	return content.filter((b) => isRecord$1(b) && b.type === "text" && typeof b.text === "string").map((b) => b.text).join("\n");
+}
+/** Parse one transcript jsonl into canonical messages + usage totals. */
+async function readClaudeTranscript(path, options = {}) {
+	return transcriptFromEntries(parseClaudeEntries(await readFile(path, "utf8")), options);
+}
+/** The messages+usage projection of already-parsed entries. */
+function transcriptFromEntries(entries, options = {}) {
+	const wantSidechain = options.includeSidechain === true;
+	const messages = [];
+	const usage = {
+		tokensIn: 0,
+		tokensOut: 0,
+		cacheRead: 0,
+		cacheWrite: 0
+	};
+	let startedAt = null;
+	let endedAt = null;
+	let model = null;
+	let lastAssistantApiId = null;
+	let lastAssistantIndex = -1;
+	for (const entry of entries) {
+		if (entry.isSidechain !== wantSidechain) continue;
+		const message = entry.message;
+		if (entry.timestamp !== null) {
+			if (startedAt === null) startedAt = entry.timestamp;
+			endedAt = entry.timestamp;
+		}
+		if (entry.type === "user") {
+			lastAssistantApiId = null;
+			lastAssistantIndex = -1;
+			const content = message.content;
+			if (typeof content === "string") {
+				messages.push({
+					role: "user",
+					content
+				});
+				continue;
+			}
+			if (!Array.isArray(content)) continue;
+			let userText = "";
+			for (const block of content) {
+				if (!isRecord$1(block)) continue;
+				if (block.type === "tool_result" && typeof block.tool_use_id === "string") messages.push({
+					role: "tool",
+					tool_call_id: block.tool_use_id,
+					content: blockText(block.content) || (typeof block.content === "string" ? block.content : "")
+				});
+				else if (block.type === "text" && typeof block.text === "string") userText += (userText.length > 0 ? "\n" : "") + block.text;
+			}
+			if (userText.length > 0) messages.push({
+				role: "user",
+				content: userText
+			});
+			continue;
+		}
+		if (typeof message.model === "string") model = message.model;
+		const apiId = typeof message.id === "string" ? message.id : null;
+		const continuesTurn = apiId !== null && apiId === lastAssistantApiId && lastAssistantIndex >= 0;
+		const msgUsage = message.usage;
+		if (isRecord$1(msgUsage) && !continuesTurn) {
+			usage.tokensIn += typeof msgUsage.input_tokens === "number" ? msgUsage.input_tokens : 0;
+			usage.tokensOut += typeof msgUsage.output_tokens === "number" ? msgUsage.output_tokens : 0;
+			usage.cacheRead += typeof msgUsage.cache_read_input_tokens === "number" ? msgUsage.cache_read_input_tokens : 0;
+			usage.cacheWrite += typeof msgUsage.cache_creation_input_tokens === "number" ? msgUsage.cache_creation_input_tokens : 0;
+		}
+		const content = message.content;
+		if (!Array.isArray(content)) continue;
+		let reasoning = "";
+		let text = "";
+		const toolCalls = [];
+		for (const block of content) {
+			if (!isRecord$1(block)) continue;
+			if (block.type === "thinking" && typeof block.thinking === "string" && block.thinking.length > 0) reasoning += (reasoning.length > 0 ? "\n" : "") + block.thinking;
+			else if (block.type === "text" && typeof block.text === "string") text += (text.length > 0 ? "\n" : "") + block.text;
+			else if (block.type === "tool_use" && typeof block.id === "string") toolCalls.push({
+				id: block.id,
+				type: "function",
+				function: {
+					name: typeof block.name === "string" ? block.name : "unknown",
+					arguments: JSON.stringify(block.input ?? {})
+				}
+			});
+		}
+		if (reasoning.length === 0 && text.length === 0 && toolCalls.length === 0) continue;
+		if (continuesTurn) {
+			const prev = messages[lastAssistantIndex];
+			if (text.length > 0) prev.content = prev.content === null ? text : `${prev.content}\n${text}`;
+			if (reasoning.length > 0) prev.reasoning_content = prev.reasoning_content === void 0 ? reasoning : `${prev.reasoning_content}\n${reasoning}`;
+			if (toolCalls.length > 0) prev.tool_calls = [...prev.tool_calls ?? [], ...toolCalls];
+			continue;
+		}
+		messages.push({
+			role: "assistant",
+			content: text.length > 0 ? text : null,
+			...reasoning.length > 0 ? { reasoning_content: reasoning } : {},
+			...toolCalls.length > 0 ? { tool_calls: toolCalls } : {}
+		});
+		lastAssistantApiId = apiId;
+		lastAssistantIndex = messages.length - 1;
+	}
+	return {
+		messages,
+		usage,
+		startedAt,
+		endedAt,
+		model
+	};
+}
+//#endregion
+//#region src/rollout/readers/opencode-sqlite.ts
+/**
+* Read-only backfill reader over the opencode sqlite store
+* (~/.local/share/opencode/opencode.db) → canonical chat-with-tools messages.
+*
+* Schema consumed (observed, 2026-07): `session` rows carry directory /
+* parent_id / agent / model / cost / tokens_*; `message` rows carry a JSON
+* `data` blob ({role, modelID, providerID, tokens, cost, finish}); `part`
+* rows carry the actual content ({type: text|reasoning|tool|step-start|
+* step-finish|snapshot…}). Tool parts hold {callID, state:{input, output,
+* status}} — both the call and its result, which we split into an assistant
+* tool_call plus a role:"tool" result message.
+*
+* The store is mutable and can be corrupt (a `.corrupt-bak` sibling ships
+* next to it in the wild), so `openOpencodeDb` returns null instead of
+* throwing — callers record a gap line, never crash the backfill.
+*/
+const DEFAULT_OPENCODE_DB = join(homedir(), ".local", "share", "opencode", "opencode.db");
+const nodeRequire = createRequire(import.meta.url);
+/** Open the store read-only; null = unavailable/corrupt (caller records a gap). */
+async function openOpencodeDb(path = DEFAULT_OPENCODE_DB) {
+	try {
+		const { DatabaseSync } = nodeRequire("node:sqlite");
+		const db = new DatabaseSync(path, { readOnly: true });
+		db.prepare("SELECT id FROM session LIMIT 1").get();
+		return db;
+	} catch {
+		return null;
+	}
+}
+const isRecord = (v) => typeof v === "object" && v !== null && !Array.isArray(v);
+function parseSessionRow(row) {
+	let model = null;
+	if (typeof row.model === "string" && row.model.length > 0) try {
+		const parsed = JSON.parse(row.model);
+		if (isRecord(parsed)) model = parsed;
+	} catch {
+		model = null;
+	}
+	return {
+		id: String(row.id),
+		parentId: row.parent_id === null || row.parent_id === void 0 ? null : String(row.parent_id),
+		directory: String(row.directory),
+		agent: row.agent === null || row.agent === void 0 ? null : String(row.agent),
+		model,
+		costUsd: Number(row.cost ?? 0),
+		tokensInput: Number(row.tokens_input ?? 0),
+		tokensOutput: Number(row.tokens_output ?? 0),
+		tokensReasoning: Number(row.tokens_reasoning ?? 0),
+		tokensCacheRead: Number(row.tokens_cache_read ?? 0),
+		tokensCacheWrite: Number(row.tokens_cache_write ?? 0),
+		timeCreated: Number(row.time_created ?? 0),
+		timeUpdated: Number(row.time_updated ?? 0)
+	};
+}
+const SESSION_COLUMNS = "id, parent_id, directory, agent, model, cost, tokens_input, tokens_output, tokens_reasoning, tokens_cache_read, tokens_cache_write, time_created, time_updated";
+/** Sessions whose cwd is `directory` (the worker-clone join key). */
+function findOpencodeSessionsByDirectory(db, directory) {
+	return db.prepare(`SELECT ${SESSION_COLUMNS} FROM session WHERE directory = ? ORDER BY time_created`).all(directory).map(parseSessionRow);
+}
+function findOpencodeSessionById(db, sessionId) {
+	const row = db.prepare(`SELECT ${SESSION_COLUMNS} FROM session WHERE id = ?`).get(sessionId);
+	return row === void 0 ? null : parseSessionRow(row);
+}
+function toolResultContent(output) {
+	if (typeof output === "string") return output;
+	if (output === null || output === void 0) return "";
+	return JSON.stringify(output);
+}
+/**
+* Convert one session's message+part rows into canonical messages.
+* An opencode assistant message row spans several model steps; each step's
+* parts (reasoning → text → tool …) become one assistant message followed by
+* the role:"tool" results of its calls, preserving order.
+*/
+function readOpencodeSessionMessages(db, sessionId) {
+	const messageRows = db.prepare("SELECT id, data FROM message WHERE session_id = ? ORDER BY time_created, id").all(sessionId);
+	const partsStmt = db.prepare("SELECT data FROM part WHERE message_id = ? ORDER BY id");
+	const messages = [];
+	for (const messageRow of messageRows) {
+		let data;
+		try {
+			const parsed = JSON.parse(messageRow.data);
+			if (!isRecord(parsed)) continue;
+			data = parsed;
+		} catch {
+			continue;
+		}
+		const parts = [];
+		for (const row of partsStmt.all(messageRow.id)) try {
+			const parsed = JSON.parse(row.data);
+			if (isRecord(parsed)) parts.push(parsed);
+		} catch {}
+		if (data.role === "user") {
+			const text = parts.filter((p) => p.type === "text" && typeof p.text === "string").map((p) => p.text).join("\n");
+			messages.push({
+				role: "user",
+				content: text
+			});
+			continue;
+		}
+		if (data.role !== "assistant") continue;
+		const steps = [];
+		let current = [];
+		for (const part of parts) {
+			if (part.type === "step-start") {
+				if (current.length > 0) steps.push(current);
+				current = [];
+				continue;
+			}
+			if (part.type === "step-finish" || part.type === "snapshot" || part.type === "patch") continue;
+			current.push(part);
+		}
+		if (current.length > 0) steps.push(current);
+		for (const step of steps) {
+			const reasoning = step.filter((p) => p.type === "reasoning" && typeof p.text === "string" && p.text.length > 0).map((p) => p.text).join("\n");
+			const text = step.filter((p) => p.type === "text" && typeof p.text === "string").map((p) => p.text).join("\n");
+			const toolParts = step.filter((p) => p.type === "tool" && typeof p.callID === "string");
+			const toolCalls = toolParts.map((p) => ({
+				id: p.callID,
+				type: "function",
+				function: {
+					name: p.tool ?? "unknown",
+					arguments: JSON.stringify(p.state?.input ?? {})
+				}
+			}));
+			if (reasoning.length === 0 && text.length === 0 && toolCalls.length === 0) continue;
+			messages.push({
+				role: "assistant",
+				content: text.length > 0 ? text : null,
+				...reasoning.length > 0 ? { reasoning_content: reasoning } : {},
+				...toolCalls.length > 0 ? { tool_calls: toolCalls } : {}
+			});
+			for (const p of toolParts) messages.push({
+				role: "tool",
+				tool_call_id: p.callID,
+				name: p.tool ?? "unknown",
+				content: toolResultContent(p.state?.output)
+			});
+		}
+	}
+	return messages;
+}
+//#endregion
+export { readOpencodeSessionMessages as a, findClaudeTranscripts as c, transcriptFromEntries as d, openOpencodeDb as i, parseClaudeEntries as l, findOpencodeSessionById as n, DEFAULT_CLAUDE_PROJECTS_DIR as o, findOpencodeSessionsByDirectory as r, claudeProjectSlug as s, DEFAULT_OPENCODE_DB as t, readClaudeTranscript as u };
+
+//# sourceMappingURL=opencode-sqlite-BGrHeDu3.js.map

package/dist/opencode-sqlite-BGrHeDu3.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"opencode-sqlite-BGrHeDu3.js","names":["isRecord"],"sources":["../src/rollout/readers/claude-jsonl.ts","../src/rollout/readers/opencode-sqlite.ts"],"sourcesContent":["/**\n * Backfill reader over Claude Code project transcripts\n * (~/.claude/projects/<cwd-slug>/<sessionId>.jsonl) → canonical\n * chat-with-tools messages plus per-session token usage.\n *\n * Transcript lines consumed: type:\"user\" (string content or content blocks —\n * text + tool_result) and type:\"assistant\" (content blocks — thinking, text,\n * tool_use; message.usage carries tokens). Sidechain lines (isSidechain=true,\n * subagent threads) are separate invocations and are excluded from the main\n * transcript. Everything else (queue-operation, attachment, last-prompt…) is\n * transport metadata, not conversation.\n */\n\nimport { readdir, readFile } from 'node:fs/promises'\nimport { homedir } from 'node:os'\nimport { join } from 'node:path'\nimport type { ChatMessage, ChatToolCall } from '../schema'\n\nexport const DEFAULT_CLAUDE_PROJECTS_DIR = join(homedir(), '.claude', 'projects')\n\n/** Claude Code's project-directory slug for a working directory. */\nexport function claudeProjectSlug(cwd: string): string {\n  return cwd.replace(/[^a-zA-Z0-9-]/g, '-')\n}\n\nexport interface ClaudeTranscriptRef {\n  sessionId: string\n  path: string\n}\n\n/** Transcript files recorded for sessions launched from `cwd`. */\nexport async function findClaudeTranscripts(\n  cwd: string,\n  projectsDir: string = DEFAULT_CLAUDE_PROJECTS_DIR,\n): Promise<ClaudeTranscriptRef[]> {\n  const dir = join(projectsDir, claudeProjectSlug(cwd))\n  const names = await readdir(dir).catch(() => [])\n  return names\n    .filter((n) => n.endsWith('.jsonl'))\n    .sort()\n    .map((n) => ({ sessionId: n.replace(/\\.jsonl$/, ''), path: join(dir, n) }))\n}\n\nexport interface ClaudeUsageTotals {\n  tokensIn: number\n  tokensOut: number\n  cacheRead: number\n  cacheWrite: number\n}\n\nexport interface ClaudeTranscript {\n  messages: ChatMessage[]\n  usage: ClaudeUsageTotals\n  /** Timestamp of the first conversation line; null = empty transcript. */\n  startedAt: string | null\n  endedAt: string | null\n  model: string | null\n}\n\nconst isRecord = (v: unknown): v is Record<string, unknown> =>\n  typeof v === 'object' && v !== null && !Array.isArray(v)\n\n/**\n * One conversation line of a transcript, still in Claude Code's own shape.\n *\n * This is the single line-level parse of the format. `readClaudeTranscript`\n * projects it to canonical messages + usage; the supervision-tree reader\n * (`src/supervisor-run/claude-code-reader.ts`) projects the SAME entries to\n * spawn/settle/steer instants. Two projections, one parser — a second\n * transcript parser is how the two views silently disagree.\n */\nexport interface ClaudeEntry {\n  readonly type: 'user' | 'assistant'\n  /** ISO instant of the line; null when the line carried none. */\n  readonly timestamp: string | null\n  /** The Anthropic message body (`role`, `content`, `model`, `usage`). */\n  readonly message: Record<string, unknown>\n  /** Claude Code's structured tool result, when the line carries one. */\n  readonly toolUseResult: unknown\n  /** True on subagent threads — a separate invocation, not this transcript's turn. */\n  readonly isSidechain: boolean\n  /** Subagent id Claude Code stamps on sidechain lines; null on main-thread lines. */\n  readonly agentId: string | null\n}\n\n/** Parse transcript jsonl text into conversation lines. Non-conversation lines are dropped. */\nexport function parseClaudeEntries(raw: string): ClaudeEntry[] {\n  const out: ClaudeEntry[] = []\n  for (const line of raw.split('\\n')) {\n    if (!line.trim()) continue\n    let entry: Record<string, unknown>\n    try {\n      const parsed: unknown = JSON.parse(line)\n      if (!isRecord(parsed)) continue\n      entry = parsed\n    } catch {\n      continue\n    }\n    if (entry.type !== 'user' && entry.type !== 'assistant') continue\n    const message = entry.message\n    if (!isRecord(message)) continue\n    out.push({\n      type: entry.type,\n      timestamp: typeof entry.timestamp === 'string' ? entry.timestamp : null,\n      message,\n      toolUseResult: entry.toolUseResult,\n      isSidechain: entry.isSidechain === true,\n      agentId: typeof entry.agentId === 'string' ? entry.agentId : null,\n    })\n  }\n  return out\n}\n\nexport interface ReadClaudeTranscriptOptions {\n  /**\n   * Read the sidechain (subagent) thread instead of skipping it. Subagent\n   * transcripts under `<session>/subagents/agent-<id>.jsonl` are sidechain\n   * lines end to end, so their usage is invisible without this.\n   */\n  readonly includeSidechain?: boolean\n}\n\nfunction blockText(content: unknown): string {\n  if (typeof content === 'string') return content\n  if (!Array.isArray(content)) return ''\n  return content\n    .filter(\n      (b): b is Record<string, unknown> =>\n        isRecord(b) && b.type === 'text' && typeof b.text === 'string',\n    )\n    .map((b) => b.text as string)\n    .join('\\n')\n}\n\n/** Parse one transcript jsonl into canonical messages + usage totals. */\nexport async function readClaudeTranscript(\n  path: string,\n  options: ReadClaudeTranscriptOptions = {},\n): Promise<ClaudeTranscript> {\n  return transcriptFromEntries(parseClaudeEntries(await readFile(path, 'utf8')), options)\n}\n\n/** The messages+usage projection of already-parsed entries. */\nexport function transcriptFromEntries(\n  entries: readonly ClaudeEntry[],\n  options: ReadClaudeTranscriptOptions = {},\n): ClaudeTranscript {\n  const wantSidechain = options.includeSidechain === true\n  const messages: ChatMessage[] = []\n  const usage: ClaudeUsageTotals = { tokensIn: 0, tokensOut: 0, cacheRead: 0, cacheWrite: 0 }\n  let startedAt: string | null = null\n  let endedAt: string | null = null\n  let model: string | null = null\n  // Claude Code writes one jsonl line PER CONTENT BLOCK of an API message,\n  // repeating message.id and usage on each — merge blocks into one canonical\n  // assistant turn and count usage once per API message id.\n  let lastAssistantApiId: string | null = null\n  let lastAssistantIndex = -1\n\n  for (const entry of entries) {\n    if (entry.isSidechain !== wantSidechain) continue\n    const message = entry.message\n    if (entry.timestamp !== null) {\n      if (startedAt === null) startedAt = entry.timestamp\n      endedAt = entry.timestamp\n    }\n\n    if (entry.type === 'user') {\n      lastAssistantApiId = null\n      lastAssistantIndex = -1\n      const content = message.content\n      if (typeof content === 'string') {\n        messages.push({ role: 'user', content })\n        continue\n      }\n      if (!Array.isArray(content)) continue\n      // A user line may interleave tool_result blocks (answers to the prior\n      // assistant tool_use) with plain text; preserve order.\n      let userText = ''\n      for (const block of content) {\n        if (!isRecord(block)) continue\n        if (block.type === 'tool_result' && typeof block.tool_use_id === 'string') {\n          messages.push({\n            role: 'tool',\n            tool_call_id: block.tool_use_id,\n            content:\n              blockText(block.content) || (typeof block.content === 'string' ? block.content : ''),\n          })\n        } else if (block.type === 'text' && typeof block.text === 'string') {\n          userText += (userText.length > 0 ? '\\n' : '') + block.text\n        }\n      }\n      if (userText.length > 0) messages.push({ role: 'user', content: userText })\n      continue\n    }\n\n    // assistant\n    if (typeof message.model === 'string') model = message.model\n    const apiId = typeof message.id === 'string' ? message.id : null\n    const continuesTurn = apiId !== null && apiId === lastAssistantApiId && lastAssistantIndex >= 0\n    const msgUsage = message.usage\n    if (isRecord(msgUsage) && !continuesTurn) {\n      usage.tokensIn += typeof msgUsage.input_tokens === 'number' ? msgUsage.input_tokens : 0\n      usage.tokensOut += typeof msgUsage.output_tokens === 'number' ? msgUsage.output_tokens : 0\n      usage.cacheRead +=\n        typeof msgUsage.cache_read_input_tokens === 'number' ? msgUsage.cache_read_input_tokens : 0\n      usage.cacheWrite +=\n        typeof msgUsage.cache_creation_input_tokens === 'number'\n          ? msgUsage.cache_creation_input_tokens\n          : 0\n    }\n    const content = message.content\n    if (!Array.isArray(content)) continue\n    let reasoning = ''\n    let text = ''\n    const toolCalls: ChatToolCall[] = []\n    for (const block of content) {\n      if (!isRecord(block)) continue\n      if (\n        block.type === 'thinking' &&\n        typeof block.thinking === 'string' &&\n        block.thinking.length > 0\n      ) {\n        reasoning += (reasoning.length > 0 ? '\\n' : '') + block.thinking\n      } else if (block.type === 'text' && typeof block.text === 'string') {\n        text += (text.length > 0 ? '\\n' : '') + block.text\n      } else if (block.type === 'tool_use' && typeof block.id === 'string') {\n        toolCalls.push({\n          id: block.id,\n          type: 'function',\n          function: {\n            name: typeof block.name === 'string' ? block.name : 'unknown',\n            arguments: JSON.stringify(block.input ?? {}),\n          },\n        })\n      }\n    }\n    if (reasoning.length === 0 && text.length === 0 && toolCalls.length === 0) continue\n    if (continuesTurn) {\n      const prev = messages[lastAssistantIndex]!\n      if (text.length > 0) prev.content = prev.content === null ? text : `${prev.content}\\n${text}`\n      if (reasoning.length > 0) {\n        prev.reasoning_content =\n          prev.reasoning_content === undefined\n            ? reasoning\n            : `${prev.reasoning_content}\\n${reasoning}`\n      }\n      if (toolCalls.length > 0) prev.tool_calls = [...(prev.tool_calls ?? []), ...toolCalls]\n      continue\n    }\n    messages.push({\n      role: 'assistant',\n      content: text.length > 0 ? text : null,\n      ...(reasoning.length > 0 ? { reasoning_content: reasoning } : {}),\n      ...(toolCalls.length > 0 ? { tool_calls: toolCalls } : {}),\n    })\n    lastAssistantApiId = apiId\n    lastAssistantIndex = messages.length - 1\n  }\n\n  return { messages, usage, startedAt, endedAt, model }\n}\n","/**\n * Read-only backfill reader over the opencode sqlite store\n * (~/.local/share/opencode/opencode.db) → canonical chat-with-tools messages.\n *\n * Schema consumed (observed, 2026-07): `session` rows carry directory /\n * parent_id / agent / model / cost / tokens_*; `message` rows carry a JSON\n * `data` blob ({role, modelID, providerID, tokens, cost, finish}); `part`\n * rows carry the actual content ({type: text|reasoning|tool|step-start|\n * step-finish|snapshot…}). Tool parts hold {callID, state:{input, output,\n * status}} — both the call and its result, which we split into an assistant\n * tool_call plus a role:\"tool\" result message.\n *\n * The store is mutable and can be corrupt (a `.corrupt-bak` sibling ships\n * next to it in the wild), so `openOpencodeDb` returns null instead of\n * throwing — callers record a gap line, never crash the backfill.\n */\n\nimport { createRequire } from 'node:module'\nimport { homedir } from 'node:os'\nimport { join } from 'node:path'\nimport type { DatabaseSync } from 'node:sqlite'\nimport type { ChatMessage, ChatToolCall } from '../schema'\n\nexport const DEFAULT_OPENCODE_DB = join(homedir(), '.local', 'share', 'opencode', 'opencode.db')\n\nexport interface OpencodeSessionRow {\n  id: string\n  parentId: string | null\n  directory: string\n  agent: string | null\n  /** Raw session.model JSON: {id, providerID, variant} where present. */\n  model: { id?: string; providerID?: string } | null\n  costUsd: number\n  tokensInput: number\n  tokensOutput: number\n  tokensReasoning: number\n  tokensCacheRead: number\n  tokensCacheWrite: number\n  timeCreated: number\n  timeUpdated: number\n}\n\n// `node:sqlite` is loaded through CommonJS `require`, not `import()`. esbuild\n// (bundling) and Vite (tests) both rewrite a dynamic import and strip the\n// `node:` prefix under an es20xx target, turning this builtin into a bogus\n// \"sqlite\" package lookup; composing the specifier at runtime does not reliably\n// defeat that (it still resolved through Vite's transform in some workers, so\n// the failure moved around as test files were added). A require obtained from\n// `createRequire` is not an analyzable module reference in either tool, so\n// neither can rewrite it.\nconst nodeRequire = createRequire(import.meta.url)\n\n/** Open the store read-only; null = unavailable/corrupt (caller records a gap). */\nexport async function openOpencodeDb(\n  path: string = DEFAULT_OPENCODE_DB,\n): Promise<DatabaseSync | null> {\n  try {\n    const { DatabaseSync } = nodeRequire('node:sqlite') as typeof import('node:sqlite')\n    const db = new DatabaseSync(path, { readOnly: true })\n    // Probe: a corrupt store can open() fine and fail on first page read.\n    db.prepare('SELECT id FROM session LIMIT 1').get()\n    return db\n  } catch {\n    return null\n  }\n}\n\nconst isRecord = (v: unknown): v is Record<string, unknown> =>\n  typeof v === 'object' && v !== null && !Array.isArray(v)\n\nfunction parseSessionRow(row: Record<string, unknown>): OpencodeSessionRow {\n  let model: OpencodeSessionRow['model'] = null\n  if (typeof row.model === 'string' && row.model.length > 0) {\n    try {\n      const parsed: unknown = JSON.parse(row.model)\n      if (isRecord(parsed)) model = parsed as { id?: string; providerID?: string }\n    } catch {\n      model = null\n    }\n  }\n  return {\n    id: String(row.id),\n    parentId: row.parent_id === null || row.parent_id === undefined ? null : String(row.parent_id),\n    directory: String(row.directory),\n    agent: row.agent === null || row.agent === undefined ? null : String(row.agent),\n    model,\n    costUsd: Number(row.cost ?? 0),\n    tokensInput: Number(row.tokens_input ?? 0),\n    tokensOutput: Number(row.tokens_output ?? 0),\n    tokensReasoning: Number(row.tokens_reasoning ?? 0),\n    tokensCacheRead: Number(row.tokens_cache_read ?? 0),\n    tokensCacheWrite: Number(row.tokens_cache_write ?? 0),\n    timeCreated: Number(row.time_created ?? 0),\n    timeUpdated: Number(row.time_updated ?? 0),\n  }\n}\n\nconst SESSION_COLUMNS =\n  'id, parent_id, directory, agent, model, cost, tokens_input, tokens_output, tokens_reasoning, tokens_cache_read, tokens_cache_write, time_created, time_updated'\n\n/** Sessions whose cwd is `directory` (the worker-clone join key). */\nexport function findOpencodeSessionsByDirectory(\n  db: DatabaseSync,\n  directory: string,\n): OpencodeSessionRow[] {\n  const rows = db\n    .prepare(`SELECT ${SESSION_COLUMNS} FROM session WHERE directory = ? ORDER BY time_created`)\n    .all(directory) as Array<Record<string, unknown>>\n  return rows.map(parseSessionRow)\n}\n\nexport function findOpencodeSessionById(\n  db: DatabaseSync,\n  sessionId: string,\n): OpencodeSessionRow | null {\n  const row = db.prepare(`SELECT ${SESSION_COLUMNS} FROM session WHERE id = ?`).get(sessionId) as\n    | Record<string, unknown>\n    | undefined\n  return row === undefined ? null : parseSessionRow(row)\n}\n\ninterface OpencodePart {\n  type?: string\n  text?: string\n  tool?: string\n  callID?: string\n  state?: { status?: string; input?: unknown; output?: unknown }\n}\n\nfunction toolResultContent(output: unknown): string {\n  if (typeof output === 'string') return output\n  if (output === null || output === undefined) return ''\n  return JSON.stringify(output)\n}\n\n/**\n * Convert one session's message+part rows into canonical messages.\n * An opencode assistant message row spans several model steps; each step's\n * parts (reasoning → text → tool …) become one assistant message followed by\n * the role:\"tool\" results of its calls, preserving order.\n */\nexport function readOpencodeSessionMessages(db: DatabaseSync, sessionId: string): ChatMessage[] {\n  const messageRows = db\n    .prepare('SELECT id, data FROM message WHERE session_id = ? ORDER BY time_created, id')\n    .all(sessionId) as Array<{ id: string; data: string }>\n  const partsStmt = db.prepare('SELECT data FROM part WHERE message_id = ? ORDER BY id')\n\n  const messages: ChatMessage[] = []\n  for (const messageRow of messageRows) {\n    let data: Record<string, unknown>\n    try {\n      const parsed: unknown = JSON.parse(messageRow.data)\n      if (!isRecord(parsed)) continue\n      data = parsed\n    } catch {\n      continue\n    }\n    const parts: OpencodePart[] = []\n    for (const row of partsStmt.all(messageRow.id) as Array<{ data: string }>) {\n      try {\n        const parsed: unknown = JSON.parse(row.data)\n        if (isRecord(parsed)) parts.push(parsed as OpencodePart)\n      } catch {\n        // Malformed part payload: skip the part, keep the message.\n      }\n    }\n\n    if (data.role === 'user') {\n      const text = parts\n        .filter((p) => p.type === 'text' && typeof p.text === 'string')\n        .map((p) => p.text as string)\n        .join('\\n')\n      messages.push({ role: 'user', content: text })\n      continue\n    }\n    if (data.role !== 'assistant') continue\n\n    // Split the row into steps at step-start boundaries; parts before the\n    // first step-start (none observed, but tolerated) form an implicit step.\n    const steps: OpencodePart[][] = []\n    let current: OpencodePart[] = []\n    for (const part of parts) {\n      if (part.type === 'step-start') {\n        if (current.length > 0) steps.push(current)\n        current = []\n        continue\n      }\n      if (part.type === 'step-finish' || part.type === 'snapshot' || part.type === 'patch') continue\n      current.push(part)\n    }\n    if (current.length > 0) steps.push(current)\n\n    for (const step of steps) {\n      const reasoning = step\n        .filter((p) => p.type === 'reasoning' && typeof p.text === 'string' && p.text.length > 0)\n        .map((p) => p.text as string)\n        .join('\\n')\n      const text = step\n        .filter((p) => p.type === 'text' && typeof p.text === 'string')\n        .map((p) => p.text as string)\n        .join('\\n')\n      const toolParts = step.filter((p) => p.type === 'tool' && typeof p.callID === 'string')\n      const toolCalls: ChatToolCall[] = toolParts.map((p) => ({\n        id: p.callID as string,\n        type: 'function',\n        function: {\n          name: p.tool ?? 'unknown',\n          arguments: JSON.stringify(p.state?.input ?? {}),\n        },\n      }))\n      if (reasoning.length === 0 && text.length === 0 && toolCalls.length === 0) continue\n      messages.push({\n        role: 'assistant',\n        content: text.length > 0 ? text : null,\n        ...(reasoning.length > 0 ? { reasoning_content: reasoning } : {}),\n        ...(toolCalls.length > 0 ? { tool_calls: toolCalls } : {}),\n      })\n      for (const p of toolParts) {\n        messages.push({\n          role: 'tool',\n          tool_call_id: p.callID as string,\n          name: p.tool ?? 'unknown',\n          content: toolResultContent(p.state?.output),\n        })\n      }\n    }\n  }\n  return messages\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAkBA,MAAa,8BAA8B,KAAK,QAAQ,GAAG,WAAW,UAAU;;AAGhF,SAAgB,kBAAkB,KAAqB;CACrD,OAAO,IAAI,QAAQ,kBAAkB,GAAG;AAC1C;;AAQA,eAAsB,sBACpB,KACA,cAAsB,6BACU;CAChC,MAAM,MAAM,KAAK,aAAa,kBAAkB,GAAG,CAAC;CAEpD,QAAO,MADa,QAAQ,GAAG,CAAC,CAAC,YAAY,CAAC,CAAC,EAAA,CAE5C,QAAQ,MAAM,EAAE,SAAS,QAAQ,CAAC,CAAC,CACnC,KAAK,CAAC,CACN,KAAK,OAAO;EAAE,WAAW,EAAE,QAAQ,YAAY,EAAE;EAAG,MAAM,KAAK,KAAK,CAAC;CAAE,EAAE;AAC9E;AAkBA,MAAMA,cAAY,MAChB,OAAO,MAAM,YAAY,MAAM,QAAQ,CAAC,MAAM,QAAQ,CAAC;;AA0BzD,SAAgB,mBAAmB,KAA4B;CAC7D,MAAM,MAAqB,CAAC;CAC5B,KAAK,MAAM,QAAQ,IAAI,MAAM,IAAI,GAAG;EAClC,IAAI,CAAC,KAAK,KAAK,GAAG;EAClB,IAAI;EACJ,IAAI;GACF,MAAM,SAAkB,KAAK,MAAM,IAAI;GACvC,IAAI,CAACA,WAAS,MAAM,GAAG;GACvB,QAAQ;EACV,QAAQ;GACN;EACF;EACA,IAAI,MAAM,SAAS,UAAU,MAAM,SAAS,aAAa;EACzD,MAAM,UAAU,MAAM;EACtB,IAAI,CAACA,WAAS,OAAO,GAAG;EACxB,IAAI,KAAK;GACP,MAAM,MAAM;GACZ,WAAW,OAAO,MAAM,cAAc,WAAW,MAAM,YAAY;GACnE;GACA,eAAe,MAAM;GACrB,aAAa,MAAM,gBAAgB;GACnC,SAAS,OAAO,MAAM,YAAY,WAAW,MAAM,UAAU;EAC/D,CAAC;CACH;CACA,OAAO;AACT;AAWA,SAAS,UAAU,SAA0B;CAC3C,IAAI,OAAO,YAAY,UAAU,OAAO;CACxC,IAAI,CAAC,MAAM,QAAQ,OAAO,GAAG,OAAO;CACpC,OAAO,QACJ,QACE,MACCA,WAAS,CAAC,KAAK,EAAE,SAAS,UAAU,OAAO,EAAE,SAAS,QAC1D,CAAC,CACA,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;AACd;;AAGA,eAAsB,qBACpB,MACA,UAAuC,CAAC,GACb;CAC3B,OAAO,sBAAsB,mBAAmB,MAAM,SAAS,MAAM,MAAM,CAAC,GAAG,OAAO;AACxF;;AAGA,SAAgB,sBACd,SACA,UAAuC,CAAC,GACtB;CAClB,MAAM,gBAAgB,QAAQ,qBAAqB;CACnD,MAAM,WAA0B,CAAC;CACjC,MAAM,QAA2B;EAAE,UAAU;EAAG,WAAW;EAAG,WAAW;EAAG,YAAY;CAAE;CAC1F,IAAI,YAA2B;CAC/B,IAAI,UAAyB;CAC7B,IAAI,QAAuB;CAI3B,IAAI,qBAAoC;CACxC,IAAI,qBAAqB;CAEzB,KAAK,MAAM,SAAS,SAAS;EAC3B,IAAI,MAAM,gBAAgB,eAAe;EACzC,MAAM,UAAU,MAAM;EACtB,IAAI,MAAM,cAAc,MAAM;GAC5B,IAAI,cAAc,MAAM,YAAY,MAAM;GAC1C,UAAU,MAAM;EAClB;EAEA,IAAI,MAAM,SAAS,QAAQ;GACzB,qBAAqB;GACrB,qBAAqB;GACrB,MAAM,UAAU,QAAQ;GACxB,IAAI,OAAO,YAAY,UAAU;IAC/B,SAAS,KAAK;KAAE,MAAM;KAAQ;IAAQ,CAAC;IACvC;GACF;GACA,IAAI,CAAC,MAAM,QAAQ,OAAO,GAAG;GAG7B,IAAI,WAAW;GACf,KAAK,MAAM,SAAS,SAAS;IAC3B,IAAI,CAACA,WAAS,KAAK,GAAG;IACtB,IAAI,MAAM,SAAS,iBAAiB,OAAO,MAAM,gBAAgB,UAC/D,SAAS,KAAK;KACZ,MAAM;KACN,cAAc,MAAM;KACpB,SACE,UAAU,MAAM,OAAO,MAAM,OAAO,MAAM,YAAY,WAAW,MAAM,UAAU;IACrF,CAAC;SACI,IAAI,MAAM,SAAS,UAAU,OAAO,MAAM,SAAS,UACxD,aAAa,SAAS,SAAS,IAAI,OAAO,MAAM,MAAM;GAE1D;GACA,IAAI,SAAS,SAAS,GAAG,SAAS,KAAK;IAAE,MAAM;IAAQ,SAAS;GAAS,CAAC;GAC1E;EACF;EAGA,IAAI,OAAO,QAAQ,UAAU,UAAU,QAAQ,QAAQ;EACvD,MAAM,QAAQ,OAAO,QAAQ,OAAO,WAAW,QAAQ,KAAK;EAC5D,MAAM,gBAAgB,UAAU,QAAQ,UAAU,sBAAsB,sBAAsB;EAC9F,MAAM,WAAW,QAAQ;EACzB,IAAIA,WAAS,QAAQ,KAAK,CAAC,eAAe;GACxC,MAAM,YAAY,OAAO,SAAS,iBAAiB,WAAW,SAAS,eAAe;GACtF,MAAM,aAAa,OAAO,SAAS,kBAAkB,WAAW,SAAS,gBAAgB;GACzF,MAAM,aACJ,OAAO,SAAS,4BAA4B,WAAW,SAAS,0BAA0B;GAC5F,MAAM,cACJ,OAAO,SAAS,gCAAgC,WAC5C,SAAS,8BACT;EACR;EACA,MAAM,UAAU,QAAQ;EACxB,IAAI,CAAC,MAAM,QAAQ,OAAO,GAAG;EAC7B,IAAI,YAAY;EAChB,IAAI,OAAO;EACX,MAAM,YAA4B,CAAC;EACnC,KAAK,MAAM,SAAS,SAAS;GAC3B,IAAI,CAACA,WAAS,KAAK,GAAG;GACtB,IACE,MAAM,SAAS,cACf,OAAO,MAAM,aAAa,YAC1B,MAAM,SAAS,SAAS,GAExB,cAAc,UAAU,SAAS,IAAI,OAAO,MAAM,MAAM;QACnD,IAAI,MAAM,SAAS,UAAU,OAAO,MAAM,SAAS,UACxD,SAAS,KAAK,SAAS,IAAI,OAAO,MAAM,MAAM;QACzC,IAAI,MAAM,SAAS,cAAc,OAAO,MAAM,OAAO,UAC1D,UAAU,KAAK;IACb,IAAI,MAAM;IACV,MAAM;IACN,UAAU;KACR,MAAM,OAAO,MAAM,SAAS,WAAW,MAAM,OAAO;KACpD,WAAW,KAAK,UAAU,MAAM,SAAS,CAAC,CAAC;IAC7C;GACF,CAAC;EAEL;EACA,IAAI,UAAU,WAAW,KAAK,KAAK,WAAW,KAAK,UAAU,WAAW,GAAG;EAC3E,IAAI,eAAe;GACjB,MAAM,OAAO,SAAS;GACtB,IAAI,KAAK,SAAS,GAAG,KAAK,UAAU,KAAK,YAAY,OAAO,OAAO,GAAG,KAAK,QAAQ,IAAI;GACvF,IAAI,UAAU,SAAS,GACrB,KAAK,oBACH,KAAK,sBAAsB,KAAA,IACvB,YACA,GAAG,KAAK,kBAAkB,IAAI;GAEtC,IAAI,UAAU,SAAS,GAAG,KAAK,aAAa,CAAC,GAAI,KAAK,cAAc,CAAC,GAAI,GAAG,SAAS;GACrF;EACF;EACA,SAAS,KAAK;GACZ,MAAM;GACN,SAAS,KAAK,SAAS,IAAI,OAAO;GAClC,GAAI,UAAU,SAAS,IAAI,EAAE,mBAAmB,UAAU,IAAI,CAAC;GAC/D,GAAI,UAAU,SAAS,IAAI,EAAE,YAAY,UAAU,IAAI,CAAC;EAC1D,CAAC;EACD,qBAAqB;EACrB,qBAAqB,SAAS,SAAS;CACzC;CAEA,OAAO;EAAE;EAAU;EAAO;EAAW;EAAS;CAAM;AACtD;;;;;;;;;;;;;;;;;;;AC9OA,MAAa,sBAAsB,KAAK,QAAQ,GAAG,UAAU,SAAS,YAAY,aAAa;AA2B/F,MAAM,cAAc,cAAc,OAAO,KAAK,GAAG;;AAGjD,eAAsB,eACpB,OAAe,qBACe;CAC9B,IAAI;EACF,MAAM,EAAE,iBAAiB,YAAY,aAAa;EAClD,MAAM,KAAK,IAAI,aAAa,MAAM,EAAE,UAAU,KAAK,CAAC;EAEpD,GAAG,QAAQ,gCAAgC,CAAC,CAAC,IAAI;EACjD,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;AAEA,MAAM,YAAY,MAChB,OAAO,MAAM,YAAY,MAAM,QAAQ,CAAC,MAAM,QAAQ,CAAC;AAEzD,SAAS,gBAAgB,KAAkD;CACzE,IAAI,QAAqC;CACzC,IAAI,OAAO,IAAI,UAAU,YAAY,IAAI,MAAM,SAAS,GACtD,IAAI;EACF,MAAM,SAAkB,KAAK,MAAM,IAAI,KAAK;EAC5C,IAAI,SAAS,MAAM,GAAG,QAAQ;CAChC,QAAQ;EACN,QAAQ;CACV;CAEF,OAAO;EACL,IAAI,OAAO,IAAI,EAAE;EACjB,UAAU,IAAI,cAAc,QAAQ,IAAI,cAAc,KAAA,IAAY,OAAO,OAAO,IAAI,SAAS;EAC7F,WAAW,OAAO,IAAI,SAAS;EAC/B,OAAO,IAAI,UAAU,QAAQ,IAAI,UAAU,KAAA,IAAY,OAAO,OAAO,IAAI,KAAK;EAC9E;EACA,SAAS,OAAO,IAAI,QAAQ,CAAC;EAC7B,aAAa,OAAO,IAAI,gBAAgB,CAAC;EACzC,cAAc,OAAO,IAAI,iBAAiB,CAAC;EAC3C,iBAAiB,OAAO,IAAI,oBAAoB,CAAC;EACjD,iBAAiB,OAAO,IAAI,qBAAqB,CAAC;EAClD,kBAAkB,OAAO,IAAI,sBAAsB,CAAC;EACpD,aAAa,OAAO,IAAI,gBAAgB,CAAC;EACzC,aAAa,OAAO,IAAI,gBAAgB,CAAC;CAC3C;AACF;AAEA,MAAM,kBACJ;;AAGF,SAAgB,gCACd,IACA,WACsB;CAItB,OAHa,GACV,QAAQ,UAAU,gBAAgB,wDAAwD,CAAC,CAC3F,IAAI,SACG,CAAC,CAAC,IAAI,eAAe;AACjC;AAEA,SAAgB,wBACd,IACA,WAC2B;CAC3B,MAAM,MAAM,GAAG,QAAQ,UAAU,gBAAgB,2BAA2B,CAAC,CAAC,IAAI,SAAS;CAG3F,OAAO,QAAQ,KAAA,IAAY,OAAO,gBAAgB,GAAG;AACvD;AAUA,SAAS,kBAAkB,QAAyB;CAClD,IAAI,OAAO,WAAW,UAAU,OAAO;CACvC,IAAI,WAAW,QAAQ,WAAW,KAAA,GAAW,OAAO;CACpD,OAAO,KAAK,UAAU,MAAM;AAC9B;;;;;;;AAQA,SAAgB,4BAA4B,IAAkB,WAAkC;CAC9F,MAAM,cAAc,GACjB,QAAQ,6EAA6E,CAAC,CACtF,IAAI,SAAS;CAChB,MAAM,YAAY,GAAG,QAAQ,wDAAwD;CAErF,MAAM,WAA0B,CAAC;CACjC,KAAK,MAAM,cAAc,aAAa;EACpC,IAAI;EACJ,IAAI;GACF,MAAM,SAAkB,KAAK,MAAM,WAAW,IAAI;GAClD,IAAI,CAAC,SAAS,MAAM,GAAG;GACvB,OAAO;EACT,QAAQ;GACN;EACF;EACA,MAAM,QAAwB,CAAC;EAC/B,KAAK,MAAM,OAAO,UAAU,IAAI,WAAW,EAAE,GAC3C,IAAI;GACF,MAAM,SAAkB,KAAK,MAAM,IAAI,IAAI;GAC3C,IAAI,SAAS,MAAM,GAAG,MAAM,KAAK,MAAsB;EACzD,QAAQ,CAER;EAGF,IAAI,KAAK,SAAS,QAAQ;GACxB,MAAM,OAAO,MACV,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,SAAS,QAAQ,CAAC,CAC9D,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,SAAS,KAAK;IAAE,MAAM;IAAQ,SAAS;GAAK,CAAC;GAC7C;EACF;EACA,IAAI,KAAK,SAAS,aAAa;EAI/B,MAAM,QAA0B,CAAC;EACjC,IAAI,UAA0B,CAAC;EAC/B,KAAK,MAAM,QAAQ,OAAO;GACxB,IAAI,KAAK,SAAS,cAAc;IAC9B,IAAI,QAAQ,SAAS,GAAG,MAAM,KAAK,OAAO;IAC1C,UAAU,CAAC;IACX;GACF;GACA,IAAI,KAAK,SAAS,iBAAiB,KAAK,SAAS,cAAc,KAAK,SAAS,SAAS;GACtF,QAAQ,KAAK,IAAI;EACnB;EACA,IAAI,QAAQ,SAAS,GAAG,MAAM,KAAK,OAAO;EAE1C,KAAK,MAAM,QAAQ,OAAO;GACxB,MAAM,YAAY,KACf,QAAQ,MAAM,EAAE,SAAS,eAAe,OAAO,EAAE,SAAS,YAAY,EAAE,KAAK,SAAS,CAAC,CAAC,CACxF,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,MAAM,OAAO,KACV,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,SAAS,QAAQ,CAAC,CAC9D,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,MAAM,YAAY,KAAK,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,WAAW,QAAQ;GACtF,MAAM,YAA4B,UAAU,KAAK,OAAO;IACtD,IAAI,EAAE;IACN,MAAM;IACN,UAAU;KACR,MAAM,EAAE,QAAQ;KAChB,WAAW,KAAK,UAAU,EAAE,OAAO,SAAS,CAAC,CAAC;IAChD;GACF,EAAE;GACF,IAAI,UAAU,WAAW,KAAK,KAAK,WAAW,KAAK,UAAU,WAAW,GAAG;GAC3E,SAAS,KAAK;IACZ,MAAM;IACN,SAAS,KAAK,SAAS,IAAI,OAAO;IAClC,GAAI,UAAU,SAAS,IAAI,EAAE,mBAAmB,UAAU,IAAI,CAAC;IAC/D,GAAI,UAAU,SAAS,IAAI,EAAE,YAAY,UAAU,IAAI,CAAC;GAC1D,CAAC;GACD,KAAK,MAAM,KAAK,WACd,SAAS,KAAK;IACZ,MAAM;IACN,cAAc,EAAE;IAChB,MAAM,EAAE,QAAQ;IAChB,SAAS,kBAAkB,EAAE,OAAO,MAAM;GAC5C,CAAC;EAEL;CACF;CACA,OAAO;AACT"}

package/dist/outcome-store-BYHIuO0e.d.ts

@@ -0,0 +1,65 @@
+//#region src/meta-eval/outcome-store.d.ts
+/**
+ * OutcomeStore — deployment outcomes attached to Run IDs.
+ *
+ * Outcomes arrive asynchronously from production telemetry after the
+ * eval run completed: user ratings, retention flags, conversion events,
+ * revenue, support-ticket rate, anything a product team can measure.
+ * The store is a peer to TraceStore — separate lifecycle, same runId
+ * foreign key.
+ *
+ * The whole point of this module is to make the meta-eval correlation
+ * question computable: `correlate(evalMetric, outcomeMetric) → r, ρ, n, CI`.
+ */
+interface DeploymentOutcome {
+  runId: string;
+  capturedAt: number;
+  /** Numeric outcomes keyed by name — retention_7d, csat, revenue_usd, etc. */
+  metrics: Record<string, number>;
+  /** Dimensions for stratified analysis — cohort, region, user_segment. */
+  labels?: Record<string, string>;
+  /** Free-form provenance (source system, pipeline version). */
+  source?: string;
+}
+interface OutcomeFilter {
+  runIds?: string[];
+  since?: number;
+  until?: number;
+  label?: {
+    key: string;
+    value: string;
+  };
+  source?: string;
+}
+interface OutcomeStore {
+  append(outcome: DeploymentOutcome): Promise<void>;
+  /** All outcomes attached to this run (a single run can have many — multiple
+   *  capture windows over deployment time). */
+  forRun(runId: string): Promise<DeploymentOutcome[]>;
+  list(filter?: OutcomeFilter): Promise<DeploymentOutcome[]>;
+}
+declare class InMemoryOutcomeStore implements OutcomeStore {
+  private items;
+  append(outcome: DeploymentOutcome): Promise<void>;
+  forRun(runId: string): Promise<DeploymentOutcome[]>;
+  list(filter?: OutcomeFilter): Promise<DeploymentOutcome[]>;
+}
+interface FileSystemOutcomeStoreOptions {
+  dir: string;
+  maxBytes?: number;
+}
+declare class FileSystemOutcomeStore implements OutcomeStore {
+  private dir;
+  private maxBytes;
+  private memo?;
+  private loaded;
+  constructor(options: FileSystemOutcomeStoreOptions);
+  private ensureDir;
+  append(outcome: DeploymentOutcome): Promise<void>;
+  private load;
+  forRun(runId: string): Promise<DeploymentOutcome[]>;
+  list(filter?: OutcomeFilter): Promise<DeploymentOutcome[]>;
+}
+//#endregion
+export { OutcomeFilter as a, InMemoryOutcomeStore as i, FileSystemOutcomeStore as n, OutcomeStore as o, FileSystemOutcomeStoreOptions as r, DeploymentOutcome as t };
+//# sourceMappingURL=outcome-store-BYHIuO0e.d.ts.map

package/dist/outcome-store-BYHIuO0e.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"outcome-store-BYHIuO0e.d.ts","names":[],"sources":["../src/meta-eval/outcome-store.ts"],"mappings":";;;;;;;;;;;;;UAaiB;EACf;EACA;;EAEA,SAAS;;EAET,SAAS;;EAET;;UAGe;EACf;EACA;EACA;EACA;IAAU;IAAa;;EACvB;;UAGe;EACf,OAAO,SAAS,oBAAoB;;;EAGpC,OAAO,gBAAgB,QAAQ;EAC/B,KAAK,SAAS,gBAAgB,QAAQ;;cAG3B,gCAAgC;UACnC;EAEF,OAAO,SAAS,oBAAoB;EAIpC,OAAO,gBAAgB,QAAQ;EAI/B,KAAK,SAAQ,gBAAqB,QAAQ;;UAKjC;EACf;EACA;;cAGW,kCAAkC;UACrC;UACA;UACA;UACA;EAER,YAAY,SAAS;UAKP;EAKR,OAAO,SAAS,oBAAoB;UAiB5B;EAuBR,OAAO,gBAAgB,QAAQ;EAI/B,KAAK,SAAS,gBAAgB,QAAQ"}