@tangle-network/agent-eval 0.144.9 → 0.144.10

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (54) hide show
  1. package/CHANGELOG.md +7 -0
  2. package/README.md +3 -0
  3. package/dist/analyst/index.d.ts +1 -1
  4. package/dist/analyst/index.js +7 -7
  5. package/dist/{benchmark-B181aMF9.js → benchmark-CWeqGl7x.js} +2 -2
  6. package/dist/{benchmark-B181aMF9.js.map → benchmark-CWeqGl7x.js.map} +1 -1
  7. package/dist/{benchmark-command-Dtym4pA-.js → benchmark-command-D3fEnKdf.js} +8 -8
  8. package/dist/{benchmark-command-Dtym4pA-.js.map → benchmark-command-D3fEnKdf.js.map} +1 -1
  9. package/dist/benchmarks/index.d.ts +1 -1
  10. package/dist/benchmarks/index.js +1 -1
  11. package/dist/{benchmarks-Ce4Ae5fX.js → benchmarks-BzqyTvAr.js} +4 -4
  12. package/dist/{benchmarks-Ce4Ae5fX.js.map → benchmarks-BzqyTvAr.js.map} +1 -1
  13. package/dist/campaign/index.d.ts +4 -4
  14. package/dist/campaign/index.js +5 -5
  15. package/dist/{campaign-FG3thH2m.js → campaign-ccOrVCuR.js} +4 -4
  16. package/dist/{campaign-FG3thH2m.js.map → campaign-ccOrVCuR.js.map} +1 -1
  17. package/dist/cli.js +1 -1
  18. package/dist/contract/index.d.ts +1 -1
  19. package/dist/contract/index.js +6 -6
  20. package/dist/{default-registry-BaQXW1Ow.js → default-registry-BmktKy8r.js} +3 -3
  21. package/dist/{default-registry-BaQXW1Ow.js.map → default-registry-BmktKy8r.js.map} +1 -1
  22. package/dist/{dspy-rlm-engine-BiN49gK6.js → dspy-rlm-engine-DbTk4JdR.js} +3 -3
  23. package/dist/{dspy-rlm-engine-BiN49gK6.js.map → dspy-rlm-engine-DbTk4JdR.js.map} +1 -1
  24. package/dist/{extract-usage-BW27f3XW.js → extract-usage-CdZdoj1s.js} +2 -2
  25. package/dist/{extract-usage-BW27f3XW.js.map → extract-usage-CdZdoj1s.js.map} +1 -1
  26. package/dist/{index-CLKMJTJF2.d.ts → index-CwYZNzMD2.d.ts} +3 -3
  27. package/dist/{index-CLKMJTJF2.d.ts.map → index-CwYZNzMD2.d.ts.map} +1 -1
  28. package/dist/{index-CBEMKO8d.d.ts → index-DvvvSho7.d.ts} +2 -2
  29. package/dist/{index-CBEMKO8d.d.ts.map → index-DvvvSho7.d.ts.map} +1 -1
  30. package/dist/index.d.ts +3 -3
  31. package/dist/index.js +11 -11
  32. package/dist/{kind-factory-BHIgPmzS.js → kind-factory-B8-r8-y8.js} +3 -16
  33. package/dist/kind-factory-B8-r8-y8.js.map +1 -0
  34. package/dist/openapi.json +1 -1
  35. package/dist/{replay-B3gACG_H.js → replay-CohS93nE.js} +4 -4
  36. package/dist/{replay-B3gACG_H.js.map → replay-CohS93nE.js.map} +1 -1
  37. package/dist/{semantic-concept-judge-Bmrq6yqU.js → semantic-concept-judge-D1z-KepS.js} +2 -2
  38. package/dist/{semantic-concept-judge-Bmrq6yqU.js.map → semantic-concept-judge-D1z-KepS.js.map} +1 -1
  39. package/dist/{single-run-lock-BMQEv1wG.js → single-run-lock-DFWHEB09.js} +123 -123
  40. package/dist/{single-run-lock-BMQEv1wG.js.map → single-run-lock-DFWHEB09.js.map} +1 -1
  41. package/dist/{skillopt-optimization-method-D8wysvU1.d.ts → skillopt-optimization-method-9J8dJbrM.d.ts} +35 -3
  42. package/dist/skillopt-optimization-method-9J8dJbrM.d.ts.map +1 -0
  43. package/dist/{skillopt-optimization-method-Dtzp7QpP.js → skillopt-optimization-method-D6Q4dLbh.js} +299 -132
  44. package/dist/skillopt-optimization-method-D6Q4dLbh.js.map +1 -0
  45. package/dist/{store-otlp-CKtTpRhv.js → store-otlp-Dw8PPIlL.js} +2 -2
  46. package/dist/{store-otlp-CKtTpRhv.js.map → store-otlp-Dw8PPIlL.js.map} +1 -1
  47. package/dist/traces.js +4 -4
  48. package/dist/{usage-receipt-EVI8B8Xu.js → usage-receipt-t7vAzCRQ.js} +15 -2
  49. package/dist/usage-receipt-t7vAzCRQ.js.map +1 -0
  50. package/package.json +1 -1
  51. package/dist/kind-factory-BHIgPmzS.js.map +0 -1
  52. package/dist/skillopt-optimization-method-D8wysvU1.d.ts.map +0 -1
  53. package/dist/skillopt-optimization-method-Dtzp7QpP.js.map +0 -1
  54. package/dist/usage-receipt-EVI8B8Xu.js.map +0 -1
@@ -1 +0,0 @@
1
- {"version":3,"file":"usage-receipt-EVI8B8Xu.js","names":[],"sources":["../src/analyst/types.ts","../src/analyst/usage-receipt.ts"],"sourcesContent":["/**\n * Analyst contract — the missing orchestration layer over agent-eval's\n * existing analyzers (analyzeTraces, MultiLayerVerifier, RunCritic,\n * SemanticConceptJudge, JudgeFn, ...).\n *\n * Each existing primitive returns its own output shape. The Analyst\n * contract is the single envelope every primitive lifts into, so a\n * registry can run N analysts against a run and a single renderer can\n * compose findings without knowing which analyzer produced them.\n *\n * The contract is intentionally domain-agnostic: nothing here knows\n * about code, voice, RAG, or any particular agent stack. Analysts\n * declare what INPUT KIND they need (a trace store, an artifact dir,\n * a RunRecord, a JudgeInput, or `custom`), and the registry routes\n * the matching input from `AnalystRunInputs`.\n */\n\nimport { createHash } from 'node:crypto'\nimport type { CostLedgerHandle } from '../cost-ledger'\nimport type { RunCostProvenance, RunRecord, RunTokenUsage } from '../run-record'\nimport type { TraceAnalysisStore } from '../trace-analyst/store'\nimport type { JudgeInput } from '../types'\nimport type { ChatClient } from './chat-client'\n\n/**\n * Unified envelope every analyst emits. Schema-versioned so renderers\n * and time-series diffs survive future field additions.\n */\nexport interface AnalystFinding {\n schema_version: '1.0.0'\n /**\n * Stable hash over identity-defining fields (analyst_id + canonical\n * claim + area + optional subject). Two findings from two runs that\n * \"are the same finding\" share this id — that's what `diffFindings`\n * uses to compute appeared/disappeared sets across runs.\n */\n finding_id: string\n analyst_id: string\n produced_at: string\n severity: AnalystSeverity\n /**\n * Coarse classification. Renderers group by this. Free-form so\n * domain-specific analysts can introduce categories without a\n * schema change ('agent-reasoning', 'verification', 'cost',\n * 'tool-use', 'safety', 'latency', 'data-quality', ...).\n */\n area: string\n claim: string\n rationale?: string\n evidence_refs: EvidenceRef[]\n recommended_action?: string\n validation_plan?: string\n /** 0..1 — the analyst's own confidence. Not calibrated across analysts. */\n confidence: number\n /**\n * Optional subject the finding is about — leaf id, agent id, request\n * id. Included in finding_id when present so per-subject findings\n * diff cleanly across runs.\n */\n subject?: string\n /** True when this finding was lifted from a judge result rather than observed\n * directly in a trace or artifact. Descriptive only: proposal access is\n * controlled by `ProposalFinding.proposal_origin`. */\n derived_from_judge?: boolean\n /** Analyst-private extras; renderers ignore unless they know the analyst. */\n metadata?: Record<string, unknown>\n}\n\nexport type AnalystSeverity = 'critical' | 'high' | 'medium' | 'low' | 'info'\n\n/** Data sources that candidate generation may intentionally learn from. */\nexport type ProposalFindingOrigin = 'search' | 'production'\n\n/** A finding explicitly admitted as candidate-generation input. */\nexport type ProposalFinding = AnalystFinding & {\n readonly proposal_origin: ProposalFindingOrigin\n}\n\nexport interface EvidenceRef {\n /**\n * Where the evidence lives. `span` and `event` refer to OTLP trace\n * elements; `artifact` to a file inside the run's artifact tree;\n * `finding` to another AnalystFinding (cross-analyst chaining);\n * `metric` to a named scalar reading the renderer knows how to read.\n */\n kind: 'span' | 'event' | 'artifact' | 'finding' | 'metric'\n uri: string\n excerpt?: string\n}\n\n// ── Analyst contract ─────────────────────────────────────────────────\n\n/**\n * The discriminator the registry uses to pass the right input.\n * `custom` is the escape hatch — analysts that need something else\n * (e.g. an embedding cache, a partner SDK handle) read it from\n * `AnalystRunInputs.custom[<analyst id>]`.\n */\nexport type AnalystInputKind =\n | 'trace-store'\n | 'artifact-dir'\n | 'run-record'\n | 'judge-input'\n | 'custom'\n\nexport interface AnalystCost {\n /** `deterministic` analysts MUST NOT call the LLM. */\n kind: 'deterministic' | 'llm'\n /** Optional declared upper bound; the registry can enforce a budget. */\n est_usd_per_run?: number\n /** Models the analyst expects to use (informational). */\n models?: string[]\n /** Maximum post-cancellation wait for provider usage. Model analysts default to 5 seconds. */\n settlement_timeout_ms?: number\n}\n\nexport interface AnalystRequirements {\n /** Min number of shots / samples the analyst needs to produce signal. */\n min_shots?: number\n /** Capabilities the runtime must supply (e.g. ['network', 'gpu']). */\n capabilities?: string[]\n}\n\n/**\n * What's passed to every analyst call. The registry resolves which\n * field the analyst's `inputKind` selects and asserts it's present.\n */\nexport interface AnalystRunInputs {\n traceStore?: TraceAnalysisStore\n artifactDir?: string\n runRecord?: RunRecord\n judgeInput?: JudgeInput\n /** Keyed by analyst id; populated by callers that registered custom analysts. */\n custom?: Record<string, unknown>\n}\n\nexport interface AnalystContext {\n runId: string\n /** Stable correlation id so logs from a single registry.run() share a tag. */\n correlationId: string\n /** Enforced wall-clock deadline (epoch ms). */\n deadlineMs?: number\n /** Per-analyst USD budget. Analysts MAY check before issuing LLM calls. */\n budgetUsd?: number\n /** Shared paid-call account when the analyst runs inside a larger campaign. */\n costLedger?: CostLedgerHandle\n /** Attribution phase used when writing to the shared paid-call account. */\n costPhase?: string\n /**\n * Shared chat client. Analysts that call an LLM go through this so\n * the operator picks transport (sandbox-sdk | router | cli-bridge |\n * direct-provider | mock) at the registry boundary without touching\n * analyst code.\n */\n chat?: ChatClient\n /**\n * Findings from a prior run the operator wants the analyst to see as\n * retrieval context. Kinds that take advantage of cross-run memory\n * (failure-mode \"I saw this cluster last run\", knowledge-gap \"the wiki\n * page I asked for is still missing\") render these into the actor's\n * working set. Filtering is the operator's job: pass the slice that\n * matches the analyst's id, or pass everything and let the kind\n * filter. Empty / absent means no cross-run context.\n */\n priorFindings?: ReadonlyArray<AnalystFinding>\n /**\n * Findings emitted by analysts that completed earlier in this registry run.\n * This is separate from `priorFindings`: upstream findings are dependency\n * context for the current pass, while prior findings are cross-run memory.\n * The registry populates this only when `RegistryRunOpts.chainFindings` is on.\n */\n upstreamFindings?: ReadonlyArray<AnalystFinding>\n /**\n * Report metered work independently of findings. This keeps an empty finding\n * set from erasing token/cost telemetry. Multiple receipts are accumulated.\n */\n recordUsage?: (receipt: AnalystUsageReceipt) => void\n /** Free-form runtime tags (env, host, op). Findings can echo these into metadata. */\n tags?: Record<string, string>\n /** Logger callback — analysts SHOULD prefer this over console.* for testability. */\n log?: (msg: string, fields?: Record<string, unknown>) => void\n /** Optional abort signal. Analysts SHOULD pass it through to LLM calls. */\n signal?: AbortSignal\n /**\n * Optional live-execution port. A runtime that owns a sandbox or checkout\n * fills it so an analyst can execute a bounded probe against the run's\n * produced state instead of reasoning about it from the trace alone. This\n * package defines only the port: no field here reaches for an agent loop,\n * and an absent probe means the analyst works from recorded evidence.\n */\n probe?: ExecutionProbe\n}\n\n// ── Live-execution port ─────────────────────────────────────────────\n\n/** One bounded command an analyst asks the probe to run. */\nexport interface ExecutionProbeRequest {\n command: string\n /** Working directory inside the probed environment. */\n cwd?: string\n /** Hard wall-clock deadline for this one execution. */\n timeoutMs: number\n /** Bytes of combined output retained; the prober truncates beyond it. */\n maxOutputBytes?: number\n signal?: AbortSignal\n}\n\n/**\n * Typed outcome of one probe execution. `succeeded: false` is a PROBE failure\n * (the environment could not run the command); a command that ran and exited\n * non-zero is a successful observation with a non-zero `exitCode`.\n */\nexport type ExecutionProbeOutcome =\n | {\n succeeded: true\n exitCode: number\n stdout: string\n stderr: string\n durationMs: number\n /** True when output was cut at `maxOutputBytes`. */\n truncated: boolean\n }\n | { succeeded: false; error: { class: string; message: string } }\n\n/**\n * The seam a runtime fills to let analysts observe produced state live.\n * Implementations own sandboxing, credentials, and cleanup; analysts only\n * submit bounded requests and read typed outcomes.\n */\nexport interface ExecutionProbe {\n /** One plain sentence naming what is being probed (e.g. a sandbox id). */\n readonly description: string\n execute(request: ExecutionProbeRequest): Promise<ExecutionProbeOutcome>\n}\n\n/**\n * The minimal contract. Concrete analysts can refine `TInput` so\n * implementations stay type-safe (e.g. a trace analyst's `TInput` is\n * `TraceAnalysisStore`); the registry passes the right field from\n * `AnalystRunInputs` based on `inputKind`.\n */\nexport interface Analyst<TInput = unknown> {\n /** Stable identifier — appears in finding_id, telemetry, and registry exclusion lists. */\n readonly id: string\n /** Human-readable. One sentence. */\n readonly description: string\n readonly inputKind: AnalystInputKind\n readonly cost: AnalystCost\n readonly requires?: AnalystRequirements\n /** Bump on breaking changes to claim wording or area so old finding_ids don't collide. */\n readonly version: string\n analyze(input: TInput, ctx: AnalystContext): Promise<AnalystFinding[]>\n}\n\n/** Metered work performed by one analyst call. */\nexport interface AnalystUsageReceipt {\n /** Number of model-usage records observed at the provider boundary. */\n calls: number | null\n /** Null when the provider did not return token accounting. */\n tokens: RunTokenUsage | null\n /** Observed, estimated, or explicitly uncaptured dollar cost. */\n cost: RunCostProvenance\n /** Known lower bound when one or more calls have uncaptured cost. */\n knownCostUsd?: number\n /**\n * Token counts the provider reported on only one side. Present exactly when\n * `tokens` is null and at least one side WAS reported: `RunTokenUsage` has no\n * nullable side, so a one-sided count cannot live in `tokens` without writing\n * a zero nobody measured. Read it as a lower bound, never as a total — the\n * field exists so a null `tokens` cannot hide a real count.\n */\n partialTokens?: { input: number | null; output: number | null }\n /**\n * True when the token counts were DERIVED by the transport (from character\n * lengths, say) rather than measured by the model provider. `cost.kind` is\n * `estimated` both for a rate estimate over exact tokens and for one over\n * derived tokens; this is the field that separates them.\n */\n tokensEstimated?: boolean\n}\n\n// ── finding_id stability ─────────────────────────────────────────────\n\n/**\n * Compute the stable finding_id from the identity-defining fields.\n * Default implementation hashes {analyst_id, area, subject, normalized claim}.\n * Analysts that emit findings whose claim text varies per run (timestamps,\n * counts) SHOULD either: (a) pass an explicit `id_basis` to fix the hash,\n * or (b) move the variable part into `rationale`/`metadata` and keep the\n * `claim` static.\n */\nexport function computeFindingId(input: {\n analyst_id: string\n area: string\n subject?: string\n claim: string\n /** Override the claim for hashing — use when the displayed claim has run-specific bits. */\n id_basis?: string\n}): string {\n const basis = JSON.stringify({\n a: input.analyst_id,\n r: input.area,\n s: input.subject ?? '',\n c: normalizeClaim(input.id_basis ?? input.claim),\n })\n return `f_${createHash('sha256').update(basis).digest('hex').slice(0, 20)}`\n}\n\nfunction normalizeClaim(c: string): string {\n // Lowercase, collapse whitespace, strip trailing punctuation. Goal:\n // \"Leaf X failed install\" and \"Leaf X failed install.\" hash the same.\n return c\n .toLowerCase()\n .replace(/\\s+/g, ' ')\n .replace(/[.!?;:,]+$/g, '')\n .trim()\n}\n\n/**\n * Convenience factory: produce a fully-formed AnalystFinding with the\n * id computed automatically. Analyst code stays terse.\n */\nexport function makeFinding(\n init: Omit<AnalystFinding, 'schema_version' | 'finding_id' | 'produced_at'> & {\n id_basis?: string\n produced_at?: string\n },\n): AnalystFinding {\n const { id_basis, produced_at, ...rest } = init\n return {\n schema_version: '1.0.0',\n finding_id: computeFindingId({\n analyst_id: rest.analyst_id,\n area: rest.area,\n subject: rest.subject,\n claim: rest.claim,\n id_basis,\n }),\n produced_at: produced_at ?? new Date().toISOString(),\n ...rest,\n }\n}\n\n/** Build a finding whose source is explicitly allowed during candidate generation. */\nexport function makeProposalFinding(\n init: Omit<ProposalFinding, 'schema_version' | 'finding_id' | 'produced_at'> & {\n id_basis?: string\n produced_at?: string\n },\n): ProposalFinding {\n const { proposal_origin, ...finding } = init\n return { ...makeFinding(finding), proposal_origin }\n}\n\n// ── Registry result envelope ────────────────────────────────────────\n\nexport interface AnalystRunSummary {\n analyst_id: string\n status: 'ok' | 'skipped' | 'failed'\n /** Why skipped — missing input, budget exceeded, capability unmet. */\n reason?: string\n findings_count: number\n latency_ms: number\n /** Additive model usage and cost provenance for this analyst. */\n usage: AnalystUsageReceipt\n /** When `status='failed'`: the error class + message, never the full stack. */\n error?: { class: string; message: string }\n}\n\nexport interface AnalystRunResult {\n run_id: string\n correlation_id: string\n started_at: string\n ended_at: string\n findings: AnalystFinding[]\n per_analyst: AnalystRunSummary[]\n /** Total LLM cost in USD across all analysts in this registry.run(). */\n total_cost_usd: number\n /**\n * Provenance for `total_cost_usd`. When uncaptured, the numeric field is only\n * the known subtotal and must not be treated as the run's total spend.\n */\n total_cost_provenance?: RunCostProvenance\n}\n\n// ── Streaming event envelope ────────────────────────────────────────\n\n/**\n * Events emitted by `AnalystRegistry.runStream(...)` in real time as\n * the registry executes. UIs subscribe via `for await (const ev of\n * registry.runStream(...))`; `registry.run(...)` is a thin collector\n * over the same stream, so the two surfaces share their invariants.\n *\n * Per-finding events are intentionally omitted — analyzers are batch\n * operations (a recursive engine returns the full `findings:json[]` at the\n * end of the responder), so streaming inside one analyst would only\n * emit partial JSON consumers can't render. The kind-completion event\n * is the right granularity; subscribers wanting per-finding rendering\n * iterate `event.findings` themselves.\n */\nexport type AnalystRunEvent =\n | {\n type: 'run-started'\n run_id: string\n correlation_id: string\n started_at: string\n /** The ordered list of analyst ids the registry will run. */\n analyst_ids: ReadonlyArray<string>\n }\n | {\n type: 'analyst-skipped'\n summary: AnalystRunSummary\n }\n | {\n type: 'analyst-started'\n analyst_id: string\n started_at: string\n }\n | {\n type: 'analyst-completed'\n /** `summary.status` is `'ok'` for clean completion or `'failed'` for thrown analysts. */\n summary: AnalystRunSummary\n findings: ReadonlyArray<AnalystFinding>\n }\n | {\n type: 'run-completed'\n result: AnalystRunResult\n }\n","import type { CostChannel, CostLedgerFilter, CostLedgerHandle } from '../cost-ledger'\nimport type { AnalystUsageReceipt } from './types'\n\nexport const DEFAULT_USAGE_SETTLEMENT_TIMEOUT_MS = 5_000\n\n/** Convert one ledger channel's complete call set into one analyst receipt. */\nexport function usageReceiptFromCostLedger(\n ledger: CostLedgerHandle,\n filter: CostChannel | CostLedgerFilter = 'analyst',\n): AnalystUsageReceipt {\n const resolvedFilter = typeof filter === 'string' ? { channel: filter } : filter\n const summary = ledger.summary(resolvedFilter)\n const receipts = ledger.list(resolvedFilter)\n const hasReasoningUsage = receipts.some((receipt) => receipt.reasoningTokens !== undefined)\n const hasCacheWriteUsage = receipts.some((receipt) => receipt.cacheWriteTokens !== undefined)\n const cost = summary.costProvenance\n return {\n calls: summary.totalCalls + summary.pendingCalls,\n tokens: summary.usageComplete\n ? {\n input: summary.inputTokens,\n output: summary.outputTokens,\n ...(hasReasoningUsage ? { reasoning: summary.reasoningTokens ?? 0 } : {}),\n ...(summary.cachedTokens > 0 ? { cached: summary.cachedTokens } : {}),\n ...(hasCacheWriteUsage ? { cacheWrite: summary.cacheWriteTokens ?? 0 } : {}),\n }\n : null,\n cost,\n ...(cost.kind === 'uncaptured' ? { knownCostUsd: summary.totalCostUsd } : {}),\n }\n}\n\nexport interface SettledUsageReceipt {\n settled: boolean\n pendingCalls: number\n receipt: AnalystUsageReceipt\n}\n\n/** Wait a bounded time for late provider receipts, then take one immutable snapshot. */\nexport async function settleUsageReceiptFromCostLedger(\n ledger: CostLedgerHandle,\n options: CostLedgerFilter & { timeoutMs?: number } = {},\n): Promise<SettledUsageReceipt> {\n const { timeoutMs: requestedTimeoutMs, ...requestedFilter } = options\n const filter: CostLedgerFilter = {\n channel: requestedFilter.channel ?? 'analyst',\n ...(requestedFilter.phase === undefined ? {} : { phase: requestedFilter.phase }),\n ...(requestedFilter.tags === undefined ? {} : { tags: requestedFilter.tags }),\n }\n const timeoutMs = validateUsageSettlementTimeout(requestedTimeoutMs)\n const initial = ledger.summary(filter)\n const waitResult =\n initial.pendingCalls === 0\n ? true\n : ledger.waitForIdle\n ? await ledger.waitForIdle({ timeoutMs })\n : false\n const pendingCalls = ledger.summary(filter).pendingCalls\n return {\n settled: waitResult && pendingCalls === 0,\n pendingCalls,\n receipt: usageReceiptFromCostLedger(ledger, filter),\n }\n}\n\nexport function validateUsageSettlementTimeout(timeoutMs?: number): number {\n const resolved = timeoutMs ?? DEFAULT_USAGE_SETTLEMENT_TIMEOUT_MS\n if (!Number.isSafeInteger(resolved) || resolved < 0 || resolved > 2_147_483_647) {\n throw new TypeError(\n 'settlementTimeoutMs must be a non-negative safe integer no greater than 2147483647',\n )\n }\n return resolved\n}\n\nexport function assertValidAnalystUsageReceipt(\n receipt: AnalystUsageReceipt,\n context = 'AnalystContext.recordUsage',\n): void {\n if (receipt.calls !== null && (!Number.isSafeInteger(receipt.calls) || receipt.calls < 0)) {\n throw new Error(`${context}: calls must be a non-negative safe integer or null`)\n }\n if (receipt.tokens) {\n assertNonNegativeSafeInteger(receipt.tokens.input, 'tokens.input', context)\n assertNonNegativeSafeInteger(receipt.tokens.output, 'tokens.output', context)\n if (receipt.tokens.reasoning !== undefined) {\n assertNonNegativeSafeInteger(receipt.tokens.reasoning, 'tokens.reasoning', context)\n if (receipt.tokens.reasoning > receipt.tokens.output) {\n throw new Error(`${context}: tokens.reasoning must not exceed tokens.output`)\n }\n }\n if (receipt.tokens.cached !== undefined) {\n assertNonNegativeSafeInteger(receipt.tokens.cached, 'tokens.cached', context)\n }\n if (receipt.tokens.cacheWrite !== undefined) {\n assertNonNegativeSafeInteger(receipt.tokens.cacheWrite, 'tokens.cacheWrite', context)\n }\n }\n if (receipt.cost.kind !== 'uncaptured') {\n assertNonNegativeFinite(receipt.cost.usd, 'cost.usd', context)\n } else if (receipt.cost.usd !== null) {\n throw new Error(`${context}: uncaptured cost.usd must be null`)\n }\n if (receipt.knownCostUsd !== undefined) {\n assertNonNegativeFinite(receipt.knownCostUsd, 'knownCostUsd', context)\n }\n if (receipt.partialTokens) {\n const { input, output } = receipt.partialTokens\n if (receipt.tokens) {\n throw new Error(`${context}: partialTokens must be absent when tokens is complete`)\n }\n if (input === null && output === null) {\n throw new Error(`${context}: partialTokens must carry at least one reported side`)\n }\n if (input !== null) assertNonNegativeSafeInteger(input, 'partialTokens.input', context)\n if (output !== null) assertNonNegativeSafeInteger(output, 'partialTokens.output', context)\n }\n}\n\nfunction assertNonNegativeSafeInteger(value: number, field: string, context: string): void {\n if (!Number.isSafeInteger(value) || value < 0) {\n throw new Error(`${context}: ${field} must be a non-negative safe integer`)\n }\n}\n\nfunction assertNonNegativeFinite(value: number, field: string, context: string): void {\n if (!Number.isFinite(value) || value < 0) {\n throw new Error(`${context}: ${field} must be a non-negative finite number`)\n }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;AAmSA,SAAgB,iBAAiB,OAOtB;CACT,MAAM,QAAQ,KAAK,UAAU;EAC3B,GAAG,MAAM;EACT,GAAG,MAAM;EACT,GAAG,MAAM,WAAW;EACpB,GAAG,eAAe,MAAM,YAAY,MAAM,KAAK;CACjD,CAAC;CACD,OAAO,KAAK,WAAW,QAAQ,CAAC,CAAC,OAAO,KAAK,CAAC,CAAC,OAAO,KAAK,CAAC,CAAC,MAAM,GAAG,EAAE;AAC1E;AAEA,SAAS,eAAe,GAAmB;CAGzC,OAAO,EACJ,YAAY,CAAC,CACb,QAAQ,QAAQ,GAAG,CAAC,CACpB,QAAQ,eAAe,EAAE,CAAC,CAC1B,KAAK;AACV;;;;;AAMA,SAAgB,YACd,MAIgB;CAChB,MAAM,EAAE,UAAU,aAAa,GAAG,SAAS;CAC3C,OAAO;EACL,gBAAgB;EAChB,YAAY,iBAAiB;GAC3B,YAAY,KAAK;GACjB,MAAM,KAAK;GACX,SAAS,KAAK;GACd,OAAO,KAAK;GACZ;EACF,CAAC;EACD,aAAa,gCAAe,IAAI,KAAK,EAAA,CAAE,YAAY;EACnD,GAAG;CACL;AACF;;AAGA,SAAgB,oBACd,MAIiB;CACjB,MAAM,EAAE,iBAAiB,GAAG,YAAY;CACxC,OAAO;EAAE,GAAG,YAAY,OAAO;EAAG;CAAgB;AACpD;;AC1VA,SAAgB,2BACd,QACA,SAAyC,WACpB;CACrB,MAAM,iBAAiB,OAAO,WAAW,WAAW,EAAE,SAAS,OAAO,IAAI;CAC1E,MAAM,UAAU,OAAO,QAAQ,cAAc;CAC7C,MAAM,WAAW,OAAO,KAAK,cAAc;CAC3C,MAAM,oBAAoB,SAAS,MAAM,YAAY,QAAQ,oBAAoB,KAAA,CAAS;CAC1F,MAAM,qBAAqB,SAAS,MAAM,YAAY,QAAQ,qBAAqB,KAAA,CAAS;CAC5F,MAAM,OAAO,QAAQ;CACrB,OAAO;EACL,OAAO,QAAQ,aAAa,QAAQ;EACpC,QAAQ,QAAQ,gBACZ;GACE,OAAO,QAAQ;GACf,QAAQ,QAAQ;GAChB,GAAI,oBAAoB,EAAE,WAAW,QAAQ,mBAAmB,EAAE,IAAI,CAAC;GACvE,GAAI,QAAQ,eAAe,IAAI,EAAE,QAAQ,QAAQ,aAAa,IAAI,CAAC;GACnE,GAAI,qBAAqB,EAAE,YAAY,QAAQ,oBAAoB,EAAE,IAAI,CAAC;EAC5E,IACA;EACJ;EACA,GAAI,KAAK,SAAS,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;CAC7E;AACF;;AASA,eAAsB,iCACpB,QACA,UAAqD,CAAC,GACxB;CAC9B,MAAM,EAAE,WAAW,oBAAoB,GAAG,oBAAoB;CAC9D,MAAM,SAA2B;EAC/B,SAAS,gBAAgB,WAAW;EACpC,GAAI,gBAAgB,UAAU,KAAA,IAAY,CAAC,IAAI,EAAE,OAAO,gBAAgB,MAAM;EAC9E,GAAI,gBAAgB,SAAS,KAAA,IAAY,CAAC,IAAI,EAAE,MAAM,gBAAgB,KAAK;CAC7E;CACA,MAAM,YAAY,+BAA+B,kBAAkB;CAEnE,MAAM,aADU,OAAO,QAAQ,MAEvB,CAAC,CAAC,iBAAiB,IACrB,OACA,OAAO,cACL,MAAM,OAAO,YAAY,EAAE,UAAU,CAAC,IACtC;CACR,MAAM,eAAe,OAAO,QAAQ,MAAM,CAAC,CAAC;CAC5C,OAAO;EACL,SAAS,cAAc,iBAAiB;EACxC;EACA,SAAS,2BAA2B,QAAQ,MAAM;CACpD;AACF;AAEA,SAAgB,+BAA+B,WAA4B;CACzE,MAAM,WAAW,aAAA;CACjB,IAAI,CAAC,OAAO,cAAc,QAAQ,KAAK,WAAW,KAAK,WAAW,YAChE,MAAM,IAAI,UACR,oFACF;CAEF,OAAO;AACT;AAEA,SAAgB,+BACd,SACA,UAAU,8BACJ;CACN,IAAI,QAAQ,UAAU,SAAS,CAAC,OAAO,cAAc,QAAQ,KAAK,KAAK,QAAQ,QAAQ,IACrF,MAAM,IAAI,MAAM,GAAG,QAAQ,oDAAoD;CAEjF,IAAI,QAAQ,QAAQ;EAClB,6BAA6B,QAAQ,OAAO,OAAO,gBAAgB,OAAO;EAC1E,6BAA6B,QAAQ,OAAO,QAAQ,iBAAiB,OAAO;EAC5E,IAAI,QAAQ,OAAO,cAAc,KAAA,GAAW;GAC1C,6BAA6B,QAAQ,OAAO,WAAW,oBAAoB,OAAO;GAClF,IAAI,QAAQ,OAAO,YAAY,QAAQ,OAAO,QAC5C,MAAM,IAAI,MAAM,GAAG,QAAQ,iDAAiD;EAEhF;EACA,IAAI,QAAQ,OAAO,WAAW,KAAA,GAC5B,6BAA6B,QAAQ,OAAO,QAAQ,iBAAiB,OAAO;EAE9E,IAAI,QAAQ,OAAO,eAAe,KAAA,GAChC,6BAA6B,QAAQ,OAAO,YAAY,qBAAqB,OAAO;CAExF;CACA,IAAI,QAAQ,KAAK,SAAS,cACxB,wBAAwB,QAAQ,KAAK,KAAK,YAAY,OAAO;MACxD,IAAI,QAAQ,KAAK,QAAQ,MAC9B,MAAM,IAAI,MAAM,GAAG,QAAQ,mCAAmC;CAEhE,IAAI,QAAQ,iBAAiB,KAAA,GAC3B,wBAAwB,QAAQ,cAAc,gBAAgB,OAAO;CAEvE,IAAI,QAAQ,eAAe;EACzB,MAAM,EAAE,OAAO,WAAW,QAAQ;EAClC,IAAI,QAAQ,QACV,MAAM,IAAI,MAAM,GAAG,QAAQ,uDAAuD;EAEpF,IAAI,UAAU,QAAQ,WAAW,MAC/B,MAAM,IAAI,MAAM,GAAG,QAAQ,sDAAsD;EAEnF,IAAI,UAAU,MAAM,6BAA6B,OAAO,uBAAuB,OAAO;EACtF,IAAI,WAAW,MAAM,6BAA6B,QAAQ,wBAAwB,OAAO;CAC3F;AACF;AAEA,SAAS,6BAA6B,OAAe,OAAe,SAAuB;CACzF,IAAI,CAAC,OAAO,cAAc,KAAK,KAAK,QAAQ,GAC1C,MAAM,IAAI,MAAM,GAAG,QAAQ,IAAI,MAAM,qCAAqC;AAE9E;AAEA,SAAS,wBAAwB,OAAe,OAAe,SAAuB;CACpF,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,QAAQ,GACrC,MAAM,IAAI,MAAM,GAAG,QAAQ,IAAI,MAAM,sCAAsC;AAE/E"}