@tangle-network/agent-eval 0.209.1 → 0.210.0

package/dist/claude-jsonl-dZir12m6.js

@@ -0,0 +1,86 @@
+import { join } from "node:path";
+import { homedir } from "node:os";
+import { readdir } from "node:fs/promises";
+import { claudeCodeReader, claudeCodeRefForFile, readSessionInput, toChatMessages } from "@tangle-network/harness-sessions";
+//#region src/rollout/readers/claude-jsonl.ts
+/**
+* Claude Code project transcripts (~/.claude/projects/<cwd-slug>/<sessionId>.jsonl) as canonical
+* chat-with-tools messages plus per-session token usage.
+*
+* The transcript is read by @tangle-network/harness-sessions, the one reader of harness sessions
+* shared with Discovery, agent-record, traces and the blog: one model call per API response
+* (`message.id`, its usage counted once), the served model from the response, tool calls with
+* their results. This module only projects that session. Sidechain lines (subagent threads in the
+* main file) are separate invocations and are excluded unless `includeSidechain` asks for a
+* subagent's own transcript.
+*/
+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 USAGE_FIELDS = [
+	["tokensIn", "input"],
+	["tokensOut", "output"],
+	["cacheRead", "cacheRead"],
+	["cacheWrite", "cacheWrite"]
+];
+/**
+* The messages+usage projection of a normalized session. A usage field is null when any answered
+* model call did not report it, so a partial sum is never mistaken for the session's usage.
+*/
+function transcriptFromSession(session) {
+	const gaps = session.integrity.gaps.filter((gap) => /is not a JSON record|still being written/u.test(gap));
+	const answered = session.modelCalls.filter((call) => call.error === null);
+	const usage = {
+		tokensIn: null,
+		tokensOut: null,
+		cacheRead: null,
+		cacheWrite: null
+	};
+	if (answered.length > 0) for (const [target, field] of USAGE_FIELDS) {
+		let total = 0;
+		for (const call of answered) {
+			const value = call.usage?.[field] ?? null;
+			if (value === null) {
+				total = null;
+				gaps.push(`model call ${call.id}: ${field} usage unavailable`);
+				break;
+			}
+			total += value;
+		}
+		usage[target] = total;
+	}
+	const model = [...answered].reverse().find((call) => call.servedModel !== null)?.servedModel ?? null;
+	return {
+		messages: toChatMessages(session),
+		usage,
+		startedAt: session.startedAt,
+		endedAt: session.endedAt,
+		model,
+		gaps
+	};
+}
+/** Read one transcript file. */
+async function readClaudeTranscript(path, options = {}) {
+	return transcriptFromSession(await claudeCodeReader.read(claudeCodeRefForFile(path, options.includeSidechain === true ? "parent" : null)));
+}
+/** Parse retained JSONL text without another file or store; source gaps remain explicit. */
+function parseClaudeTranscript(raw, options = {}) {
+	return transcriptFromSession(readSessionInput("claude-code", { text: raw }, {
+		parentNativeSessionId: options.includeSidechain === true ? "parent" : null,
+		label: "retained transcript"
+	}));
+}
+//#endregion
+export { readClaudeTranscript as a, parseClaudeTranscript as i, claudeProjectSlug as n, transcriptFromSession as o, findClaudeTranscripts as r, DEFAULT_CLAUDE_PROJECTS_DIR as t };
+
+//# sourceMappingURL=claude-jsonl-dZir12m6.js.map

package/dist/claude-jsonl-dZir12m6.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"claude-jsonl-dZir12m6.js","names":[],"sources":["../src/rollout/readers/claude-jsonl.ts"],"sourcesContent":["/**\n * Claude Code project transcripts (~/.claude/projects/<cwd-slug>/<sessionId>.jsonl) as canonical\n * chat-with-tools messages plus per-session token usage.\n *\n * The transcript is read by @tangle-network/harness-sessions, the one reader of harness sessions\n * shared with Discovery, agent-record, traces and the blog: one model call per API response\n * (`message.id`, its usage counted once), the served model from the response, tool calls with\n * their results. This module only projects that session. Sidechain lines (subagent threads in the\n * main file) are separate invocations and are excluded unless `includeSidechain` asks for a\n * subagent's own transcript.\n */\n\nimport { readdir } from 'node:fs/promises'\nimport { homedir } from 'node:os'\nimport { join } from 'node:path'\nimport {\n  claudeCodeReader,\n  claudeCodeRefForFile,\n  type HarnessSession,\n  readSessionInput,\n  toChatMessages,\n} from '@tangle-network/harness-sessions'\nimport type { ChatMessage } 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 | null\n  tokensOut: number | null\n  cacheRead: number | null\n  cacheWrite: number | null\n}\n\nexport interface ClaudeTranscript {\n  messages: ChatMessage[]\n  usage: ClaudeUsageTotals\n  /** Timestamp of the first conversation record; null = empty transcript. */\n  startedAt: string | null\n  endedAt: string | null\n  /** The model that answered the last response (never Claude Code's `<synthetic>` error turn). */\n  model: string | null\n  /** Every parse or usage gap retained beside this semantic projection. */\n  gaps: string[]\n}\n\nexport interface ReadClaudeTranscriptOptions {\n  /**\n   * Read a subagent's own transcript (`<session>/subagents/agent-<id>.jsonl`), whose records are\n   * sidechain end to end, instead of skipping sidechain records.\n   */\n  readonly includeSidechain?: boolean\n}\n\nconst USAGE_FIELDS = [\n  ['tokensIn', 'input'],\n  ['tokensOut', 'output'],\n  ['cacheRead', 'cacheRead'],\n  ['cacheWrite', 'cacheWrite'],\n] as const\n\n/**\n * The messages+usage projection of a normalized session. A usage field is null when any answered\n * model call did not report it, so a partial sum is never mistaken for the session's usage.\n */\nexport function transcriptFromSession(session: HarnessSession): ClaudeTranscript {\n  const gaps = session.integrity.gaps.filter((gap) =>\n    /is not a JSON record|still being written/u.test(gap),\n  )\n  const answered = session.modelCalls.filter((call) => call.error === null)\n  const usage: ClaudeUsageTotals = {\n    tokensIn: null,\n    tokensOut: null,\n    cacheRead: null,\n    cacheWrite: null,\n  }\n  if (answered.length > 0) {\n    for (const [target, field] of USAGE_FIELDS) {\n      let total: number | null = 0\n      for (const call of answered) {\n        const value = call.usage?.[field] ?? null\n        if (value === null) {\n          total = null\n          gaps.push(`model call ${call.id}: ${field} usage unavailable`)\n          break\n        }\n        total += value\n      }\n      usage[target] = total\n    }\n  }\n  const model =\n    [...answered].reverse().find((call) => call.servedModel !== null)?.servedModel ?? null\n  return {\n    messages: toChatMessages(session),\n    usage,\n    startedAt: session.startedAt,\n    endedAt: session.endedAt,\n    model,\n    gaps,\n  }\n}\n\n/** Read one transcript file. */\nexport async function readClaudeTranscript(\n  path: string,\n  options: ReadClaudeTranscriptOptions = {},\n): Promise<ClaudeTranscript> {\n  const session = await claudeCodeReader.read(\n    claudeCodeRefForFile(path, options.includeSidechain === true ? 'parent' : null),\n  )\n  return transcriptFromSession(session)\n}\n\n/** Parse retained JSONL text without another file or store; source gaps remain explicit. */\nexport function parseClaudeTranscript(\n  raw: string,\n  options: ReadClaudeTranscriptOptions = {},\n): ClaudeTranscript {\n  const session = readSessionInput(\n    'claude-code',\n    { text: raw },\n    {\n      parentNativeSessionId: options.includeSidechain === true ? 'parent' : null,\n      label: 'retained transcript',\n    },\n  )\n  return transcriptFromSession(session)\n}\n"],"mappings":";;;;;;;;;;;;;;;;AAwBA,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;AA6BA,MAAM,eAAe;CACnB,CAAC,YAAY,OAAO;CACpB,CAAC,aAAa,QAAQ;CACtB,CAAC,aAAa,WAAW;CACzB,CAAC,cAAc,YAAY;AAC7B;;;;;AAMA,SAAgB,sBAAsB,SAA2C;CAC/E,MAAM,OAAO,QAAQ,UAAU,KAAK,QAAQ,QAC1C,4CAA4C,KAAK,GAAG,CACtD;CACA,MAAM,WAAW,QAAQ,WAAW,QAAQ,SAAS,KAAK,UAAU,IAAI;CACxE,MAAM,QAA2B;EAC/B,UAAU;EACV,WAAW;EACX,WAAW;EACX,YAAY;CACd;CACA,IAAI,SAAS,SAAS,GACpB,KAAK,MAAM,CAAC,QAAQ,UAAU,cAAc;EAC1C,IAAI,QAAuB;EAC3B,KAAK,MAAM,QAAQ,UAAU;GAC3B,MAAM,QAAQ,KAAK,QAAQ,UAAU;GACrC,IAAI,UAAU,MAAM;IAClB,QAAQ;IACR,KAAK,KAAK,cAAc,KAAK,GAAG,IAAI,MAAM,mBAAmB;IAC7D;GACF;GACA,SAAS;EACX;EACA,MAAM,UAAU;CAClB;CAEF,MAAM,QACJ,CAAC,GAAG,QAAQ,CAAC,CAAC,QAAQ,CAAC,CAAC,MAAM,SAAS,KAAK,gBAAgB,IAAI,CAAC,EAAE,eAAe;CACpF,OAAO;EACL,UAAU,eAAe,OAAO;EAChC;EACA,WAAW,QAAQ;EACnB,SAAS,QAAQ;EACjB;EACA;CACF;AACF;;AAGA,eAAsB,qBACpB,MACA,UAAuC,CAAC,GACb;CAI3B,OAAO,sBAAsB,MAHP,iBAAiB,KACrC,qBAAqB,MAAM,QAAQ,qBAAqB,OAAO,WAAW,IAAI,CAChF,CACoC;AACtC;;AAGA,SAAgB,sBACd,KACA,UAAuC,CAAC,GACtB;CASlB,OAAO,sBARS,iBACd,eACA,EAAE,MAAM,IAAI,GACZ;EACE,uBAAuB,QAAQ,qBAAqB,OAAO,WAAW;EACtE,OAAO;CACT,CAEiC,CAAC;AACtC"}

package/dist/cli.js

@@ -1,13 +1,13 @@
 #!/usr/bin/env node
-import { _ as INSUFFICIENT_FROM, g as operatorYield } from "./search-kernel-Bfe3kzRk.js";
+import { _ as INSUFFICIENT_FROM, g as operatorYield } from "./search-kernel-BIXWM7YV.js";
 import { i as replaySearchLedgerText } from "./search-ledger-aMTNQn_m.js";
-import { t as renderSearchSummary } from "./search-summary-DK1bwTAf.js";
+import { t as renderSearchSummary } from "./search-summary-nwrtXxC_.js";
 import { c as hostedTenantFromEnv, g as SearchRunKindSchema, r as shipSearchLedger, t as SearchShipConflictError, u as SEARCH_LEDGER_BATCH_MAX_BYTES } from "./search-shipper-DKXWuLnp.js";
 import { i as runRolloutReleaseCli } from "./hf-dataset-BR2yhuNn.js";
 import { t as createChatClient } from "./chat-client-CydfmLdO.js";
-import { t as runAnalystBenchmarkCommand } from "./benchmark-command-Cve7mHNC.js";
+import { t as runAnalystBenchmarkCommand } from "./benchmark-command-S8tCtkxh.js";
 import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-BZOt7eNC.js";
-import { C as editCredit, S as front, _ as surfaceTextEdits, c as metaSearch, d as renderMetaSearchText, m as landscape, n as taskMatrix, o as skillManifold, p as formatLandscape, r as formatSkillManifold, t as tree, w as editCreditText } from "./tree-CZ78tiPg.js";
+import { C as editCredit, S as front, _ as surfaceTextEdits, c as metaSearch, d as renderMetaSearchText, m as landscape, n as taskMatrix, o as skillManifold, p as formatLandscape, r as formatSkillManifold, t as tree, w as editCreditText } from "./tree-kc1d80yF.js";
 import { t as runSupervisorRunCommand } from "./report-command-CtIIsQKP.js";
 import { createHash } from "node:crypto";
 import { existsSync, readFileSync, writeFileSync } from "node:fs";

package/dist/contract/index.d.ts

@@ -1,14 +1,14 @@
 import { c as CostLedgerHandle } from "../cost-ledger-DbQdN3nO.js";
 import { _ as ProposalFinding, i as AnalystFinding, v as ProposalFindingOrigin, x as makeProposalFinding } from "../types-KEkKkETv.js";
 import { A as createChatClient, T as CreateChatClientOpts, x as ChatClient } from "../types-Lub7o6bN.js";
-import { $ as paretoSignificanceGate, G as BuildEvidenceVectorOptions, J as ParetoSignificanceGateOptions, K as EvidenceVector, Q as paretoPolicy, U as AxisEvidence, W as AxisVerdict, X as PromotionPolicy, Y as PromotionObjective, Z as buildEvidenceVector, q as ObjectiveSource } from "../statistical-heldout-D0rX_IDl.js";
+import { $ as paretoSignificanceGate, G as BuildEvidenceVectorOptions, J as ParetoSignificanceGateOptions, K as EvidenceVector, Q as paretoPolicy, U as AxisEvidence, W as AxisVerdict, X as PromotionPolicy, Y as PromotionObjective, Z as buildEvidenceVector, q as ObjectiveSource } from "../statistical-heldout-8RGPIeyo.js";
 import { C as JudgeDimension, H as SurfaceProposer, M as OptimizerConfig, R as Scenario, S as JudgeConfig, V as SessionScript, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, h as GateContext, i as CampaignCostMeter, j as MutableSurface, l as CodeSurface, m as GateCheckStatus, n as CampaignArtifactWriter, p as Gate, r as CampaignCellResult, t as CampaignAggregates, v as GateResult, w as JudgeScore, y as GenerationCandidate } from "../types-Dv9sYwBB.js";
-import { $i as RunImprovementLoopOptions, $r as GepaOptimizationRecipe, Bi as ReferenceEquivalenceJudgeInput, Ci as compareOptimizationMethods, Dr as DefaultProductionGateOptions, Er as DefaultProductionGateCheck, Fa as campaignSplitDigest, Gi as runReferenceEquivalenceJudge, Gr as SkillOptOptimizationMethodConfig, Hi as ReferenceEquivalenceJudgeResult, Ji as llmJudge, Jr as skillOptOptimizationMethod, Ki as LlmJudgeDimension, Kr as SkillOptRunnerCommand, Oa as FinalEvidencePolicy, Or as DefaultProductionRewardHackingOptions, Qi as transientDispatchFailure, Qr as GepaOptimizationMethodConfig, Ri as REFERENCE_EQUIVALENCE_INPUT_LIMITS, Si as OptimizationTokenUsage, Ti as ComparisonCost, Tr as heldOutGate, Ui as ReferenceEquivalenceScenario, Vi as ReferenceEquivalenceJudgeOptions, Wi as createReferenceEquivalenceJudge, Wr as composeGate, Xr as GepaEngineOptions, Yr as GepaAdaptiveEngineRun, Zr as GepaEngineRun, _a as runCampaign, _i as OptimizationMethodProvenance, ai as ExternalTextOptimizationMethodConfig, ci as ExternalOptimizationExample, di as CompareOptimizationMethodsOptions, ea as RunImprovementLoopResult, ei as GepaRunnerCommand, fa as RunEvalOptions, fi as OptimizationMethod, ga as RunCampaignOptions, ha as CampaignCellRetryPolicy, hi as OptimizationMethodInput, ii as externalTextOptimizationMethod, ka as FinalEvidenceUse, kr as defaultProductionGate, li as ExternalTextEvaluationResponse, ma as CampaignCellFailureReceipt, ni as OpenAICompatibleOptimizerModel, oi as ExternalTextOptimizerContext, pa as runEval, pi as OptimizationMethodComparison, qi as LlmJudgeOptions, qr as SkillOptTrainerConfig, ri as OptimizerModelBudget, si as ExternalTextOptimizerResult, ta as runImprovementLoop, ti as gepaOptimizationMethod, vi as OptimizationMethodResult, wr as HeldOutGateOptions, xi as OptimizationPackageSource, zi as REFERENCE_EQUIVALENCE_JUDGE_VERSION } from "../index-Ddi5M-TQ.js";
+import { $i as RunImprovementLoopOptions, $r as GepaOptimizationRecipe, Bi as ReferenceEquivalenceJudgeInput, Ci as compareOptimizationMethods, Dr as DefaultProductionGateOptions, Er as DefaultProductionGateCheck, Fa as campaignSplitDigest, Gi as runReferenceEquivalenceJudge, Gr as SkillOptOptimizationMethodConfig, Hi as ReferenceEquivalenceJudgeResult, Ji as llmJudge, Jr as skillOptOptimizationMethod, Ki as LlmJudgeDimension, Kr as SkillOptRunnerCommand, Oa as FinalEvidencePolicy, Or as DefaultProductionRewardHackingOptions, Qi as transientDispatchFailure, Qr as GepaOptimizationMethodConfig, Ri as REFERENCE_EQUIVALENCE_INPUT_LIMITS, Si as OptimizationTokenUsage, Ti as ComparisonCost, Tr as heldOutGate, Ui as ReferenceEquivalenceScenario, Vi as ReferenceEquivalenceJudgeOptions, Wi as createReferenceEquivalenceJudge, Wr as composeGate, Xr as GepaEngineOptions, Yr as GepaAdaptiveEngineRun, Zr as GepaEngineRun, _a as runCampaign, _i as OptimizationMethodProvenance, ai as ExternalTextOptimizationMethodConfig, ci as ExternalOptimizationExample, di as CompareOptimizationMethodsOptions, ea as RunImprovementLoopResult, ei as GepaRunnerCommand, fa as RunEvalOptions, fi as OptimizationMethod, ga as RunCampaignOptions, ha as CampaignCellRetryPolicy, hi as OptimizationMethodInput, ii as externalTextOptimizationMethod, ka as FinalEvidenceUse, kr as defaultProductionGate, li as ExternalTextEvaluationResponse, ma as CampaignCellFailureReceipt, ni as OpenAICompatibleOptimizerModel, oi as ExternalTextOptimizerContext, pa as runEval, pi as OptimizationMethodComparison, qi as LlmJudgeOptions, qr as SkillOptTrainerConfig, ri as OptimizerModelBudget, si as ExternalTextOptimizerResult, ta as runImprovementLoop, ti as gepaOptimizationMethod, vi as OptimizationMethodResult, wr as HeldOutGateOptions, xi as OptimizationPackageSource, zi as REFERENCE_EQUIVALENCE_JUDGE_VERSION } from "../index-xnhRdqik.js";
 import { $t as fsCampaignStorage, Zt as CampaignStorage, en as inMemoryCampaignStorage } from "../search-kernel-ah2GESe3.js";
 import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "../outcome-store-CNt4iZ67.js";
 import { r as HostedTenant } from "../client-BzL9niqi.js";
 import { _ as ReleaseSummary, a as summarizeExecution, c as ExecutionInsight, d as InsightReport, f as InterRaterInsight, g as Recommendation, h as OutcomeCorrelationInsight, i as analyzeRuns, l as FailureClassTally, m as LiftInsight, n as ExecutionReport, o as CostProvenanceSummary, p as JudgeInsight, r as SummarizeExecutionOptions, s as ExecutionErrorOutcomeCell, t as AnalyzeRunsOptions, u as FailureClusterInsight, v as ScalarDistribution, y as TokenUsageInsight } from "../analyze-runs-BlslTkOJ.js";
-import { _ as SelfImproveMethodResult, a as DefinedAgentEval, c as SelfImproveMethodOptions, d as SelfImproveProposerOptions, f as SelfImproveProposerResult, g as SelfImproveMethodProvenance, h as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveOptions, m as SelfImproveRunError, n as AgentEvalEvaluateOptions, o as defineAgentEval, p as SelfImproveResult, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveProgressEvent } from "../define-agent-eval-_nyV5nev.js";
-import { $ as FromRunRecordDirResult, A as CodeAgentSessionActionStatus, B as AgentTraceFile, C as fromOpenCodeSession, Ct as measuredComparisonFromCandidateExperiment, D as streamCodeAgentJsonlFile, Dt as sealCandidateExperiment, E as parseCodeAgentJsonlFile, Et as sealCandidateBenchmarkTask, F as CodeAgentSessionTerminalStatus, G as PartitionByAuthoringModelResult, H as AgentTraceRange, I as observeCodeAgentSession, J as EvalReportingSuiteInput, K as parseAgentTrace, L as AgentTraceContributor, M as CodeAgentSessionExecutionReceipt, N as CodeAgentSessionObservation, O as CodeAgentSessionAction, Ot as verifyCandidateExperiment, P as CodeAgentSessionSource, Q as FromRunRecordDirOptions, R as AgentTraceContributorType, S as fromKimiCodeSession, St as evaluatePairedMeasurements, T as parseCodeAgentJsonl, Tt as sealCandidateBenchmarkSuite, U as AgentTraceRecord, V as AgentTraceIndex, W as AuthoringProvenance, X as EvalReportingSuiteResult, Y as EvalReportingSuiteOptions, Z as evalReportingSuite, _ as CodeAgentSessionIntakeResult, _t as PairedMeasurement, a as FromFeedbackTableOptions, at as RunAgentProfileImprovementExperimentOptions, b as fromClaudeCodeSession, bt as RunCandidateExperimentOptions, c as CodeAgentStoreObservation, ct as runAgentProfileImprovementExperiment, d as CodeAgentStoreScope, dt as sealAgentProfileImprovementTask, et as RunRecordRejection, f as CodeAgentStoreSession, ft as verifyAgentProfileImprovementExperimentComparison, g as CodeAgentSessionIntakeOptions, gt as EvaluatePairedMeasurementsOptions, h as CodeAgentSessionDiagnostic, ht as CompareCandidateExperimentOptions, i as FeedbackTableRow, it as CompareAgentProfileImprovementExperimentOptions, j as CodeAgentSessionActionSurface, k as CodeAgentSessionActionKind, kt as verifyCandidateExperimentComparison, l as CodeAgentStoreRejection, lt as sealAgentProfileImprovementExperiment, m as CodeAgentJsonlLine, mt as CandidateExperimentRun, n as fromOtelSpans, nt as AgentProfileImprovementExperimentExecutionInput, o as FromFeedbackTableResult, ot as SealAgentProfileImprovementSuiteOptions, p as observeCodeAgentStore, pt as CandidateExperimentExecutionInput, q as partitionRunsByAuthoringModel, r as FeedbackTableMeta, rt as AgentProfileImprovementExperimentRun, s as fromFeedbackTable, st as measuredComparisonFromAgentProfileImprovementExperiment, t as FromOtelSpansOptions, tt as fromRunRecordDir, u as CodeAgentStoreRejections, ut as sealAgentProfileImprovementSuite, v as CodeAgentSessionMetrics, vt as PairedMeasurementAdapter, w as fromPiSession, wt as runCandidateExperiment, x as fromCodexSession, xt as SealCandidateBenchmarkSuiteOptions, y as ParsedCodeAgentJsonl, yt as PairedMeasurementEvaluation, z as AgentTraceConversation } from "../index-BHh9Z9El.js";
+import { _ as SelfImproveMethodResult, a as DefinedAgentEval, c as SelfImproveMethodOptions, d as SelfImproveProposerOptions, f as SelfImproveProposerResult, g as SelfImproveMethodProvenance, h as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveOptions, m as SelfImproveRunError, n as AgentEvalEvaluateOptions, o as defineAgentEval, p as SelfImproveResult, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveProgressEvent } from "../define-agent-eval-B_zZY2AZ.js";
+import { $ as FromRunRecordDirOptions, A as CodeAgentSessionActionKind, At as verifyCandidateExperimentComparison, B as AgentTraceConversation, C as fromOpenCodeSession, Ct as evaluatePairedMeasurements, D as parseCodeAgentJsonlFile, Dt as sealCandidateBenchmarkTask, E as parseCodeAgentJsonl, Et as sealCandidateBenchmarkSuite, F as CodeAgentSessionSource, G as AuthoringProvenance, H as AgentTraceIndex, I as CodeAgentSessionTerminalStatus, J as partitionRunsByAuthoringModel, K as PartitionByAuthoringModelResult, L as observeCodeAgentSession, M as CodeAgentSessionActionSurface, N as CodeAgentSessionExecutionReceipt, O as streamCodeAgentJsonlFile, Ot as sealCandidateExperiment, P as CodeAgentSessionObservation, Q as evalReportingSuite, R as AgentTraceContributor, S as fromKimiCodeSession, St as SealCandidateBenchmarkSuiteOptions, T as fromPiSession, Tt as runCandidateExperiment, U as AgentTraceRange, V as AgentTraceFile, W as AgentTraceRecord, X as EvalReportingSuiteOptions, Y as EvalReportingSuiteInput, Z as EvalReportingSuiteResult, _ as CodeAgentSessionIntakeResult, _t as EvaluatePairedMeasurementsOptions, a as FromFeedbackTableOptions, at as CompareAgentProfileImprovementExperimentOptions, b as fromClaudeCodeSession, bt as PairedMeasurementEvaluation, c as CodeAgentStoreObservation, ct as measuredComparisonFromAgentProfileImprovementExperiment, d as CodeAgentStoreScope, dt as sealAgentProfileImprovementSuite, et as FromRunRecordDirResult, f as CodeAgentStoreSession, ft as sealAgentProfileImprovementTask, g as CodeAgentSessionIntakeOptions, gt as CompareCandidateExperimentOptions, h as CodeAgentSessionDiagnostic, ht as CandidateExperimentRun, i as FeedbackTableRow, it as AgentProfileImprovementExperimentRun, j as CodeAgentSessionActionStatus, k as CodeAgentSessionAction, kt as verifyCandidateExperiment, l as CodeAgentStoreRejection, lt as runAgentProfileImprovementExperiment, m as CodeAgentJsonlLine, mt as CandidateExperimentExecutionInput, n as fromOtelSpans, nt as fromRunRecordDir, o as FromFeedbackTableResult, ot as RunAgentProfileImprovementExperimentOptions, p as observeCodeAgentStore, pt as verifyAgentProfileImprovementExperimentComparison, q as parseAgentTrace, r as FeedbackTableMeta, rt as AgentProfileImprovementExperimentExecutionInput, s as fromFeedbackTable, st as SealAgentProfileImprovementSuiteOptions, t as FromOtelSpansOptions, tt as RunRecordRejection, u as CodeAgentStoreRejections, ut as sealAgentProfileImprovementExperiment, v as CodeAgentSessionMetrics, vt as PairedMeasurement, w as fromPiGraphSession, wt as measuredComparisonFromCandidateExperiment, x as fromCodexSession, xt as RunCandidateExperimentOptions, y as ParsedCodeAgentJsonl, yt as PairedMeasurementAdapter, z as AgentTraceContributorType } from "../index-DnZdBxCL.js";
 import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-Crk3QEy1.js";
-export { type AgentEvalAgent, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, type AgentProfileImprovementExperimentExecutionInput, type AgentProfileImprovementExperimentRun, type AgentTraceContributor, type AgentTraceContributorType, type AgentTraceConversation, type AgentTraceFile, type AgentTraceIndex, type AgentTraceRange, type AgentTraceRecord, type AnalystFinding, type AnalyzeRunsOptions, type AuthoringProvenance, type AxisEvidence, type AxisVerdict, type BuildEvidenceVectorOptions, type CampaignAggregates, type CampaignArtifactWriter, type CampaignCellFailureReceipt, type CampaignCellResult, type CampaignCellRetryPolicy, type CampaignCostMeter, type CampaignResult, type CampaignStorage, type CampaignTraceWriter, type CandidateExperimentExecutionInput, type CandidateExperimentRun, type ChatClient, type CodeAgentJsonlLine, type CodeAgentSessionAction, type CodeAgentSessionActionKind, type CodeAgentSessionActionStatus, type CodeAgentSessionActionSurface, type CodeAgentSessionDiagnostic, type CodeAgentSessionExecutionReceipt, type CodeAgentSessionIntakeOptions, type CodeAgentSessionIntakeResult, type CodeAgentSessionMetrics, type CodeAgentSessionObservation, type CodeAgentSessionSource, type CodeAgentSessionTerminalStatus, type CodeAgentStoreObservation, type CodeAgentStoreRejection, type CodeAgentStoreRejections, type CodeAgentStoreScope, type CodeAgentStoreSession, type CodeSurface, type CompareAgentProfileImprovementExperimentOptions, type CompareCandidateExperimentOptions, type CompareOptimizationMethodsOptions, type ComparisonCost, type CostLedgerHandle, type CostProvenanceSummary, type CreateChatClientOpts, type DefaultAnalystRegistryOptions, type DefaultProductionGateCheck, type DefaultProductionGateOptions, type DefaultProductionRewardHackingOptions, type DefineAgentEvalOptions, type DefinedAgentEval, type DeploymentOutcome, type DispatchFn as Dispatch, type DispatchContext, type EvalReportingSuiteInput, type EvalReportingSuiteOptions, type EvalReportingSuiteResult, type EvaluatePairedMeasurementsOptions, type EvidenceVector, type ExecutionErrorOutcomeCell, type ExecutionInsight, type ExecutionReport, type ExternalOptimizationExample, type ExternalTextEvaluationResponse, type ExternalTextOptimizationMethodConfig, type ExternalTextOptimizerContext, type ExternalTextOptimizerResult, type FailureClassTally, type FailureClusterInsight, type FeedbackTableMeta, type FeedbackTableRow, FileSystemOutcomeStore, type FileSystemOutcomeStoreOptions, type FinalEvidencePolicy, type FinalEvidenceUse, type FromFeedbackTableOptions, type FromFeedbackTableResult, type FromOtelSpansOptions, type FromRunRecordDirOptions, type FromRunRecordDirResult, type Gate, type GateCheckStatus, type GateContext, type GateContribution, type GateDecision, type GateResult, type GenerationCandidate, type GenerationRecord, type GepaAdaptiveEngineRun, type GepaEngineOptions, type GepaEngineRun, type GepaOptimizationMethodConfig, type GepaOptimizationRecipe, type GepaRunnerCommand, type HeldOutGateOptions, type HostedTenant, InMemoryOutcomeStore, type InsightReport, type InterRaterInsight, type JudgeConfig, type JudgeDimension, type JudgeInsight, type JudgeScore, type LiftInsight, type LlmJudgeDimension, type LlmJudgeOptions, type MutableSurface, type ObjectiveSource, type OpenAICompatibleOptimizerModel, type OptimizationMethod, type OptimizationMethodComparison, type OptimizationMethodInput, type OptimizationMethodProvenance, type OptimizationMethodResult, type OptimizationPackageSource, type OptimizationTokenUsage, type OptimizerConfig, type OptimizerModelBudget, type OutcomeCorrelationInsight, type OutcomeStore, type PairedMeasurement, type PairedMeasurementAdapter, type PairedMeasurementEvaluation, type ParetoSignificanceGateOptions, type ParsedCodeAgentJsonl, type PartitionByAuthoringModelResult, type PromotionObjective, type PromotionPolicy, type ProposalFinding, type ProposalFindingOrigin, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, type Recommendation, type ReferenceEquivalenceJudgeInput, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceJudgeResult, type ReferenceEquivalenceScenario, type ReleaseSummary, type RunAgentProfileImprovementExperimentOptions, type RunCampaignOptions, type RunCandidateExperimentOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunRecordRejection, type ScalarDistribution, type Scenario, type SealAgentProfileImprovementSuiteOptions, type SealCandidateBenchmarkSuiteOptions, type SelfImproveBudget, type SelfImproveMethodOptions, type SelfImproveMethodProvenance, type SelfImproveMethodResult, type SelfImproveOptions, type SelfImproveProgressEvent, type SelfImproveProposerOptions, type SelfImproveProposerResult, type SelfImproveResult, SelfImproveRunError, type SessionScript, type SkillOptOptimizationMethodConfig, type SkillOptRunnerCommand, type SkillOptTrainerConfig, type SummarizeExecutionOptions, type SurfaceProposer, type TokenUsageInsight, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, createReferenceEquivalenceJudge, defaultProductionGate, defineAgentEval, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, makeProposalFinding, measuredComparisonFromAgentProfileImprovementExperiment, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, observeCodeAgentStore, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, parseCodeAgentJsonlFile, partitionRunsByAuthoringModel, runAgentProfileImprovementExperiment, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealAgentProfileImprovementExperiment, sealAgentProfileImprovementSuite, sealAgentProfileImprovementTask, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, streamCodeAgentJsonlFile, summarizeExecution, transientDispatchFailure, verifyAgentProfileImprovementExperimentComparison, verifyCandidateExperiment, verifyCandidateExperimentComparison };
+export { type AgentEvalAgent, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, type AgentProfileImprovementExperimentExecutionInput, type AgentProfileImprovementExperimentRun, type AgentTraceContributor, type AgentTraceContributorType, type AgentTraceConversation, type AgentTraceFile, type AgentTraceIndex, type AgentTraceRange, type AgentTraceRecord, type AnalystFinding, type AnalyzeRunsOptions, type AuthoringProvenance, type AxisEvidence, type AxisVerdict, type BuildEvidenceVectorOptions, type CampaignAggregates, type CampaignArtifactWriter, type CampaignCellFailureReceipt, type CampaignCellResult, type CampaignCellRetryPolicy, type CampaignCostMeter, type CampaignResult, type CampaignStorage, type CampaignTraceWriter, type CandidateExperimentExecutionInput, type CandidateExperimentRun, type ChatClient, type CodeAgentJsonlLine, type CodeAgentSessionAction, type CodeAgentSessionActionKind, type CodeAgentSessionActionStatus, type CodeAgentSessionActionSurface, type CodeAgentSessionDiagnostic, type CodeAgentSessionExecutionReceipt, type CodeAgentSessionIntakeOptions, type CodeAgentSessionIntakeResult, type CodeAgentSessionMetrics, type CodeAgentSessionObservation, type CodeAgentSessionSource, type CodeAgentSessionTerminalStatus, type CodeAgentStoreObservation, type CodeAgentStoreRejection, type CodeAgentStoreRejections, type CodeAgentStoreScope, type CodeAgentStoreSession, type CodeSurface, type CompareAgentProfileImprovementExperimentOptions, type CompareCandidateExperimentOptions, type CompareOptimizationMethodsOptions, type ComparisonCost, type CostLedgerHandle, type CostProvenanceSummary, type CreateChatClientOpts, type DefaultAnalystRegistryOptions, type DefaultProductionGateCheck, type DefaultProductionGateOptions, type DefaultProductionRewardHackingOptions, type DefineAgentEvalOptions, type DefinedAgentEval, type DeploymentOutcome, type DispatchFn as Dispatch, type DispatchContext, type EvalReportingSuiteInput, type EvalReportingSuiteOptions, type EvalReportingSuiteResult, type EvaluatePairedMeasurementsOptions, type EvidenceVector, type ExecutionErrorOutcomeCell, type ExecutionInsight, type ExecutionReport, type ExternalOptimizationExample, type ExternalTextEvaluationResponse, type ExternalTextOptimizationMethodConfig, type ExternalTextOptimizerContext, type ExternalTextOptimizerResult, type FailureClassTally, type FailureClusterInsight, type FeedbackTableMeta, type FeedbackTableRow, FileSystemOutcomeStore, type FileSystemOutcomeStoreOptions, type FinalEvidencePolicy, type FinalEvidenceUse, type FromFeedbackTableOptions, type FromFeedbackTableResult, type FromOtelSpansOptions, type FromRunRecordDirOptions, type FromRunRecordDirResult, type Gate, type GateCheckStatus, type GateContext, type GateContribution, type GateDecision, type GateResult, type GenerationCandidate, type GenerationRecord, type GepaAdaptiveEngineRun, type GepaEngineOptions, type GepaEngineRun, type GepaOptimizationMethodConfig, type GepaOptimizationRecipe, type GepaRunnerCommand, type HeldOutGateOptions, type HostedTenant, InMemoryOutcomeStore, type InsightReport, type InterRaterInsight, type JudgeConfig, type JudgeDimension, type JudgeInsight, type JudgeScore, type LiftInsight, type LlmJudgeDimension, type LlmJudgeOptions, type MutableSurface, type ObjectiveSource, type OpenAICompatibleOptimizerModel, type OptimizationMethod, type OptimizationMethodComparison, type OptimizationMethodInput, type OptimizationMethodProvenance, type OptimizationMethodResult, type OptimizationPackageSource, type OptimizationTokenUsage, type OptimizerConfig, type OptimizerModelBudget, type OutcomeCorrelationInsight, type OutcomeStore, type PairedMeasurement, type PairedMeasurementAdapter, type PairedMeasurementEvaluation, type ParetoSignificanceGateOptions, type ParsedCodeAgentJsonl, type PartitionByAuthoringModelResult, type PromotionObjective, type PromotionPolicy, type ProposalFinding, type ProposalFindingOrigin, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, type Recommendation, type ReferenceEquivalenceJudgeInput, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceJudgeResult, type ReferenceEquivalenceScenario, type ReleaseSummary, type RunAgentProfileImprovementExperimentOptions, type RunCampaignOptions, type RunCandidateExperimentOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunRecordRejection, type ScalarDistribution, type Scenario, type SealAgentProfileImprovementSuiteOptions, type SealCandidateBenchmarkSuiteOptions, type SelfImproveBudget, type SelfImproveMethodOptions, type SelfImproveMethodProvenance, type SelfImproveMethodResult, type SelfImproveOptions, type SelfImproveProgressEvent, type SelfImproveProposerOptions, type SelfImproveProposerResult, type SelfImproveResult, SelfImproveRunError, type SessionScript, type SkillOptOptimizationMethodConfig, type SkillOptRunnerCommand, type SkillOptTrainerConfig, type SummarizeExecutionOptions, type SurfaceProposer, type TokenUsageInsight, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, createReferenceEquivalenceJudge, defaultProductionGate, defineAgentEval, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiGraphSession, fromPiSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, makeProposalFinding, measuredComparisonFromAgentProfileImprovementExperiment, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, observeCodeAgentStore, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, parseCodeAgentJsonlFile, partitionRunsByAuthoringModel, runAgentProfileImprovementExperiment, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealAgentProfileImprovementExperiment, sealAgentProfileImprovementSuite, sealAgentProfileImprovementTask, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, streamCodeAgentJsonlFile, summarizeExecution, transientDispatchFailure, verifyAgentProfileImprovementExperimentComparison, verifyCandidateExperiment, verifyCandidateExperimentComparison };

package/dist/contract/index.js

@@ -1,29 +1,30 @@
 import { a as hashCanonical } from "../canonical-BsZVgBhW.js";
 import { r as pairedBootstrap } from "../paired-tests-jbwPdSz5.js";
-import { t as decidePairedPromotion } from "../paired-promotion-decision-2deJ7sDX.js";
-import { a as heldoutSignificance } from "../statistical-heldout-BP6iEJvR.js";
-import { n as summarizeExecution, t as analyzeRuns } from "../analyze-runs-0cdYMgob.js";
+import { t as decidePairedPromotion } from "../paired-promotion-decision-BN6IQTiu.js";
+import { a as heldoutSignificance } from "../statistical-heldout-1Gbmex7v.js";
+import { n as summarizeExecution, t as analyzeRuns } from "../analyze-runs-3xpby3p2.js";
 import { a as parseRunRecordSafe } from "../run-record-DlTZ_lqe.js";
-import { K as defaultProductionGate, a as transientDispatchFailure, c as compareOptimizationMethods, h as runImprovementLoop, nt as runEval, rt as runCampaign, t as llmJudge } from "../llm-judge-C0256VT_.js";
+import { K as defaultProductionGate, a as transientDispatchFailure, c as compareOptimizationMethods, h as runImprovementLoop, nt as runEval, rt as runCampaign, t as llmJudge } from "../llm-judge-Tr7MBUnf.js";
 import { i as isModelPriced, n as estimateCost } from "../metrics-DOMpkoeb.js";
 import { i as CostLedger } from "../cost-ledger-_IbkejCF.js";
-import { G as campaignSplitDigest } from "../campaign-evidence-B5qsCHEn.js";
+import { G as campaignSplitDigest } from "../campaign-evidence-OcUciD-N.js";
 import { h as mapConcurrentRange } from "../journal-DvKCvjty.js";
 import { C as inMemoryCampaignStorage, S as fsCampaignStorage } from "../external-optimizer-subprocess-C0EsxEbI.js";
-import { n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-Csd47ur_.js";
+import { n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-CoXXDCMu.js";
 import { t as buildDefaultAnalystRegistry } from "../default-registry-DF4Yp7HQ.js";
 import { t as extractUsage } from "../extract-usage-BrnzRJ6H.js";
 import { i as makeProposalFinding } from "../types-CW5b90JH.js";
 import { t as createChatClient } from "../chat-client-CydfmLdO.js";
 import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "../outcome-store-DUcMj-3D.js";
-import { c as heldOutGate, g as runReferenceEquivalenceJudge, h as createReferenceEquivalenceJudge, l as composeGate, m as REFERENCE_EQUIVALENCE_JUDGE_VERSION, n as gepaOptimizationMethod, p as REFERENCE_EQUIVALENCE_INPUT_LIMITS, t as skillOptOptimizationMethod, u as externalTextOptimizationMethod } from "../skillopt-optimization-method-BCojvoP1.js";
-import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-CCmyEH_c.js";
+import { c as heldOutGate, g as runReferenceEquivalenceJudge, h as createReferenceEquivalenceJudge, l as composeGate, m as REFERENCE_EQUIVALENCE_JUDGE_VERSION, n as gepaOptimizationMethod, p as REFERENCE_EQUIVALENCE_INPUT_LIMITS, t as skillOptOptimizationMethod, u as externalTextOptimizationMethod } from "../skillopt-optimization-method-DujPOiwD.js";
+import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-CWW35UWq.js";
 import { t as fromOtelSpans } from "../otel-spans-C8YIJWy1.js";
 import { createHash } from "node:crypto";
 import { agentCandidateBenchmarkSuiteSchema, agentCandidateBenchmarkTaskSchema, agentCandidateBundleSchema, agentCandidateEvaluationPolicySchema, agentCandidateExperimentSchema, agentImprovementMeasuredComparisonSchema, agentProfileImprovementExperimentSchema, agentProfileImprovementMeasuredComparisonSchema, agentProfileImprovementRunCellSchema, agentProfileImprovementRunReceiptSchema, agentProfileImprovementSuiteInputsSchema, agentProfileImprovementSuiteSchema, agentProfileImprovementTaskSchema, candidateExecutionEvidenceSchema, canonicalCandidateDigest, canonicalCandidateJson, numbersApproximatelyEqual, omitTopLevelDigest } from "@tangle-network/agent-interface";
 import { createReadStream } from "node:fs";
 import { dirname, isAbsolute, join, relative, resolve, sep } from "node:path";
 import { mkdir, readFile, readdir, stat, writeFile } from "node:fs/promises";
+import { readSessionInput, summarizeSessionInput } from "@tangle-network/harness-sessions";
 //#region src/contract/fixed-spend.ts
 function addFixedSpend(left, right) {
 	return {
@@ -1403,6 +1404,11 @@ function partitionRunsByAuthoringModel(runs, index) {
 //#endregion
 //#region src/contract/intake/code-agent-observation.ts
 /**
+* Where a session came from. `pi` is a native Pi session (`~/.pi/agent/sessions/*.jsonl`), read
+* through @tangle-network/harness-sessions; `pi-graph` is the graph IR a Pi graph run exports
+* (`nodes[].ir.kind`), which is not a Pi session.
+*/
+/**
 * Project one provider session into the exact user-visible answer and a
 * provider-neutral action stream. Raw prompts, tool inputs, and tool outputs
 * stay out of this projection; callers retain the source JSONL as evidence.
@@ -1425,7 +1431,8 @@ function projectionFor(source, entries) {
 		case "claude-code": return claudeProjection(entries);
 		case "opencode": return openCodeProjection(entries);
 		case "kimi-code": return kimiProjection(entries);
-		case "pi": return piProjection(entries);
+		case "pi": return piSessionProjection(entries);
+		case "pi-graph": return piGraphProjection(entries);
 	}
 }
 function codexProjection(entries) {
@@ -1799,7 +1806,30 @@ function kimiProjection(entries) {
 		actions
 	};
 }
-function piProjection(entries) {
+/** A native Pi session, folded by the shared reader: one action per tool call, the last answer, the ending. */
+function piSessionProjection(entries) {
+	const session = readSessionInput("pi", { records: entries });
+	const actions = session.toolCalls.map((call, index) => actionFor({
+		id: call.id,
+		stepIndex: index,
+		kind: "tool",
+		surface: surfaceForTool(call.name),
+		name: call.name,
+		status: call.status === "completed" ? "completed" : call.status === "error" ? "failed" : call.status === "pending" ? "started" : "unknown",
+		timestampMs: timestamp(call.startedAt),
+		metadata: { sourceEventType: "pi-tool-call" }
+	}));
+	const answer = [...session.messages].reverse().find((m) => m.role === "assistant" && m.actor === "agent" && m.parts.some((p) => p.type === "text"));
+	const finalText = answer ? answer.parts.flatMap((p) => p.type === "text" ? [p.text] : []).join("\n") : null;
+	const status = session.ending.status;
+	return {
+		finalText,
+		terminal: status === "completed" ? "completed" : status === "error" || status === "aborted" ? "failed" : "unknown",
+		explicitTerminal: status !== "open",
+		actions
+	};
+}
+function piGraphProjection(entries) {
 	const actions = [];
 	let terminal = "unknown";
 	let explicitTerminal = false;
@@ -2037,6 +2067,10 @@ function compactMetadata(input) {
 }
 function sessionIdFromEntries$1(source, entries) {
 	for (const entry of entries) {
+		if (source === "pi" && entry.type === "session") {
+			const id = stringField$1(entry, "id");
+			if (id) return id;
+		}
 		if (source === "codex") {
 			const threadId = stringField$1(entry, "thread_id");
 			if (threadId) return threadId;
@@ -2178,9 +2212,14 @@ function fromOpenCodeSession(options) {
 function fromKimiCodeSession(options) {
 	return fromCodeAgentSession("kimi-code", options);
 }
+/** A native Pi session (`~/.pi/agent/sessions/<cwd>/<ts>_<id>.jsonl`), parsed entries in order. */
 function fromPiSession(options) {
 	return fromCodeAgentSession("pi", options);
 }
+/** The graph IR a Pi graph run exports (`nodes[].ir`), which is not a Pi session. */
+function fromPiGraphSession(options) {
+	return fromCodeAgentSession("pi-graph", options);
+}
 function fromCodeAgentSession(source, options) {
 	const entries = options.entries.filter(isRecord);
 	if (entries.length === 0) return {
@@ -2347,7 +2386,8 @@ function metricsFor(source, entries) {
 		case "claude-code": return claudeCodeMetrics(entries);
 		case "opencode": return openCodeMetrics(entries);
 		case "kimi-code": return kimiCodeMetrics(entries);
-		case "pi": return piMetrics(entries);
+		case "pi": return piSessionMetrics(entries);
+		case "pi-graph": return piGraphMetrics(entries);
 	}
 }
 function codexMetrics(entries) {
@@ -2488,7 +2528,6 @@ function claudeCodeMetrics(entries) {
 		if (type === "pr-link") metrics.prLinks += 1;
 		if (type === "file-history-snapshot") metrics.fileSnapshots += 1;
 		const message = record(entry.message);
-		if (message) addUsage(metrics, record(message.usage));
 		const content = Array.isArray(message?.content) ? message.content : [];
 		for (const item of content) {
 			const part = record(item);
@@ -2502,6 +2541,12 @@ function claudeCodeMetrics(entries) {
 			}
 		}
 	}
+	const usage = summarizeSessionInput("claude-code", { records: entries }).usage;
+	metrics.inputTokens = usage?.input ?? 0;
+	metrics.outputTokens = usage?.output ?? 0;
+	metrics.reasoningTokens = usage?.reasoning ?? 0;
+	metrics.cachedTokens = usage?.cacheRead ?? 0;
+	metrics.cacheWriteTokens = usage?.cacheWrite ?? 0;
 	if (startedAt !== void 0 && completedAt !== void 0) metrics.wallMs = Math.max(0, completedAt - startedAt);
 	metrics.processScore = claudeProcessScore(metrics);
 	return metrics;
@@ -2617,7 +2662,33 @@ function kimiCodeMetrics(entries) {
 	metrics.processScore = terminalProcessScore(metrics);
 	return metrics;
 }
-function piMetrics(entries) {
+/** A native Pi session's metrics, from the shared reader's fold. */
+function piSessionMetrics(entries) {
+	const metrics = emptyMetrics(entries.length);
+	const session = readSessionInput("pi", { records: entries });
+	metrics.userMessages = session.messages.filter((m) => m.role === "user" && m.actor === "human").length;
+	metrics.assistantMessages = session.messages.filter((m) => m.role === "assistant" && m.actor === "agent").length;
+	metrics.reasoningItems = session.messages.reduce((n, m) => n + m.parts.filter((p) => p.type === "reasoning").length, 0);
+	metrics.toolCalls = session.toolCalls.length;
+	metrics.toolOutputs = session.toolCalls.filter((c) => c.result !== null).length;
+	metrics.toolErrors = session.toolCalls.filter((c) => c.result?.isError === true).length;
+	metrics.unclassifiedErrors = session.modelCalls.filter((c) => c.error !== null).length;
+	metrics.turnsStarted = metrics.userMessages;
+	if (session.ending.status === "completed") metrics.turnsCompleted = 1;
+	if (session.ending.status === "aborted" || session.ending.status === "error") metrics.turnsAborted = 1;
+	metrics.inputTokens = session.usage?.input ?? 0;
+	metrics.outputTokens = session.usage?.output ?? 0;
+	metrics.reasoningTokens = session.usage?.reasoning ?? 0;
+	metrics.cachedTokens = session.usage?.cacheRead ?? 0;
+	metrics.cacheWriteTokens = session.usage?.cacheWrite ?? 0;
+	const costs = session.modelCalls.flatMap((c) => c.costUsd === null ? [] : [c.costUsd]);
+	metrics.observedCostUsd = costs.reduce((a, b) => a + b, 0);
+	metrics.observedCostCaptured = costs.length > 0;
+	if (session.startedAt !== null && session.endedAt !== null) metrics.wallMs = Math.max(0, Date.parse(session.endedAt) - Date.parse(session.startedAt));
+	metrics.processScore = terminalProcessScore(metrics);
+	return metrics;
+}
+function piGraphMetrics(entries) {
 	const metrics = emptyMetrics(entries.length);
 	let bestReliabilityScore;
 	for (const entry of entries) {
@@ -2826,6 +2897,10 @@ function applyObservedSurfaceMetrics(metrics, observation) {
 }
 function sessionIdFromEntries(source, entries) {
 	for (const entry of entries) {
+		if (source === "pi" && entry.type === "session") {
+			const id = stringField(entry, "id");
+			if (id) return id;
+		}
 		if (source === "codex") {
 			const threadId = stringField(entry, "thread_id");
 			if (threadId) return threadId;
@@ -2839,6 +2914,7 @@ function sessionIdFromEntries(source, entries) {
 	}
 }
 function modelFromEntries(source, entries) {
+	if (source === "pi") return summarizeSessionInput("pi", { records: entries }).lastServedModel ?? void 0;
 	let providerModel;
 	for (const entry of entries) {
 		if (source === "codex") {
@@ -3224,6 +3300,6 @@ function fromFeedbackTable(opts) {
 	};
 }
 //#endregion
-export { FileSystemOutcomeStore, InMemoryOutcomeStore, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, SelfImproveRunError, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, createReferenceEquivalenceJudge, defaultProductionGate, defineAgentEval, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, makeProposalFinding, measuredComparisonFromAgentProfileImprovementExperiment, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, observeCodeAgentStore, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, parseCodeAgentJsonlFile, partitionRunsByAuthoringModel, runAgentProfileImprovementExperiment, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealAgentProfileImprovementExperiment, sealAgentProfileImprovementSuite, sealAgentProfileImprovementTask, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, streamCodeAgentJsonlFile, summarizeExecution, transientDispatchFailure, verifyAgentProfileImprovementExperimentComparison, verifyCandidateExperiment, verifyCandidateExperimentComparison };
+export { FileSystemOutcomeStore, InMemoryOutcomeStore, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, SelfImproveRunError, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, createReferenceEquivalenceJudge, defaultProductionGate, defineAgentEval, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiGraphSession, fromPiSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, makeProposalFinding, measuredComparisonFromAgentProfileImprovementExperiment, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, observeCodeAgentStore, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, parseCodeAgentJsonlFile, partitionRunsByAuthoringModel, runAgentProfileImprovementExperiment, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealAgentProfileImprovementExperiment, sealAgentProfileImprovementSuite, sealAgentProfileImprovementTask, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, streamCodeAgentJsonlFile, summarizeExecution, transientDispatchFailure, verifyAgentProfileImprovementExperimentComparison, verifyCandidateExperiment, verifyCandidateExperimentComparison };
 
 //# sourceMappingURL=index.js.map