npm - @tangle-network/agent-eval - Versions diffs - 0.202.2 → 0.203.1 - Mend

@tangle-network/agent-eval 0.202.2 → 0.203.1

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 (172) hide show

package/dist/{produced-state-CR2tAPLf.js.map → produced-state-g9BPgu4D.js.map} RENAMED Viewed

	@@ -1 +1 @@
1	- {"version":3,"file":"produced-state-CR2tAPLf.js","names":[],"sources":["../src/agent-profile.ts","../src/completion-verifier.ts","../src/produced-state.ts"],"sourcesContent":["import { createHash } from 'node:crypto'\nimport type { AgentProfile } from '@tangle-network/agent-interface'\nimport { type HarnessType, harnessSupportsModel } from '@tangle-network/agent-interface'\nimport { ValidationError } from './errors'\nimport { canonicalString } from './ledger-core/canonical'\n\nexport type { AgentProfile, HarnessType } from '@tangle-network/agent-interface'\n\n/*\n The agentic coding harnesses an eval sweeps by default — the ones we care about\n * ranking. This is the SINGLE source of that list; consumers import it instead of\n * re-declaring their own (a re-declared list is how the fleet drifts). Pass an\n * explicit `harnesses` (e.g. `harnessTypeSchema.options` for literally every known\n * harness) to widen beyond these.\n /\nexport const CODING_HARNESSES: readonly HarnessType[] = [\n 'opencode',\n 'claude-code',\n 'codex',\n 'kimi-code',\n]\n\nexport interface ProfileAxisSpec {\n /* The domain profile to sweep. Its prompt/tools/skills are held fixed; only the\n * harness and model vary. `model.default` is the fallback model. /\n base: AgentProfile\n /* Harnesses to cross. Default: {@link CODING_HARNESSES}. /\n harnesses?: readonly HarnessType[]\n /* Models to cross. Default: `[base.model.default]` — one model, i.e. today's\n * single-model behaviour, so omitting this never changes an existing run. /\n models?: readonly string[]\n /* Force every (harness, model) pair verbatim, even ones the harness can't run —\n * for deliberately testing failure modes. Default (false): SNAP instead — a\n * vendor-locked harness runs only the swept models in its family, or its native\n * default when it supports none, so no harness is dropped and none gets a\n * guaranteed-failing foreign-model cell. /\n keepIncompatible?: boolean\n}\n\n/* Model sentinel for a vendor-locked harness that supports none of the swept models:\n * it carries no provider prefix, so `harnessSupportsModel` accepts it and the harness\n * resolves it to its own native default model at runtime (e.g. kimi-code → its Kimi\n * model). Lets `expandProfileAxes` snap-instead-of-drop without a per-harness flagship\n * table that would rot as router catalogs change. /\nexport const HARNESS_NATIVE_MODEL = 'default'\n\n/\n Expand a base profile across the harness × model matrix into the `AgentProfile[]`\n * that `runProfileMatrix` / `selfImprove` score — the ONE place \"which harnesses ×\n * which models do we evaluate\" lives, so no product hand-rolls its own harness list\n * or column→profile mapping (the pattern that let those copies drift and silently\n * break the harness pivot).\n \n Each cell clones `base`, sets the canonical top-level `harness` and `model.default`,\n * and stamps `metadata.harness` + `metadata.harnessModel` for matrix grouping. Both\n * metadata fields are hash-bearing, so every cell gets a distinct `agentProfileId` row\n * and results join back by harness/model via {@link harnessAxisOf} with no\n * hand-recomputed key. A vendor-locked harness snaps to its family's swept models — or\n * its native default ({@link HARNESS_NATIVE_MODEL}) when it supports none — so every\n * requested harness runs; `keepIncompatible` forces every pair verbatim.\n \n Omit `harnesses`/`models` to sweep the full default set — the \"turn it on for\n * everything we care about\" switch, identical in shape whether one harness or all.\n /\nexport function expandProfileAxes(spec: ProfileAxisSpec): AgentProfile[] {\n const harnesses = spec.harnesses ?? CODING_HARNESSES\n if (harnesses.length === 0) throw new ValidationError('expandProfileAxes: no harnesses to sweep')\n const baseModel = spec.base.model?.default\n const models = spec.models ?? (baseModel ? [baseModel] : [])\n if (models.length === 0) {\n throw new ValidationError(\n 'expandProfileAxes: no models to sweep — base profile has no model.default and none were supplied',\n )\n }\n const out: AgentProfile[] = []\n const seen = new Set<string>()\n for (const harness of harnesses) {\n // A universal (router-backed) harness — opencode/pi/claudish — runs every swept\n // model. A vendor-locked harness — codex/claude-code/kimi-code — runs only the\n // swept models in its own family; when it supports NONE of them it snaps to its\n // native default (the `HARNESS_NATIVE_MODEL` sentinel it resolves at runtime)\n // rather than being dropped, so every requested harness still appears in the\n // sweep on a model it can actually run — e.g. sweeping `deepseek/x` puts opencode\n // on deepseek and kimi-code on its own Kimi model, a real head-to-head.\n // `keepIncompatible` forces every (harness, model) pair verbatim (failure-mode runs).\n const supported = spec.keepIncompatible\n ? models\n : models.filter((model) => harnessSupportsModel(harness, model))\n const effective = supported.length > 0 ? supported : [HARNESS_NATIVE_MODEL]\n for (const model of effective) {\n const profile: AgentProfile = {\n ...spec.base,\n harness,\n name: `${spec.base.name ?? 'agent'}/${harness}/${model}`,\n model: { ...spec.base.model, default: model },\n metadata: { ...(spec.base.metadata ?? {}), harness, harnessModel: model },\n }\n const id = agentProfileId(profile)\n if (seen.has(id)) continue\n seen.add(id)\n out.push(profile)\n }\n }\n if (out.length === 0) {\n // Unreachable in normal use — snapping guarantees ≥1 cell per harness — but keep a\n // fail-closed guard so a future refactor can't silently produce an empty sweep.\n throw new ValidationError(\n `expandProfileAxes: produced no profiles (harnesses=[${harnesses.join(', ')}], models=[${models.join(', ')}]).`,\n )\n }\n return out\n}\n\n/\n Read the (harness, model) a matrix cell ran under, off a profile or a result row's\n * profile — the join-back for a `byHarness` pivot. Returns undefined when the profile\n * wasn't produced by {@link expandProfileAxes}. Callers group `result.byProfile` by\n * this instead of recomputing an id (recomputing the wrong key is what broke the pivot\n * in the hand-rolled copies).\n /\nexport function harnessAxisOf(\n profile: Pick<AgentProfile, 'metadata'>,\n): { harness: HarnessType; model: string } \| undefined {\n const m = profile.metadata as Record<string, unknown> \| undefined\n const harness = m?.harness\n const model = m?.harnessModel\n if (typeof harness === 'string' && typeof model === 'string') {\n return { harness: harness as HarnessType, model }\n }\n return undefined\n}\n\n/\n Collision-resistant, path-safe, human-readable profile id for eval artifacts.\n * Scorecard joins still use `agentProfileHash`; this id is for run ids, matrix\n * keys, and directory names where two profiles must not collapse onto one row.\n * The suffix is the first 64 bits of the behaviour hash, enough for ordinary\n * eval matrices while keeping filenames readable.\n /\nexport function agentProfileId(profile: AgentProfile): string {\n const label = pathSafeProfileLabel(agentProfileDisplayLabel(profile)) ?? 'profile'\n return `${label}-${agentProfileHash(profile).slice(0, 16)}`\n}\n\n/\n Model snapshot used for `RunRecord.model`. Eval surfaces require a concrete\n * model id because run records reject bare/missing model aliases.\n /\nexport function agentProfileModelId(profile: AgentProfile): string {\n const model = profile.model?.default?.trim()\n if (!model) {\n const label = agentProfileDisplayLabel(profile) ?? 'unnamed profile'\n throw new ValidationError(\n `AgentProfile \"${label}\" has no model.default — cannot record eval run`,\n )\n }\n return model\n}\n\nfunction agentProfileDisplayLabel(profile: AgentProfile): string \| undefined {\n return profile.name?.trim() \|\| profile.version?.trim() \|\| undefined\n}\n\nfunction pathSafeProfileLabel(label: string \| undefined): string \| undefined {\n const safe = label\n ?.trim()\n .replace(/[^A-Za-z0-9._-]+/g, '-')\n .replace(/-+/g, '-')\n .replace(/^-\|-$/g, '')\n return safe \|\| undefined\n}\n\nfunction compact<T extends Record<string, unknown>>(input: T): Partial<T> {\n const out: Record<string, unknown> = {}\n for (const [key, value] of Object.entries(input)) {\n if (value !== undefined) out[key] = value\n }\n return out as Partial<T>\n}\n\n/\n Deterministic behaviour identity for the canonical\n * `@tangle-network/agent-interface` AgentProfile.\n \n `name` and `description` are labels and do not affect the hash. Profile\n * `version`, prompt, model hints, tools, resources, hooks, modes, permissions,\n * and extensions do affect the hash. Resource array order is hash-bearing\n * because mount order can change agent behaviour. Undefined fields are treated\n * as absent; explicit `null` fields remain hash-bearing.\n /\nexport function agentProfileHash(profile: AgentProfile): string {\n const model = agentProfileModelId(profile)\n const behaviour = compact({\n ...profile,\n name: undefined,\n description: undefined,\n tags: profile.tags ? [...profile.tags].sort() : undefined,\n model: compact({ ...profile.model, default: model }),\n })\n return createHash('sha256').update(canonicalString(behaviour)).digest('hex')\n}\n","/\n Completion verifier — the task-completion oracle.\n \n Answers the only eval question that is not a proxy: did the agent actually\n * COMPLETE the task — produce every required deliverable, persisted and\n * correct — rather than describe what should be done. A fluent transcript\n * that never produces the artifact scores zero here.\n \n Per requirement, a two-stage check:\n * 1. Structural — a produced item (vault artifact / approved proposal /\n * tool call) of the right kind is matched against the requirement and\n * carries non-empty content. Deterministic; no LLM.\n * 2. Correctness — only if structurally present AND the matched item\n * carries content, one targeted check decides whether that item\n * actually fulfils the requirement. A hallucinated artifact fails here;\n * an absent one already failed stage 1.\n \n `completionRate` is satisfied / MEASURABLE requirements (unmeasured rows —\n * checker failures — are excluded from the denominator, never scored as\n * zeros). Quality dimensions are meaningless on an incomplete task — callers\n * gate on `fullyComplete` / `completionRate` before scoring quality.\n /\n\nimport { randomUUID } from 'node:crypto'\nimport type { ChatClient, ChatRequest } from './analyst/chat-client'\nimport type { Artifact } from './artifact-validator'\nimport { CostLedger, type CostLedgerHandle } from './cost-ledger'\nimport { assertServedModel, ModelSubstitutionError } from './integrity/served-model'\nimport { recoverTruncatedJson } from './json-recovery'\nimport { JudgeParseError } from './judges'\nimport {\n costReceiptFromLlm,\n costReceiptFromLlmError,\n maximumChargeForLlmRequest,\n} from './llm-client'\nimport { packageVersion } from './package-version'\nimport type { RawProviderEvent, RawProviderSink } from './trace/raw-provider-sink'\nimport {\n certificationEvidenceDigest,\n type DefaultVerdict,\n type VerdictCertification,\n} from './verdict'\nimport type { CheckerIdentity, VerificationStrategySource } from './verification-strategy'\n\n/* What kind of produced state can satisfy a requirement structurally. /\nexport type SatisfiedBy = 'artifact' \| 'proposal' \| 'tool-call' \| 'any'\n\nexport interface CompletionRequirement {\n /* Stable id from the task gold (e.g. a persona's `expected_requirements[].req_id`). /\n reqId: string\n /* Human-readable description of the required deliverable. /\n title: string\n /* Optional kind/category hint, matched against a produced item's kind. /\n category?: string\n /* What produced state satisfies this requirement. Defaults to 'any'. /\n satisfiedBy?: SatisfiedBy\n}\n\nexport interface TaskGold {\n taskId: string\n requirements: CompletionRequirement[]\n}\n\nexport interface ProducedProposal {\n id: string\n title: string\n status: 'pending' \| 'approved' \| 'rejected'\n /* Optional persisted body — when present, enables a correctness check. /\n content?: string\n}\n\n/* Everything observable about what a run actually produced. /\nexport interface ProducedState {\n /* Persisted vault artifacts. Reuses the shared `Artifact` shape. /\n artifacts: Artifact[]\n /* Proposals / filings the agent created. /\n proposals: ProducedProposal[]\n /* Names of tools the agent invoked. /\n toolCalls: string[]\n}\n\nexport interface RequirementCheck {\n reqId: string\n title: string\n /* A produced item of the right kind matched the requirement, non-empty. /\n structurallyPresent: boolean\n /\n Whether the matched item actually fulfils the requirement. `null` when\n * not structurally present, when the matched item carries no content\n * to assess, or when the correctness check itself failed (`unmeasured`).\n /\n correct: boolean \| null\n /* structurallyPresent && !unmeasured && correct !== false. /\n satisfied: boolean\n /\n Set when the correctness check itself errored (LLM call failure or an\n * unparseable response after retry). The requirement's fulfilment is\n * UNKNOWN — `correct` stays null, `satisfied` is false, and\n * `completionVerdict` excludes the row from `completionRate`'s\n * denominator. Never folded into a zero: a synthetic zero is\n * indistinguishable from a real failure (see `JudgeParseError`).\n /\n unmeasured?: true\n /* Why the correctness check could not be measured (present iff `unmeasured`). /\n unmeasuredReason?: string\n /* Human-readable evidence for the verdict. /\n evidence: string[]\n}\n\n/* Extends the substrate verdict spine: `valid` = `fullyComplete` and\n * `score` = `completionRate` — derived in `completionVerdict()`, the one\n * place those equalities hold by construction. /\nexport interface CompletionVerdict extends DefaultVerdict {\n taskId: string\n requirements: RequirementCheck[]\n /* satisfied / MEASURABLE requirements (unmeasured rows leave the denominator). /\n completionRate: number\n /* Every measurable requirement satisfied (false when anything is unmeasured). /\n fullyComplete: boolean\n /* Requirements whose correctness check errored — reported, never scored as zero. /\n unmeasuredCount: number\n}\n\n/\n Construct a `CompletionVerdict` from the per-requirement checks, deriving\n * `completionRate` / `fullyComplete` and the spine fields (`valid` =\n * `fullyComplete`, `score` = `completionRate`) in one place. Throws on zero\n * requirements — a verdict over nothing is a misconfiguration, mirroring\n * `verifyCompletion`'s gold-spec guard.\n /\nexport function completionVerdict(input: {\n taskId: string\n requirements: RequirementCheck[]\n /* What certified the correctness stage, when anything did. Omitted =\n * an uncertified verdict — the honest default for a bare checker. /\n certification?: VerdictCertification\n}): CompletionVerdict {\n if (input.requirements.length === 0) {\n throw new Error(\n `completionVerdict: task '${input.taskId}' has no requirement checks — nothing to derive a verdict from`,\n )\n }\n const measurable = input.requirements.filter((r) => !r.unmeasured)\n const unmeasuredCount = input.requirements.length - measurable.length\n if (measurable.length === 0) {\n // Every check errored: this is an infrastructure failure, not a scored\n // run. A 0-rate verdict here would be a fabricated measurement.\n throw new Error(\n `completionVerdict: task '${input.taskId}' has no measurable requirements — all ${input.requirements.length} correctness checks failed (${input.requirements[0]?.unmeasuredReason ?? 'unknown reason'})`,\n )\n }\n const satisfiedCount = measurable.filter((r) => r.satisfied).length\n const completionRate = satisfiedCount / measurable.length\n // A run with unmeasured rows can still report a rate over what WAS\n // measured, but must not claim full completion.\n const fullyComplete = unmeasuredCount === 0 && satisfiedCount === measurable.length\n return {\n taskId: input.taskId,\n requirements: input.requirements,\n completionRate,\n fullyComplete,\n unmeasuredCount,\n valid: fullyComplete,\n score: completionRate,\n ...(input.certification === undefined ? {} : { certification: input.certification }),\n }\n}\n\n/\n What a correctness checker declares about itself so the completion\n * verdict can carry a certification: which strategy member it discharges,\n * its exact identity, and the steps its answers rest on unverified.\n /\nexport interface CorrectnessCheckerAttestation {\n strategy: VerificationStrategySource\n checker: CheckerIdentity\n assumptions: string[]\n}\n\n/\n Decides whether a produced item's content actually fulfils a requirement.\n * Injected so the structural verifier stays pure and unit-testable; the\n * production implementation is `createLlmCorrectnessChecker`.\n \n `attestation` is optional metadata on the function value: a checker that\n * carries one yields certified completion verdicts; a bare function yields\n * the same verdict uncertified. A plain arrow function remains a valid\n * checker.\n /\nexport interface CorrectnessChecker {\n (\n requirement: CompletionRequirement,\n content: string,\n ): Promise<{ correct: boolean; reason: string }>\n attestation?: CorrectnessCheckerAttestation\n}\n\nconst STOPWORDS = new Set([\n 'the',\n 'a',\n 'an',\n 'of',\n 'for',\n 'and',\n 'or',\n 'to',\n 'in',\n 'on',\n 'with',\n 'by',\n])\n\n// Deliverable-FORM vocabulary — words that name the SHAPE of an output, not its\n// domain content. A correct \"swap-comparison view persisted as a ui/* artifact\"\n// is an OpenUI JSON whose body says nothing about \"artifact\" / \"persisted\" /\n// \"view\"; the discriminative tokens are the domain nouns (swap, comparison).\n// Stripped from the REQUIREMENT side of structural recall so a deliverable is\n// matched on what it IS about, not on the boilerplate describing its form. The\n// correctness checker strips the same class via TITLE_STOPWORDS. Anti-game holds:\n// the distinctive domain tokens remain, so an off-topic item still fails.\nconst REQUIREMENT_FORM_STOPWORDS = new Set([\n 'generated',\n 'generate',\n 'view',\n 'render',\n 'rendered',\n 'persisted',\n 'persist',\n 'artifact',\n 'file',\n 'document',\n 'note',\n 'proposal',\n 'deliverable',\n 'output',\n 'created',\n 'create',\n 'produce',\n 'produced',\n 'flag',\n])\n\nconst MATCH_THRESHOLD = 0.5\nconst MIN_CONTENT_CHARS = 50\n\nfunction tokens(s: string, extraStop?: Set<string>): Set<string> {\n return new Set(\n s\n .toLowerCase()\n .split(/[^a-z0-9]+/)\n .filter((t) => t.length > 1 && !STOPWORDS.has(t) && !extraStop?.has(t)),\n )\n}\n\n/*\n Recall of the requirement's tokens within a candidate's identifying text.\n * Recall, not Jaccard — a candidate's path/id legitimately carries extra\n * tokens the requirement does not name. The requirement side drops\n * deliverable-FORM vocabulary so recall keys on the distinctive domain tokens.\n /\nfunction tokenRecall(requirementText: string, candidateText: string): number {\n const req = tokens(requirementText, REQUIREMENT_FORM_STOPWORDS)\n if (req.size === 0) return 0\n const cand = tokens(candidateText)\n let hit = 0\n for (const t of req) if (cand.has(t)) hit++\n return hit / req.size\n}\n\ninterface Candidate {\n reqIndex: number\n /* Unique key for a produced item — each item satisfies at most one requirement. /\n itemKey: string\n score: number\n evidence: string\n /* Content to correctness-check, or null when the matched item has none. /\n content: string \| null\n}\n\nfunction artifactCandidates(\n req: CompletionRequirement,\n reqIndex: number,\n artifacts: Artifact[],\n): Candidate[] {\n const reqText = `${req.title} ${req.category ?? ''}`\n const out: Candidate[] = []\n artifacts.forEach((a, i) => {\n if ((a.content ?? '').trim().length < MIN_CONTENT_CHARS) return\n // Match against the artifact CONTENT too, not just its path + kind — a\n // generated view / note whose path is generic still satisfies a requirement\n // when its body covers it (e.g. an OpenUI comparison grounded in the on-file\n // figures). Bounded slice keeps the recall text cheap; MATCH_THRESHOLD holds.\n let score = tokenRecall(\n reqText,\n `${a.path ?? ''} ${a.kind} ${(a.content ?? '').slice(0, 4000)}`,\n )\n if (req.category && a.kind && req.category.toLowerCase() === a.kind.toLowerCase()) {\n score = Math.max(score, 1)\n }\n if (score < MATCH_THRESHOLD) return\n out.push({\n reqIndex,\n itemKey: `artifact:${i}`,\n score,\n evidence: `artifact '${a.path ?? a.kind}' matched (token recall ${score.toFixed(2)})`,\n content: a.content ?? null,\n })\n })\n return out\n}\n\nfunction proposalCandidates(\n req: CompletionRequirement,\n reqIndex: number,\n proposals: ProducedProposal[],\n): Candidate[] {\n const reqText = `${req.title} ${req.category ?? ''}`\n const out: Candidate[] = []\n for (const p of proposals) {\n // Pending or rejected work is not a completed deliverable.\n if (p.status !== 'approved') continue\n // A proposal needs an assessable BODY to be a deliverable. A bare title is\n // not completion: correctness cannot be judged on it, so a title-only match\n // would auto-pass the oracle (structurallyPresent && correct===null →\n // satisfied) with no verifiable content. Tool calls are the only\n // legitimately content-less deliverable (`toolCallCandidates`).\n const body = (p.content ?? '').trim()\n if (body.length < MIN_CONTENT_CHARS) continue\n // Match against the body as well as the (often short) title — a refusal /\n // flag / analysis proposal whose title is a label still satisfies a\n // descriptively-worded requirement when its content covers it. MATCH_THRESHOLD\n // + the requirement's distinctive tokens keep an off-topic proposal out;\n // correctness (a SEMANTIC checker, NOT this lexical pass) then judges\n // polarity/fulfilment, so a negation that merely contains the tokens fails.\n // Structural and correctness must use different evidence or the two-stage\n // check collapses to one lexical gate.\n const score = tokenRecall(reqText, `${p.title} ${body}`)\n if (score < MATCH_THRESHOLD) continue\n out.push({\n reqIndex,\n itemKey: `proposal:${p.id}`,\n score,\n evidence: `approved proposal '${p.title}' matched (token recall ${score.toFixed(2)})`,\n content: body,\n })\n }\n return out\n}\n\nfunction toolCallCandidates(\n req: CompletionRequirement,\n reqIndex: number,\n toolCalls: string[],\n): Candidate[] {\n const out: Candidate[] = []\n toolCalls.forEach((name, i) => {\n const score = tokenRecall(req.title, name)\n if (score < MATCH_THRESHOLD) return\n out.push({\n reqIndex,\n itemKey: `tool:${i}`,\n score,\n evidence: `tool call '${name}' matched (token recall ${score.toFixed(2)})`,\n content: null,\n })\n })\n return out\n}\n\n/\n Verify whether a run completed the task. `checkCorrectness` is injected —\n * `createLlmCorrectnessChecker` for production, a deterministic stub in tests.\n \n Throws on a gold spec with no requirements: an eval task that requires\n * nothing is a misconfiguration, not a vacuously-complete task.\n /\nexport async function verifyCompletion(\n gold: TaskGold,\n state: ProducedState,\n checkCorrectness: CorrectnessChecker,\n): Promise<CompletionVerdict> {\n if (gold.requirements.length === 0) {\n throw new Error(\n `verifyCompletion: task '${gold.taskId}' has no requirements — malformed gold spec`,\n )\n }\n\n // Collect every above-threshold (requirement, produced-item) candidate, then\n // assign greedily by descending score: each requirement and each produced\n // item is used at most once. One deliverable fulfils one requirement.\n const candidates: Candidate[] = []\n gold.requirements.forEach((req, i) => {\n const by = req.satisfiedBy ?? 'any'\n if (by === 'artifact' \|\| by === 'any') {\n candidates.push(...artifactCandidates(req, i, state.artifacts))\n }\n if (by === 'proposal' \|\| by === 'any') {\n candidates.push(...proposalCandidates(req, i, state.proposals))\n }\n if (by === 'tool-call' \|\| by === 'any') {\n candidates.push(...toolCallCandidates(req, i, state.toolCalls))\n }\n })\n candidates.sort((a, b) => b.score - a.score)\n\n const assigned = new Map<number, Candidate>()\n const itemTaken = new Set<string>()\n for (const c of candidates) {\n if (assigned.has(c.reqIndex) \|\| itemTaken.has(c.itemKey)) continue\n assigned.set(c.reqIndex, c)\n itemTaken.add(c.itemKey)\n }\n\n const requirements: RequirementCheck[] = []\n for (let i = 0; i < gold.requirements.length; i++) {\n const req = gold.requirements[i]!\n const match = assigned.get(i)\n const evidence: string[] = []\n let correct: boolean \| null = null\n let unmeasuredReason: string \| undefined\n\n if (match) {\n evidence.push(match.evidence)\n if (match.content !== null) {\n try {\n const r = await checkCorrectness(req, match.content)\n correct = r.correct\n evidence.push(`correctness: ${r.correct ? 'pass' : 'fail'} — ${r.reason}`)\n } catch (err) {\n // The CHECKER failed, not the requirement. Recording this as a\n // zero would fabricate a model failure out of an infrastructure\n // one; the requirement is unmeasured and leaves the denominator.\n unmeasuredReason =\n err instanceof JudgeParseError\n ? `checker response unparseable after retry: ${err.raw.slice(0, 200)}`\n : `checker call failed: ${err instanceof Error ? err.message : String(err)}`\n evidence.push(`correctness: UNMEASURED — ${unmeasuredReason}`)\n }\n } else {\n evidence.push('correctness: not assessed — matched item carries no content')\n }\n } else {\n const by = req.satisfiedBy ?? 'any'\n const kind = by === 'any' ? 'artifact/proposal/tool-call' : by\n evidence.push(`no produced ${kind} matched this requirement`)\n }\n\n const structurallyPresent = match !== undefined\n const unmeasured = unmeasuredReason !== undefined\n const satisfied = structurallyPresent && !unmeasured && correct !== false\n requirements.push({\n reqId: req.reqId,\n title: req.title,\n structurallyPresent,\n correct,\n satisfied,\n ...(unmeasured ? { unmeasured: true as const, unmeasuredReason } : {}),\n evidence,\n })\n }\n\n // The verdict blends the deterministic structural stage with the injected\n // correctness stage, so the certification names the checker's own member\n // and carries the structural stage's lexical nature as an assumption. A\n // checker without an attestation yields an uncertified verdict.\n const attestation = checkCorrectness.attestation\n const certification: VerdictCertification \| undefined = attestation\n ? {\n strategy: attestation.strategy,\n checker: attestation.checker,\n assumptions: [\n 'structural matching is lexical token recall over produced items — the correctness stage only sees items it matched',\n ...attestation.assumptions,\n ],\n evidenceDigest: certificationEvidenceDigest({ taskId: gold.taskId, requirements }),\n }\n : undefined\n\n return completionVerdict({\n taskId: gold.taskId,\n requirements,\n ...(certification === undefined ? {} : { certification }),\n })\n}\n\nexport interface LlmCorrectnessCheckerOpts {\n model?: string\n /* Optional ledger for direct use. /\n costLedger?: CostLedgerHandle\n costPhase?: string\n costTags?: Record<string, string>\n signal?: AbortSignal\n /* Max chars of artifact content sent to the checker. /\n maxContentChars?: number\n /\n Checker LLM calls per requirement before giving up (parse failures and\n * call errors both consume attempts). The failure then surfaces as an\n * `unmeasured` requirement, never a zero.\n /\n maxAttempts?: number\n /\n Forensic capture of every checker request/response/error — without it a\n * checker failure is unauditable (the agent-turn raws never contain the\n * checker's own calls). Same sink contract as `LlmClient`.\n /\n rawSink?: RawProviderSink\n}\n\n/\n Parse the correctness checker's model response. Tolerates a response\n * truncated mid-JSON (max_tokens cap) by auto-closing the prefix — the\n * verdict boolean usually lands in the first few tokens, so a recovered\n * prefix with a boolean `correct` is a real measurement, not a guess.\n * Fails loud (JudgeParseError) when no boolean verdict is recoverable.\n /\nexport function parseCorrectnessResponse(raw: string): { correct: boolean; reason: string } {\n const readVerdict = (candidate: unknown): { correct: boolean; reason: string } \| null => {\n if (candidate === null \|\| typeof candidate !== 'object') return null\n const { correct, reason } = candidate as { correct?: unknown; reason?: unknown }\n if (typeof correct !== 'boolean') return null\n return { correct, reason: typeof reason === 'string' ? reason : '' }\n }\n\n const match = raw.match(/\\{[\\s\\S]\\}/)\n if (match) {\n try {\n const strict = readVerdict(JSON.parse(match[0]))\n if (strict) return strict\n } catch {\n // fall through to truncation recovery\n }\n }\n // The strict path needs a closing `}`; a cap-hit response has none. Take\n // everything from the first `{` and auto-close it.\n const start = raw.indexOf('{')\n if (start !== -1) {\n const recovered = readVerdict(recoverTruncatedJson(raw.slice(start)))\n if (recovered) return recovered\n }\n throw new JudgeParseError('correctness-checker', raw)\n}\n\n/*\n Production `CorrectnessChecker` — one LLM call per matched artifact,\n * deterministic (temperature 0), structured JSON out. Judges fulfilment\n * only: a plan, a gesture, or a description of what should be done does not\n * fulfil a requirement — the artifact must BE the deliverable.\n /\nexport function createLlmCorrectnessChecker(\n chat: ChatClient,\n opts: LlmCorrectnessCheckerOpts = {},\n): CorrectnessChecker {\n const model = opts.model ?? 'claude-sonnet-4-6'\n const maxContentChars = opts.maxContentChars ?? 8000\n const maxAttempts = opts.maxAttempts ?? 2\n const costLedger = opts.costLedger ?? new CostLedger()\n const sink = opts.rawSink\n const record = async (event: RawProviderEvent): Promise<void> => {\n // Forensic capture is best-effort; the verdict is the system of record.\n try {\n await sink?.record(event)\n } catch {\n // Intentionally swallowed.\n }\n }\n const checker: CorrectnessChecker = async (requirement, content) => {\n const request = {\n model,\n messages: [\n {\n role: 'system' as const,\n content:\n 'You verify whether a produced work artifact actually fulfils a stated requirement. Judge fulfilment only — is the deliverable substantively present and on-point — not polish. A plan to do it later, a vague gesture, or a description of what should be done does NOT fulfil a requirement; the artifact must BE the deliverable. Respond with a single JSON object: {\"correct\": boolean, \"reason\": string (<= 30 words)}.',\n },\n {\n role: 'user' as const,\n content: `Requirement: ${requirement.title}\\n${\n requirement.category ? `Category: ${requirement.category}\\n` : ''\n }\\nProduced artifact:\\n${content.slice(0, maxContentChars)}`,\n },\n ],\n temperature: 0,\n maxTokens: 200,\n } satisfies ChatRequest\n let lastErr: unknown\n for (let attempt = 0; attempt < maxAttempts; attempt++) {\n const started = Date.now()\n await record({\n eventId: randomUUID(),\n provider: chat.transport,\n model,\n endpoint: '/chat',\n baseUrl: '',\n attemptIndex: attempt,\n direction: 'request',\n timestamp: started,\n requestBody: request,\n redactedFields: [],\n })\n try {\n const paid = await costLedger.runPaidCall({\n channel: 'verifier',\n phase: opts.costPhase ?? 'completion.correctness',\n actor: 'correctness-checker',\n model,\n maximumCharge:\n chat.maximumAttempts === undefined\n ? undefined\n : maximumChargeForLlmRequest(request, { maximumAttempts: chat.maximumAttempts }),\n tags: {\n ...opts.costTags,\n requirementId: requirement.reqId,\n attempt: String(attempt),\n },\n signal: opts.signal,\n execute: (signal, callId) => chat.chat(request, { signal, idempotencyKey: callId }),\n receipt: costReceiptFromLlm,\n receiptFromError: costReceiptFromLlmError,\n })\n if (!paid.succeeded) throw paid.error\n const resp = paid.value\n // Hold the transport to its own word: a verdict produced by a model\n // other than the one requested is not this model's verdict. A\n // transport that echoes no id cannot be made to prove identity here —\n // callers needing that proof enable it at the client\n // (`LlmClientOptions.assertServedModel`) or gate on `assertModelsServed`.\n assertServedModel(model, resp.servedModel, {\n allowUnreported: true,\n context: `correctness checker for requirement ${requirement.reqId}`,\n })\n const raw = resp.content\n await record({\n eventId: randomUUID(),\n provider: chat.transport,\n model,\n endpoint: '/chat',\n baseUrl: '',\n attemptIndex: attempt,\n direction: 'response',\n timestamp: Date.now(),\n durationMs: Date.now() - started,\n responseBody: resp,\n redactedFields: [],\n })\n return parseCorrectnessResponse(raw)\n } catch (err) {\n lastErr = err\n await record({\n eventId: randomUUID(),\n provider: chat.transport,\n model,\n endpoint: '/chat',\n baseUrl: '',\n attemptIndex: attempt,\n direction: 'error',\n timestamp: Date.now(),\n durationMs: Date.now() - started,\n errorMessage: err instanceof Error ? err.message : String(err),\n redactedFields: [],\n })\n // Substitution is a verdict about identity, not a transient fault:\n // another attempt reaches the same wrong model and spends more tokens.\n if (err instanceof ModelSubstitutionError) throw err\n }\n }\n throw lastErr instanceof Error ? lastErr : new Error(String(lastErr))\n }\n checker.attestation = {\n strategy: 'judge',\n checker: { name: 'llm-correctness-checker', version: model },\n assumptions: [\n 'served-model identity is accepted when the transport reports none (assertServedModel allowUnreported)',\n ],\n }\n return checker\n}\n\n/* Stopwords for requirement-title tokenization — drops the imperative verbs\n * ('review', 'update', …) common to deliverable titles so recall keys on the\n * substantive nouns, not the boilerplate ask. /\nconst TITLE_STOPWORDS = new Set([\n 'the',\n 'a',\n 'an',\n 'and',\n 'or',\n 'for',\n 'to',\n 'of',\n 'in',\n 'on',\n 'with',\n 'review',\n 'update',\n 'new',\n 'proposed',\n])\n\n/\n Deterministic `CorrectnessChecker` — the no-LLM counterpart to\n * `createLlmCorrectnessChecker`. A produced item fulfils a requirement when its\n * content is substantive (≥ `minContentLength` chars) AND recalls ≥ `minRecall`\n * of the requirement title's significant tokens. No network.\n \n Polarity-blind: token recall credits a negation that contains the\n * requirement's tokens (\"I will NOT produce the comparison\" recalls every token\n * of \"produce the comparison\"). The structural match stage is ALSO lexical, so\n * pairing the two collapses to a single gameable gate. Use this only as an\n * opt-in structural pre-filter or for tasks whose requirements have no polarity\n * to invert; for produced-state grading the correctness checker MUST be semantic\n * (`createLlmCorrectnessChecker`). See the anti-game fixtures in the test suite.\n /\nexport function createTokenRecallChecker(\n opts: { minRecall?: number; minContentLength?: number } = {},\n): CorrectnessChecker {\n const minRecall = opts.minRecall ?? 0.5\n const minLen = opts.minContentLength ?? 120\n const checker: CorrectnessChecker = async (requirement, content) => {\n const body = content.trim()\n if (body.length < minLen)\n return {\n correct: false,\n reason: `content too thin (${body.length} chars) to be the deliverable`,\n }\n const titleTokens = requirement.title\n .toLowerCase()\n .split(/[^a-z0-9]+/)\n .filter((t) => t.length > 2 && !TITLE_STOPWORDS.has(t))\n if (titleTokens.length === 0)\n return {\n correct: true,\n reason: 'requirement title has no significant tokens — structural match accepted',\n }\n const lower = body.toLowerCase()\n const hits = titleTokens.filter((t) => lower.includes(t)).length\n const recall = hits / titleTokens.length\n return recall >= minRecall\n ? {\n correct: true,\n reason: `content recalls ${hits}/${titleTokens.length} requirement tokens`,\n }\n : {\n correct: false,\n reason: `content recalls only ${hits}/${titleTokens.length} requirement tokens`,\n }\n }\n // 'schema': token recall checks shape, not meaning — the member whose\n // documented failure mode (\"a well-formed wrong answer passes\") is\n // exactly this checker's polarity blindness.\n checker.attestation = {\n strategy: 'schema',\n checker: { name: 'token-recall-checker', version: packageVersion() },\n assumptions: ['polarity-blind: a negation that recalls the requirement tokens passes'],\n }\n return checker\n}\n","/\n Produced-state extraction — normalize a run's runtime event stream into the\n * typed `ProducedState` the completion oracle consumes.\n \n `ProducedState` answers \"what did the agent actually produce\" — vault\n * artifacts, proposals, tool calls. The runtime emits these as a stream of\n * events; this module is the single normalization point from that stream to\n * the shape `verifyCompletion` expects.\n \n Input is structurally typed (`RuntimeEventLike`) so this module does not\n * depend on agent-runtime — agent-runtime's `RuntimeStreamEvent` satisfies it\n * structurally. The `content` on `ArtifactEventLike` and the whole\n * `proposal_created` variant are the runtime-side enrichments this contract\n * requires; the runtime emits them, this module consumes them.\n /\n\nimport type { Artifact } from './artifact-validator'\nimport type { ProducedProposal, ProducedState } from './completion-verifier'\n\n/* A tool the agent invoked. /\nexport interface ToolCallEventLike {\n type: 'tool_call'\n toolName: string\n}\n\n/\n An artifact the agent produced. `content` is the enriched field — the\n * runtime's base `artifact` event carries only metadata; the completion\n * oracle needs the body to verify the deliverable, so the runtime emits it.\n /\nexport interface ArtifactEventLike {\n type: 'artifact'\n artifactId: string\n name?: string\n mimeType?: string\n uri?: string\n content?: string\n}\n\n/* A proposal / filing the agent created. /\nexport interface ProposalEventLike {\n type: 'proposal_created'\n proposalId: string\n title: string\n status?: 'pending' \| 'approved' \| 'rejected'\n // body of the proposal (e.g. a submit_proposal `description`). When present,\n // the completion oracle correctness-checks it like artifact content; absent,\n // the proposal is graded presence-only.\n content?: string\n}\n\n/\n The subset of runtime stream events `extractProducedState` consumes.\n * agent-runtime's full `RuntimeStreamEvent` union satisfies this structurally;\n * the `{ type: string }` catch-all keeps the input permissive so callers can\n * pass the whole unfiltered telemetry stream — unrecognized events are skipped.\n /\nexport type RuntimeEventLike =\n \| ToolCallEventLike\n \| ArtifactEventLike\n \| ProposalEventLike\n \| { type: string }\n\nfunction artifactKind(mimeType: string \| undefined): string {\n if (!mimeType) return 'file'\n if (mimeType.includes('json')) return 'json'\n if (mimeType.startsWith('text/')) return 'text'\n return 'file'\n}\n\n/\n Normalize an emission-ordered stream into its latest observed produced state.\n * Artifacts are keyed by their exact output path (name, then URI, then id);\n * proposals by id. Later observations replace earlier ones, including missing\n * content or a rejected status. An obsolete version cannot prove completion.\n * Distinct paths remain distinct; no path normalization or aliasing is inferred.\n * Results retain first-seen identity order. Tool names describe invocation,\n * not success, and are deduplicated in first-seen order.\n \n An artifact without observed content yields empty content, which the\n * completion oracle rejects. This projection does not verify external effects,\n * recover dropped events, or establish freshness beyond the supplied stream.\n */\nexport function extractProducedState(events: readonly RuntimeEventLike[]): ProducedState {\n const artifacts = new Map<string, Artifact>()\n const proposals = new Map<string, ProducedProposal>()\n const toolCalls: string[] = []\n const seenTools = new Set<string>()\n\n for (const ev of events) {\n if (ev.type === 'tool_call') {\n const name = (ev as ToolCallEventLike).toolName\n if (name && !seenTools.has(name)) {\n seenTools.add(name)\n toolCalls.push(name)\n }\n } else if (ev.type === 'artifact') {\n const a = ev as ArtifactEventLike\n const path = a.name ?? a.uri ?? a.artifactId\n artifacts.set(path, {\n kind: artifactKind(a.mimeType),\n path,\n content: a.content ?? '',\n })\n } else if (ev.type === 'proposal_created') {\n const p = ev as ProposalEventLike\n proposals.set(p.proposalId, {\n id: p.proposalId,\n title: p.title,\n status: p.status ?? 'pending',\n ...(p.content !== undefined ? { content: p.content } : {}),\n })\n }\n }\n\n return { artifacts: [...artifacts.values()], proposals: [...proposals.values()], toolCalls }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAeA,MAAa,mBAA2C;CACtD;CACA;CACA;CACA;AACF;;;;;;AAwBA,MAAa,uBAAuB;;;;;;;;;;;;;;;;;;;AAoBpC,SAAgB,kBAAkB,MAAuC;CACvE,MAAM,YAAY,KAAK,aAAa;CACpC,IAAI,UAAU,WAAW,GAAG,MAAM,IAAI,gBAAgB,0CAA0C;CAChG,MAAM,YAAY,KAAK,KAAK,OAAO;CACnC,MAAM,SAAS,KAAK,WAAW,YAAY,CAAC,SAAS,IAAI,CAAC;CAC1D,IAAI,OAAO,WAAW,GACpB,MAAM,IAAI,gBACR,kGACF;CAEF,MAAM,MAAsB,CAAC;CAC7B,MAAM,uBAAO,IAAI,IAAY;CAC7B,KAAK,MAAM,WAAW,WAAW;EAS/B,MAAM,YAAY,KAAK,mBACnB,SACA,OAAO,QAAQ,UAAU,qBAAqB,SAAS,KAAK,CAAC;EACjE,MAAM,YAAY,UAAU,SAAS,IAAI,YAAY,CAAC,oBAAoB;EAC1E,KAAK,MAAM,SAAS,WAAW;GAC7B,MAAM,UAAwB;IAC5B,GAAG,KAAK;IACR;IACA,MAAM,GAAG,KAAK,KAAK,QAAQ,QAAQ,GAAG,QAAQ,GAAG;IACjD,OAAO;KAAE,GAAG,KAAK,KAAK;KAAO,SAAS;IAAM;IAC5C,UAAU;KAAE,GAAI,KAAK,KAAK,YAAY,CAAC;KAAI;KAAS,cAAc;IAAM;GAC1E;GACA,MAAM,KAAK,eAAe,OAAO;GACjC,IAAI,KAAK,IAAI,EAAE,GAAG;GAClB,KAAK,IAAI,EAAE;GACX,IAAI,KAAK,OAAO;EAClB;CACF;CACA,IAAI,IAAI,WAAW,GAGjB,MAAM,IAAI,gBACR,uDAAuD,UAAU,KAAK,IAAI,EAAE,aAAa,OAAO,KAAK,IAAI,EAAE,IAC7G;CAEF,OAAO;AACT;;;;;;;;AASA,SAAgB,cACd,SACqD;CACrD,MAAM,IAAI,QAAQ;CAClB,MAAM,UAAU,GAAG;CACnB,MAAM,QAAQ,GAAG;CACjB,IAAI,OAAO,YAAY,YAAY,OAAO,UAAU,UAClD,OAAO;EAAW;EAAwB;CAAM;AAGpD;;;;;;;;AASA,SAAgB,eAAe,SAA+B;CAE5D,OAAO,GADO,qBAAqB,yBAAyB,OAAO,CAAC,KAAK,UACzD,GAAG,iBAAiB,OAAO,CAAC,CAAC,MAAM,GAAG,EAAE;AAC1D;;;;;AAMA,SAAgB,oBAAoB,SAA+B;CACjE,MAAM,QAAQ,QAAQ,OAAO,SAAS,KAAK;CAC3C,IAAI,CAAC,OAAO;EACV,MAAM,QAAQ,yBAAyB,OAAO,KAAK;EACnD,MAAM,IAAI,gBACR,iBAAiB,MAAM,gDACzB;CACF;CACA,OAAO;AACT;AAEA,SAAS,yBAAyB,SAA2C;CAC3E,OAAO,QAAQ,MAAM,KAAK,KAAK,QAAQ,SAAS,KAAK,KAAK,KAAA;AAC5D;AAEA,SAAS,qBAAqB,OAA+C;CAM3E,OALa,OACT,KAAK,CAAC,CACP,QAAQ,qBAAqB,GAAG,CAAC,CACjC,QAAQ,OAAO,GAAG,CAAC,CACnB,QAAQ,UAAU,EAAE,KACR,KAAA;AACjB;AAEA,SAAS,QAA2C,OAAsB;CACxE,MAAM,MAA+B,CAAC;CACtC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,KAAK,GAC7C,IAAI,UAAU,KAAA,GAAW,IAAI,OAAO;CAEtC,OAAO;AACT;;;;;;;;;;;AAYA,SAAgB,iBAAiB,SAA+B;CAC9D,MAAM,QAAQ,oBAAoB,OAAO;CACzC,MAAM,YAAY,QAAQ;EACxB,GAAG;EACH,MAAM,KAAA;EACN,aAAa,KAAA;EACb,MAAM,QAAQ,OAAO,CAAC,GAAG,QAAQ,IAAI,CAAC,CAAC,KAAK,IAAI,KAAA;EAChD,OAAO,QAAQ;GAAE,GAAG,QAAQ;GAAO,SAAS;EAAM,CAAC;CACrD,CAAC;CACD,OAAO,WAAW,QAAQ,CAAC,CAAC,OAAO,gBAAgB,SAAS,CAAC,CAAC,CAAC,OAAO,KAAK;AAC7E;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACtEA,SAAgB,kBAAkB,OAMZ;CACpB,IAAI,MAAM,aAAa,WAAW,GAChC,MAAM,IAAI,MACR,4BAA4B,MAAM,OAAO,+DAC3C;CAEF,MAAM,aAAa,MAAM,aAAa,QAAQ,MAAM,CAAC,EAAE,UAAU;CACjE,MAAM,kBAAkB,MAAM,aAAa,SAAS,WAAW;CAC/D,IAAI,WAAW,WAAW,GAGxB,MAAM,IAAI,MACR,4BAA4B,MAAM,OAAO,yCAAyC,MAAM,aAAa,OAAO,8BAA8B,MAAM,aAAa,EAAE,EAAE,oBAAoB,iBAAiB,EACxM;CAEF,MAAM,iBAAiB,WAAW,QAAQ,MAAM,EAAE,SAAS,CAAC,CAAC;CAC7D,MAAM,iBAAiB,iBAAiB,WAAW;CAGnD,MAAM,gBAAgB,oBAAoB,KAAK,mBAAmB,WAAW;CAC7E,OAAO;EACL,QAAQ,MAAM;EACd,cAAc,MAAM;EACpB;EACA;EACA;EACA,OAAO;EACP,OAAO;EACP,GAAI,MAAM,kBAAkB,KAAA,IAAY,CAAC,IAAI,EAAE,eAAe,MAAM,cAAc;CACpF;AACF;AA+BA,MAAM,4BAAY,IAAI,IAAI;CACxB;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAUD,MAAM,6CAA6B,IAAI,IAAI;CACzC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAED,MAAM,kBAAkB;AACxB,MAAM,oBAAoB;AAE1B,SAAS,OAAO,GAAW,WAAsC;CAC/D,OAAO,IAAI,IACT,EACG,YAAY,CAAC,CACb,MAAM,YAAY,CAAC,CACnB,QAAQ,MAAM,EAAE,SAAS,KAAK,CAAC,UAAU,IAAI,CAAC,KAAK,CAAC,WAAW,IAAI,CAAC,CAAC,CAC1E;AACF;;;;;;;AAQA,SAAS,YAAY,iBAAyB,eAA+B;CAC3E,MAAM,MAAM,OAAO,iBAAiB,0BAA0B;CAC9D,IAAI,IAAI,SAAS,GAAG,OAAO;CAC3B,MAAM,OAAO,OAAO,aAAa;CACjC,IAAI,MAAM;CACV,KAAK,MAAM,KAAK,KAAK,IAAI,KAAK,IAAI,CAAC,GAAG;CACtC,OAAO,MAAM,IAAI;AACnB;AAYA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,UAAU,GAAG,IAAI,MAAM,GAAG,IAAI,YAAY;CAChD,MAAM,MAAmB,CAAC;CAC1B,UAAU,SAAS,GAAG,MAAM;EAC1B,KAAK,EAAE,WAAW,GAAA,CAAI,KAAK,CAAC,CAAC,SAAS,mBAAmB;EAKzD,IAAI,QAAQ,YACV,SACA,GAAG,EAAE,QAAQ,GAAG,GAAG,EAAE,KAAK,IAAI,EAAE,WAAW,GAAA,CAAI,MAAM,GAAG,GAAI,GAC9D;EACA,IAAI,IAAI,YAAY,EAAE,QAAQ,IAAI,SAAS,YAAY,MAAM,EAAE,KAAK,YAAY,GAC9E,QAAQ,KAAK,IAAI,OAAO,CAAC;EAE3B,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,YAAY;GACrB;GACA,UAAU,aAAa,EAAE,QAAQ,EAAE,KAAK,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACnF,SAAS,EAAE,WAAW;EACxB,CAAC;CACH,CAAC;CACD,OAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,UAAU,GAAG,IAAI,MAAM,GAAG,IAAI,YAAY;CAChD,MAAM,MAAmB,CAAC;CAC1B,KAAK,MAAM,KAAK,WAAW;EAEzB,IAAI,EAAE,WAAW,YAAY;EAM7B,MAAM,QAAQ,EAAE,WAAW,GAAA,CAAI,KAAK;EACpC,IAAI,KAAK,SAAS,mBAAmB;EASrC,MAAM,QAAQ,YAAY,SAAS,GAAG,EAAE,MAAM,GAAG,MAAM;EACvD,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,YAAY,EAAE;GACvB;GACA,UAAU,sBAAsB,EAAE,MAAM,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACnF,SAAS;EACX,CAAC;CACH;CACA,OAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,MAAmB,CAAC;CAC1B,UAAU,SAAS,MAAM,MAAM;EAC7B,MAAM,QAAQ,YAAY,IAAI,OAAO,IAAI;EACzC,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,QAAQ;GACjB;GACA,UAAU,cAAc,KAAK,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACxE,SAAS;EACX,CAAC;CACH,CAAC;CACD,OAAO;AACT;;;;;;;;AASA,eAAsB,iBACpB,MACA,OACA,kBAC4B;CAC5B,IAAI,KAAK,aAAa,WAAW,GAC/B,MAAM,IAAI,MACR,2BAA2B,KAAK,OAAO,4CACzC;CAMF,MAAM,aAA0B,CAAC;CACjC,KAAK,aAAa,SAAS,KAAK,MAAM;EACpC,MAAM,KAAK,IAAI,eAAe;EAC9B,IAAI,OAAO,cAAc,OAAO,OAC9B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;EAEhE,IAAI,OAAO,cAAc,OAAO,OAC9B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;EAEhE,IAAI,OAAO,eAAe,OAAO,OAC/B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;CAElE,CAAC;CACD,WAAW,MAAM,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;CAE3C,MAAM,2BAAW,IAAI,IAAuB;CAC5C,MAAM,4BAAY,IAAI,IAAY;CAClC,KAAK,MAAM,KAAK,YAAY;EAC1B,IAAI,SAAS,IAAI,EAAE,QAAQ,KAAK,UAAU,IAAI,EAAE,OAAO,GAAG;EAC1D,SAAS,IAAI,EAAE,UAAU,CAAC;EAC1B,UAAU,IAAI,EAAE,OAAO;CACzB;CAEA,MAAM,eAAmC,CAAC;CAC1C,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,aAAa,QAAQ,KAAK;EACjD,MAAM,MAAM,KAAK,aAAa;EAC9B,MAAM,QAAQ,SAAS,IAAI,CAAC;EAC5B,MAAM,WAAqB,CAAC;EAC5B,IAAI,UAA0B;EAC9B,IAAI;EAEJ,IAAI,OAAO;GACT,SAAS,KAAK,MAAM,QAAQ;GAC5B,IAAI,MAAM,YAAY,MACpB,IAAI;IACF,MAAM,IAAI,MAAM,iBAAiB,KAAK,MAAM,OAAO;IACnD,UAAU,EAAE;IACZ,SAAS,KAAK,gBAAgB,EAAE,UAAU,SAAS,OAAO,KAAK,EAAE,QAAQ;GAC3E,SAAS,KAAK;IAIZ,mBACE,eAAe,kBACX,6CAA6C,IAAI,IAAI,MAAM,GAAG,GAAG,MACjE,wBAAwB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IAC7E,SAAS,KAAK,6BAA6B,kBAAkB;GAC/D;QAEA,SAAS,KAAK,6DAA6D;EAE/E,OAAO;GACL,MAAM,KAAK,IAAI,eAAe;GAC9B,MAAM,OAAO,OAAO,QAAQ,gCAAgC;GAC5D,SAAS,KAAK,eAAe,KAAK,0BAA0B;EAC9D;EAEA,MAAM,sBAAsB,UAAU,KAAA;EACtC,MAAM,aAAa,qBAAqB,KAAA;EACxC,MAAM,YAAY,uBAAuB,CAAC,cAAc,YAAY;EACpE,aAAa,KAAK;GAChB,OAAO,IAAI;GACX,OAAO,IAAI;GACX;GACA;GACA;GACA,GAAI,aAAa;IAAE,YAAY;IAAe;GAAiB,IAAI,CAAC;GACpE;EACF,CAAC;CACH;CAMA,MAAM,cAAc,iBAAiB;CACrC,MAAM,gBAAkD,cACpD;EACE,UAAU,YAAY;EACtB,SAAS,YAAY;EACrB,aAAa,CACX,sHACA,GAAG,YAAY,WACjB;EACA,gBAAgB,4BAA4B;GAAE,QAAQ,KAAK;GAAQ;EAAa,CAAC;CACnF,IACA,KAAA;CAEJ,OAAO,kBAAkB;EACvB,QAAQ,KAAK;EACb;EACA,GAAI,kBAAkB,KAAA,IAAY,CAAC,IAAI,EAAE,cAAc;CACzD,CAAC;AACH;;;;;;;;AAgCA,SAAgB,yBAAyB,KAAmD;CAC1F,MAAM,eAAe,cAAoE;EACvF,IAAI,cAAc,QAAQ,OAAO,cAAc,UAAU,OAAO;EAChE,MAAM,EAAE,SAAS,WAAW;EAC5B,IAAI,OAAO,YAAY,WAAW,OAAO;EACzC,OAAO;GAAE;GAAS,QAAQ,OAAO,WAAW,WAAW,SAAS;EAAG;CACrE;CAEA,MAAM,QAAQ,IAAI,MAAM,aAAa;CACrC,IAAI,OACF,IAAI;EACF,MAAM,SAAS,YAAY,KAAK,MAAM,MAAM,EAAE,CAAC;EAC/C,IAAI,QAAQ,OAAO;CACrB,QAAQ,CAER;CAIF,MAAM,QAAQ,IAAI,QAAQ,GAAG;CAC7B,IAAI,UAAU,IAAI;EAChB,MAAM,YAAY,YAAY,qBAAqB,IAAI,MAAM,KAAK,CAAC,CAAC;EACpE,IAAI,WAAW,OAAO;CACxB;CACA,MAAM,IAAI,gBAAgB,uBAAuB,GAAG;AACtD;;;;;;;AAQA,SAAgB,4BACd,MACA,OAAkC,CAAC,GACf;CACpB,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,kBAAkB,KAAK,mBAAmB;CAChD,MAAM,cAAc,KAAK,eAAe;CACxC,MAAM,aAAa,KAAK,cAAc,IAAI,WAAW;CACrD,MAAM,OAAO,KAAK;CAClB,MAAM,SAAS,OAAO,UAA2C;EAE/D,IAAI;GACF,MAAM,MAAM,OAAO,KAAK;EAC1B,QAAQ,CAER;CACF;CACA,MAAM,UAA8B,OAAO,aAAa,YAAY;EAClE,MAAM,UAAU;GACd;GACA,UAAU,CACR;IACE,MAAM;IACN,SACE;GACJ,GACA;IACE,MAAM;IACN,SAAS,gBAAgB,YAAY,MAAM,IACzC,YAAY,WAAW,aAAa,YAAY,SAAS,MAAM,GAChE,wBAAwB,QAAQ,MAAM,GAAG,eAAe;GAC3D,CACF;GACA,aAAa;GACb,WAAW;EACb;EACA,IAAI;EACJ,KAAK,IAAI,UAAU,GAAG,UAAU,aAAa,WAAW;GACtD,MAAM,UAAU,KAAK,IAAI;GACzB,MAAM,OAAO;IACX,SAAS,WAAW;IACpB,UAAU,KAAK;IACf;IACA,UAAU;IACV,SAAS;IACT,cAAc;IACd,WAAW;IACX,WAAW;IACX,aAAa;IACb,gBAAgB,CAAC;GACnB,CAAC;GACD,IAAI;IACF,MAAM,OAAO,MAAM,WAAW,YAAY;KACxC,SAAS;KACT,OAAO,KAAK,aAAa;KACzB,OAAO;KACP;KACA,eACE,KAAK,oBAAoB,KAAA,IACrB,KAAA,IACA,2BAA2B,SAAS,EAAE,iBAAiB,KAAK,gBAAgB,CAAC;KACnF,MAAM;MACJ,GAAG,KAAK;MACR,eAAe,YAAY;MAC3B,SAAS,OAAO,OAAO;KACzB;KACA,QAAQ,KAAK;KACb,UAAU,QAAQ,WAAW,KAAK,KAAK,SAAS;MAAE;MAAQ,gBAAgB;KAAO,CAAC;KAClF,SAAS;KACT,kBAAkB;IACpB,CAAC;IACD,IAAI,CAAC,KAAK,WAAW,MAAM,KAAK;IAChC,MAAM,OAAO,KAAK;IAMlB,kBAAkB,OAAO,KAAK,aAAa;KACzC,iBAAiB;KACjB,SAAS,uCAAuC,YAAY;IAC9D,CAAC;IACD,MAAM,MAAM,KAAK;IACjB,MAAM,OAAO;KACX,SAAS,WAAW;KACpB,UAAU,KAAK;KACf;KACA,UAAU;KACV,SAAS;KACT,cAAc;KACd,WAAW;KACX,WAAW,KAAK,IAAI;KACpB,YAAY,KAAK,IAAI,IAAI;KACzB,cAAc;KACd,gBAAgB,CAAC;IACnB,CAAC;IACD,OAAO,yBAAyB,GAAG;GACrC,SAAS,KAAK;IACZ,UAAU;IACV,MAAM,OAAO;KACX,SAAS,WAAW;KACpB,UAAU,KAAK;KACf;KACA,UAAU;KACV,SAAS;KACT,cAAc;KACd,WAAW;KACX,WAAW,KAAK,IAAI;KACpB,YAAY,KAAK,IAAI,IAAI;KACzB,cAAc,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;KAC7D,gBAAgB,CAAC;IACnB,CAAC;IAGD,IAAI,eAAe,wBAAwB,MAAM;GACnD;EACF;EACA,MAAM,mBAAmB,QAAQ,UAAU,IAAI,MAAM,OAAO,OAAO,CAAC;CACtE;CACA,QAAQ,cAAc;EACpB,UAAU;EACV,SAAS;GAAE,MAAM;GAA2B,SAAS;EAAM;EAC3D,aAAa,CACX,uGACF;CACF;CACA,OAAO;AACT;;;ACpmBA,SAAS,aAAa,UAAsC;CAC1D,IAAI,CAAC,UAAU,OAAO;CACtB,IAAI,SAAS,SAAS,MAAM,GAAG,OAAO;CACtC,IAAI,SAAS,WAAW,OAAO,GAAG,OAAO;CACzC,OAAO;AACT;;;;;;;;;;;;;;AAeA,SAAgB,qBAAqB,QAAoD;CACvF,MAAM,4BAAY,IAAI,IAAsB;CAC5C,MAAM,4BAAY,IAAI,IAA8B;CACpD,MAAM,YAAsB,CAAC;CAC7B,MAAM,4BAAY,IAAI,IAAY;CAElC,KAAK,MAAM,MAAM,QACf,IAAI,GAAG,SAAS,aAAa;EAC3B,MAAM,OAAQ,GAAyB;EACvC,IAAI,QAAQ,CAAC,UAAU,IAAI,IAAI,GAAG;GAChC,UAAU,IAAI,IAAI;GAClB,UAAU,KAAK,IAAI;EACrB;CACF,OAAO,IAAI,GAAG,SAAS,YAAY;EACjC,MAAM,IAAI;EACV,MAAM,OAAO,EAAE,QAAQ,EAAE,OAAO,EAAE;EAClC,UAAU,IAAI,MAAM;GAClB,MAAM,aAAa,EAAE,QAAQ;GAC7B;GACA,SAAS,EAAE,WAAW;EACxB,CAAC;CACH,OAAO,IAAI,GAAG,SAAS,oBAAoB;EACzC,MAAM,IAAI;EACV,UAAU,IAAI,EAAE,YAAY;GAC1B,IAAI,EAAE;GACN,OAAO,EAAE;GACT,QAAQ,EAAE,UAAU;GACpB,GAAI,EAAE,YAAY,KAAA,IAAY,EAAE,SAAS,EAAE,QAAQ,IAAI,CAAC;EAC1D,CAAC;CACH;CAGF,OAAO;EAAE,WAAW,CAAC,GAAG,UAAU,OAAO,CAAC;EAAG,WAAW,CAAC,GAAG,UAAU,OAAO,CAAC;EAAG;CAAU;AAC7F"}
1	+ {"version":3,"file":"produced-state-g9BPgu4D.js","names":[],"sources":["../src/agent-profile.ts","../src/completion-verifier.ts","../src/produced-state.ts"],"sourcesContent":["import { createHash } from 'node:crypto'\nimport type { AgentProfile } from '@tangle-network/agent-interface'\nimport { type HarnessType, harnessSupportsModel } from '@tangle-network/agent-interface'\nimport { ValidationError } from './errors'\nimport { canonicalString } from './ledger-core/canonical'\n\nexport type { AgentProfile, HarnessType } from '@tangle-network/agent-interface'\n\n/*\n The agentic coding harnesses an eval sweeps by default — the ones we care about\n * ranking. This is the SINGLE source of that list; consumers import it instead of\n * re-declaring their own (a re-declared list is how the fleet drifts). Pass an\n * explicit `harnesses` (e.g. `harnessTypeSchema.options` for literally every known\n * harness) to widen beyond these.\n /\nexport const CODING_HARNESSES: readonly HarnessType[] = [\n 'opencode',\n 'claude-code',\n 'codex',\n 'kimi-code',\n]\n\nexport interface ProfileAxisSpec {\n /* The domain profile to sweep. Its prompt/tools/skills are held fixed; only the\n * harness and model vary. `model.default` is the fallback model. /\n base: AgentProfile\n /* Harnesses to cross. Default: {@link CODING_HARNESSES}. /\n harnesses?: readonly HarnessType[]\n /* Models to cross. Default: `[base.model.default]` — one model, i.e. today's\n * single-model behaviour, so omitting this never changes an existing run. /\n models?: readonly string[]\n /* Force every (harness, model) pair verbatim, even ones the harness can't run —\n * for deliberately testing failure modes. Default (false): SNAP instead — a\n * vendor-locked harness runs only the swept models in its family, or its native\n * default when it supports none, so no harness is dropped and none gets a\n * guaranteed-failing foreign-model cell. /\n keepIncompatible?: boolean\n}\n\n/* Model sentinel for a vendor-locked harness that supports none of the swept models:\n * it carries no provider prefix, so `harnessSupportsModel` accepts it and the harness\n * resolves it to its own native default model at runtime (e.g. kimi-code → its Kimi\n * model). Lets `expandProfileAxes` snap-instead-of-drop without a per-harness flagship\n * table that would rot as router catalogs change. /\nexport const HARNESS_NATIVE_MODEL = 'default'\n\n/\n Expand a base profile across the harness × model matrix into the `AgentProfile[]`\n * that `runProfileMatrix` / `selfImprove` score — the ONE place \"which harnesses ×\n * which models do we evaluate\" lives, so no product hand-rolls its own harness list\n * or column→profile mapping (the pattern that let those copies drift and silently\n * break the harness pivot).\n \n Each cell clones `base`, sets the canonical top-level `harness` and `model.default`,\n * and stamps `metadata.harness` + `metadata.harnessModel` for matrix grouping. Both\n * metadata fields are hash-bearing, so every cell gets a distinct `agentProfileId` row\n * and results join back by harness/model via {@link harnessAxisOf} with no\n * hand-recomputed key. A vendor-locked harness snaps to its family's swept models — or\n * its native default ({@link HARNESS_NATIVE_MODEL}) when it supports none — so every\n * requested harness runs; `keepIncompatible` forces every pair verbatim.\n \n Omit `harnesses`/`models` to sweep the full default set — the \"turn it on for\n * everything we care about\" switch, identical in shape whether one harness or all.\n /\nexport function expandProfileAxes(spec: ProfileAxisSpec): AgentProfile[] {\n const harnesses = spec.harnesses ?? CODING_HARNESSES\n if (harnesses.length === 0) throw new ValidationError('expandProfileAxes: no harnesses to sweep')\n const baseModel = spec.base.model?.default\n const models = spec.models ?? (baseModel ? [baseModel] : [])\n if (models.length === 0) {\n throw new ValidationError(\n 'expandProfileAxes: no models to sweep — base profile has no model.default and none were supplied',\n )\n }\n const out: AgentProfile[] = []\n const seen = new Set<string>()\n for (const harness of harnesses) {\n // A universal (router-backed) harness — opencode/pi/claudish — runs every swept\n // model. A vendor-locked harness — codex/claude-code/kimi-code — runs only the\n // swept models in its own family; when it supports NONE of them it snaps to its\n // native default (the `HARNESS_NATIVE_MODEL` sentinel it resolves at runtime)\n // rather than being dropped, so every requested harness still appears in the\n // sweep on a model it can actually run — e.g. sweeping `deepseek/x` puts opencode\n // on deepseek and kimi-code on its own Kimi model, a real head-to-head.\n // `keepIncompatible` forces every (harness, model) pair verbatim (failure-mode runs).\n const supported = spec.keepIncompatible\n ? models\n : models.filter((model) => harnessSupportsModel(harness, model))\n const effective = supported.length > 0 ? supported : [HARNESS_NATIVE_MODEL]\n for (const model of effective) {\n const profile: AgentProfile = {\n ...spec.base,\n harness,\n name: `${spec.base.name ?? 'agent'}/${harness}/${model}`,\n model: { ...spec.base.model, default: model },\n metadata: { ...(spec.base.metadata ?? {}), harness, harnessModel: model },\n }\n const id = agentProfileId(profile)\n if (seen.has(id)) continue\n seen.add(id)\n out.push(profile)\n }\n }\n if (out.length === 0) {\n // Unreachable in normal use — snapping guarantees ≥1 cell per harness — but keep a\n // fail-closed guard so a future refactor can't silently produce an empty sweep.\n throw new ValidationError(\n `expandProfileAxes: produced no profiles (harnesses=[${harnesses.join(', ')}], models=[${models.join(', ')}]).`,\n )\n }\n return out\n}\n\n/\n Read the (harness, model) a matrix cell ran under, off a profile or a result row's\n * profile — the join-back for a `byHarness` pivot. Returns undefined when the profile\n * wasn't produced by {@link expandProfileAxes}. Callers group `result.byProfile` by\n * this instead of recomputing an id (recomputing the wrong key is what broke the pivot\n * in the hand-rolled copies).\n /\nexport function harnessAxisOf(\n profile: Pick<AgentProfile, 'metadata'>,\n): { harness: HarnessType; model: string } \| undefined {\n const m = profile.metadata as Record<string, unknown> \| undefined\n const harness = m?.harness\n const model = m?.harnessModel\n if (typeof harness === 'string' && typeof model === 'string') {\n return { harness: harness as HarnessType, model }\n }\n return undefined\n}\n\n/\n Collision-resistant, path-safe, human-readable profile id for eval artifacts.\n * Scorecard joins still use `agentProfileHash`; this id is for run ids, matrix\n * keys, and directory names where two profiles must not collapse onto one row.\n * The suffix is the first 64 bits of the behaviour hash, enough for ordinary\n * eval matrices while keeping filenames readable.\n /\nexport function agentProfileId(profile: AgentProfile): string {\n const label = pathSafeProfileLabel(agentProfileDisplayLabel(profile)) ?? 'profile'\n return `${label}-${agentProfileHash(profile).slice(0, 16)}`\n}\n\n/\n Model snapshot used for `RunRecord.model`. Eval surfaces require a concrete\n * model id because run records reject bare/missing model aliases.\n /\nexport function agentProfileModelId(profile: AgentProfile): string {\n const model = profile.model?.default?.trim()\n if (!model) {\n const label = agentProfileDisplayLabel(profile) ?? 'unnamed profile'\n throw new ValidationError(\n `AgentProfile \"${label}\" has no model.default — cannot record eval run`,\n )\n }\n return model\n}\n\nfunction agentProfileDisplayLabel(profile: AgentProfile): string \| undefined {\n return profile.name?.trim() \|\| profile.version?.trim() \|\| undefined\n}\n\nfunction pathSafeProfileLabel(label: string \| undefined): string \| undefined {\n const safe = label\n ?.trim()\n .replace(/[^A-Za-z0-9._-]+/g, '-')\n .replace(/-+/g, '-')\n .replace(/^-\|-$/g, '')\n return safe \|\| undefined\n}\n\nfunction compact<T extends Record<string, unknown>>(input: T): Partial<T> {\n const out: Record<string, unknown> = {}\n for (const [key, value] of Object.entries(input)) {\n if (value !== undefined) out[key] = value\n }\n return out as Partial<T>\n}\n\n/\n Deterministic behaviour identity for the canonical\n * `@tangle-network/agent-interface` AgentProfile.\n \n `name` and `description` are labels and do not affect the hash. Profile\n * `version`, prompt, model hints, tools, resources, hooks, modes, permissions,\n * and extensions do affect the hash. Resource array order is hash-bearing\n * because mount order can change agent behaviour. Undefined fields are treated\n * as absent; explicit `null` fields remain hash-bearing.\n /\nexport function agentProfileHash(profile: AgentProfile): string {\n const model = agentProfileModelId(profile)\n const behaviour = compact({\n ...profile,\n name: undefined,\n description: undefined,\n tags: profile.tags ? [...profile.tags].sort() : undefined,\n model: compact({ ...profile.model, default: model }),\n })\n return createHash('sha256').update(canonicalString(behaviour)).digest('hex')\n}\n","/\n Completion verifier — the task-completion oracle.\n \n Answers the only eval question that is not a proxy: did the agent actually\n * COMPLETE the task — produce every required deliverable, persisted and\n * correct — rather than describe what should be done. A fluent transcript\n * that never produces the artifact scores zero here.\n \n Per requirement, a two-stage check:\n * 1. Structural — a produced item (vault artifact / approved proposal /\n * tool call) of the right kind is matched against the requirement and\n * carries non-empty content. Deterministic; no LLM.\n * 2. Correctness — only if structurally present AND the matched item\n * carries content, one targeted check decides whether that item\n * actually fulfils the requirement. A hallucinated artifact fails here;\n * an absent one already failed stage 1.\n \n `completionRate` is satisfied / MEASURABLE requirements (unmeasured rows —\n * checker failures — are excluded from the denominator, never scored as\n * zeros). Quality dimensions are meaningless on an incomplete task — callers\n * gate on `fullyComplete` / `completionRate` before scoring quality.\n /\n\nimport { randomUUID } from 'node:crypto'\nimport type { ChatClient, ChatRequest } from './analyst/chat-client'\nimport type { Artifact } from './artifact-validator'\nimport { CostLedger, type CostLedgerHandle } from './cost-ledger'\nimport { assertServedModel, ModelSubstitutionError } from './integrity/served-model'\nimport { recoverTruncatedJson } from './json-recovery'\nimport { JudgeParseError } from './judges'\nimport {\n costReceiptFromLlm,\n costReceiptFromLlmError,\n maximumChargeForLlmRequest,\n} from './llm-client'\nimport { packageVersion } from './package-version'\nimport type { RawProviderEvent, RawProviderSink } from './trace/raw-provider-sink'\nimport {\n certificationEvidenceDigest,\n type DefaultVerdict,\n type VerdictCertification,\n} from './verdict'\nimport type { CheckerIdentity, VerificationStrategySource } from './verification-strategy'\n\n/* What kind of produced state can satisfy a requirement structurally. /\nexport type SatisfiedBy = 'artifact' \| 'proposal' \| 'tool-call' \| 'any'\n\nexport interface CompletionRequirement {\n /* Stable id from the task gold (e.g. a persona's `expected_requirements[].req_id`). /\n reqId: string\n /* Human-readable description of the required deliverable. /\n title: string\n /* Optional kind/category hint, matched against a produced item's kind. /\n category?: string\n /* What produced state satisfies this requirement. Defaults to 'any'. /\n satisfiedBy?: SatisfiedBy\n}\n\nexport interface TaskGold {\n taskId: string\n requirements: CompletionRequirement[]\n}\n\nexport interface ProducedProposal {\n id: string\n title: string\n status: 'pending' \| 'approved' \| 'rejected'\n /* Optional persisted body — when present, enables a correctness check. /\n content?: string\n}\n\n/* Everything observable about what a run actually produced. /\nexport interface ProducedState {\n /* Persisted vault artifacts. Reuses the shared `Artifact` shape. /\n artifacts: Artifact[]\n /* Proposals / filings the agent created. /\n proposals: ProducedProposal[]\n /* Names of tools the agent invoked. /\n toolCalls: string[]\n}\n\nexport interface RequirementCheck {\n reqId: string\n title: string\n /* A produced item of the right kind matched the requirement, non-empty. /\n structurallyPresent: boolean\n /\n Whether the matched item actually fulfils the requirement. `null` when\n * not structurally present, when the matched item carries no content\n * to assess, or when the correctness check itself failed (`unmeasured`).\n /\n correct: boolean \| null\n /* structurallyPresent && !unmeasured && correct !== false. /\n satisfied: boolean\n /\n Set when the correctness check itself errored (LLM call failure or an\n * unparseable response after retry). The requirement's fulfilment is\n * UNKNOWN — `correct` stays null, `satisfied` is false, and\n * `completionVerdict` excludes the row from `completionRate`'s\n * denominator. Never folded into a zero: a synthetic zero is\n * indistinguishable from a real failure (see `JudgeParseError`).\n /\n unmeasured?: true\n /* Why the correctness check could not be measured (present iff `unmeasured`). /\n unmeasuredReason?: string\n /* Human-readable evidence for the verdict. /\n evidence: string[]\n}\n\n/* Extends the substrate verdict spine: `valid` = `fullyComplete` and\n * `score` = `completionRate` — derived in `completionVerdict()`, the one\n * place those equalities hold by construction. /\nexport interface CompletionVerdict extends DefaultVerdict {\n taskId: string\n requirements: RequirementCheck[]\n /* satisfied / MEASURABLE requirements (unmeasured rows leave the denominator). /\n completionRate: number\n /* Every measurable requirement satisfied (false when anything is unmeasured). /\n fullyComplete: boolean\n /* Requirements whose correctness check errored — reported, never scored as zero. /\n unmeasuredCount: number\n}\n\n/\n Construct a `CompletionVerdict` from the per-requirement checks, deriving\n * `completionRate` / `fullyComplete` and the spine fields (`valid` =\n * `fullyComplete`, `score` = `completionRate`) in one place. Throws on zero\n * requirements — a verdict over nothing is a misconfiguration, mirroring\n * `verifyCompletion`'s gold-spec guard.\n /\nexport function completionVerdict(input: {\n taskId: string\n requirements: RequirementCheck[]\n /* What certified the correctness stage, when anything did. Omitted =\n * an uncertified verdict — the honest default for a bare checker. /\n certification?: VerdictCertification\n}): CompletionVerdict {\n if (input.requirements.length === 0) {\n throw new Error(\n `completionVerdict: task '${input.taskId}' has no requirement checks — nothing to derive a verdict from`,\n )\n }\n const measurable = input.requirements.filter((r) => !r.unmeasured)\n const unmeasuredCount = input.requirements.length - measurable.length\n if (measurable.length === 0) {\n // Every check errored: this is an infrastructure failure, not a scored\n // run. A 0-rate verdict here would be a fabricated measurement.\n throw new Error(\n `completionVerdict: task '${input.taskId}' has no measurable requirements — all ${input.requirements.length} correctness checks failed (${input.requirements[0]?.unmeasuredReason ?? 'unknown reason'})`,\n )\n }\n const satisfiedCount = measurable.filter((r) => r.satisfied).length\n const completionRate = satisfiedCount / measurable.length\n // A run with unmeasured rows can still report a rate over what WAS\n // measured, but must not claim full completion.\n const fullyComplete = unmeasuredCount === 0 && satisfiedCount === measurable.length\n return {\n taskId: input.taskId,\n requirements: input.requirements,\n completionRate,\n fullyComplete,\n unmeasuredCount,\n valid: fullyComplete,\n score: completionRate,\n ...(input.certification === undefined ? {} : { certification: input.certification }),\n }\n}\n\n/\n What a correctness checker declares about itself so the completion\n * verdict can carry a certification: which strategy member it discharges,\n * its exact identity, and the steps its answers rest on unverified.\n /\nexport interface CorrectnessCheckerAttestation {\n strategy: VerificationStrategySource\n checker: CheckerIdentity\n assumptions: string[]\n}\n\n/\n Decides whether a produced item's content actually fulfils a requirement.\n * Injected so the structural verifier stays pure and unit-testable; the\n * production implementation is `createLlmCorrectnessChecker`.\n \n `attestation` is optional metadata on the function value: a checker that\n * carries one yields certified completion verdicts; a bare function yields\n * the same verdict uncertified. A plain arrow function remains a valid\n * checker.\n /\nexport interface CorrectnessChecker {\n (\n requirement: CompletionRequirement,\n content: string,\n ): Promise<{ correct: boolean; reason: string }>\n attestation?: CorrectnessCheckerAttestation\n}\n\nconst STOPWORDS = new Set([\n 'the',\n 'a',\n 'an',\n 'of',\n 'for',\n 'and',\n 'or',\n 'to',\n 'in',\n 'on',\n 'with',\n 'by',\n])\n\n// Deliverable-FORM vocabulary — words that name the SHAPE of an output, not its\n// domain content. A correct \"swap-comparison view persisted as a ui/* artifact\"\n// is an OpenUI JSON whose body says nothing about \"artifact\" / \"persisted\" /\n// \"view\"; the discriminative tokens are the domain nouns (swap, comparison).\n// Stripped from the REQUIREMENT side of structural recall so a deliverable is\n// matched on what it IS about, not on the boilerplate describing its form. The\n// correctness checker strips the same class via TITLE_STOPWORDS. Anti-game holds:\n// the distinctive domain tokens remain, so an off-topic item still fails.\nconst REQUIREMENT_FORM_STOPWORDS = new Set([\n 'generated',\n 'generate',\n 'view',\n 'render',\n 'rendered',\n 'persisted',\n 'persist',\n 'artifact',\n 'file',\n 'document',\n 'note',\n 'proposal',\n 'deliverable',\n 'output',\n 'created',\n 'create',\n 'produce',\n 'produced',\n 'flag',\n])\n\nconst MATCH_THRESHOLD = 0.5\nconst MIN_CONTENT_CHARS = 50\n\nfunction tokens(s: string, extraStop?: Set<string>): Set<string> {\n return new Set(\n s\n .toLowerCase()\n .split(/[^a-z0-9]+/)\n .filter((t) => t.length > 1 && !STOPWORDS.has(t) && !extraStop?.has(t)),\n )\n}\n\n/*\n Recall of the requirement's tokens within a candidate's identifying text.\n * Recall, not Jaccard — a candidate's path/id legitimately carries extra\n * tokens the requirement does not name. The requirement side drops\n * deliverable-FORM vocabulary so recall keys on the distinctive domain tokens.\n /\nfunction tokenRecall(requirementText: string, candidateText: string): number {\n const req = tokens(requirementText, REQUIREMENT_FORM_STOPWORDS)\n if (req.size === 0) return 0\n const cand = tokens(candidateText)\n let hit = 0\n for (const t of req) if (cand.has(t)) hit++\n return hit / req.size\n}\n\ninterface Candidate {\n reqIndex: number\n /* Unique key for a produced item — each item satisfies at most one requirement. /\n itemKey: string\n score: number\n evidence: string\n /* Content to correctness-check, or null when the matched item has none. /\n content: string \| null\n}\n\nfunction artifactCandidates(\n req: CompletionRequirement,\n reqIndex: number,\n artifacts: Artifact[],\n): Candidate[] {\n const reqText = `${req.title} ${req.category ?? ''}`\n const out: Candidate[] = []\n artifacts.forEach((a, i) => {\n if ((a.content ?? '').trim().length < MIN_CONTENT_CHARS) return\n // Match against the artifact CONTENT too, not just its path + kind — a\n // generated view / note whose path is generic still satisfies a requirement\n // when its body covers it (e.g. an OpenUI comparison grounded in the on-file\n // figures). Bounded slice keeps the recall text cheap; MATCH_THRESHOLD holds.\n let score = tokenRecall(\n reqText,\n `${a.path ?? ''} ${a.kind} ${(a.content ?? '').slice(0, 4000)}`,\n )\n if (req.category && a.kind && req.category.toLowerCase() === a.kind.toLowerCase()) {\n score = Math.max(score, 1)\n }\n if (score < MATCH_THRESHOLD) return\n out.push({\n reqIndex,\n itemKey: `artifact:${i}`,\n score,\n evidence: `artifact '${a.path ?? a.kind}' matched (token recall ${score.toFixed(2)})`,\n content: a.content ?? null,\n })\n })\n return out\n}\n\nfunction proposalCandidates(\n req: CompletionRequirement,\n reqIndex: number,\n proposals: ProducedProposal[],\n): Candidate[] {\n const reqText = `${req.title} ${req.category ?? ''}`\n const out: Candidate[] = []\n for (const p of proposals) {\n // Pending or rejected work is not a completed deliverable.\n if (p.status !== 'approved') continue\n // A proposal needs an assessable BODY to be a deliverable. A bare title is\n // not completion: correctness cannot be judged on it, so a title-only match\n // would auto-pass the oracle (structurallyPresent && correct===null →\n // satisfied) with no verifiable content. Tool calls are the only\n // legitimately content-less deliverable (`toolCallCandidates`).\n const body = (p.content ?? '').trim()\n if (body.length < MIN_CONTENT_CHARS) continue\n // Match against the body as well as the (often short) title — a refusal /\n // flag / analysis proposal whose title is a label still satisfies a\n // descriptively-worded requirement when its content covers it. MATCH_THRESHOLD\n // + the requirement's distinctive tokens keep an off-topic proposal out;\n // correctness (a SEMANTIC checker, NOT this lexical pass) then judges\n // polarity/fulfilment, so a negation that merely contains the tokens fails.\n // Structural and correctness must use different evidence or the two-stage\n // check collapses to one lexical gate.\n const score = tokenRecall(reqText, `${p.title} ${body}`)\n if (score < MATCH_THRESHOLD) continue\n out.push({\n reqIndex,\n itemKey: `proposal:${p.id}`,\n score,\n evidence: `approved proposal '${p.title}' matched (token recall ${score.toFixed(2)})`,\n content: body,\n })\n }\n return out\n}\n\nfunction toolCallCandidates(\n req: CompletionRequirement,\n reqIndex: number,\n toolCalls: string[],\n): Candidate[] {\n const out: Candidate[] = []\n toolCalls.forEach((name, i) => {\n const score = tokenRecall(req.title, name)\n if (score < MATCH_THRESHOLD) return\n out.push({\n reqIndex,\n itemKey: `tool:${i}`,\n score,\n evidence: `tool call '${name}' matched (token recall ${score.toFixed(2)})`,\n content: null,\n })\n })\n return out\n}\n\n/\n Verify whether a run completed the task. `checkCorrectness` is injected —\n * `createLlmCorrectnessChecker` for production, a deterministic stub in tests.\n \n Throws on a gold spec with no requirements: an eval task that requires\n * nothing is a misconfiguration, not a vacuously-complete task.\n /\nexport async function verifyCompletion(\n gold: TaskGold,\n state: ProducedState,\n checkCorrectness: CorrectnessChecker,\n): Promise<CompletionVerdict> {\n if (gold.requirements.length === 0) {\n throw new Error(\n `verifyCompletion: task '${gold.taskId}' has no requirements — malformed gold spec`,\n )\n }\n\n // Collect every above-threshold (requirement, produced-item) candidate, then\n // assign greedily by descending score: each requirement and each produced\n // item is used at most once. One deliverable fulfils one requirement.\n const candidates: Candidate[] = []\n gold.requirements.forEach((req, i) => {\n const by = req.satisfiedBy ?? 'any'\n if (by === 'artifact' \|\| by === 'any') {\n candidates.push(...artifactCandidates(req, i, state.artifacts))\n }\n if (by === 'proposal' \|\| by === 'any') {\n candidates.push(...proposalCandidates(req, i, state.proposals))\n }\n if (by === 'tool-call' \|\| by === 'any') {\n candidates.push(...toolCallCandidates(req, i, state.toolCalls))\n }\n })\n candidates.sort((a, b) => b.score - a.score)\n\n const assigned = new Map<number, Candidate>()\n const itemTaken = new Set<string>()\n for (const c of candidates) {\n if (assigned.has(c.reqIndex) \|\| itemTaken.has(c.itemKey)) continue\n assigned.set(c.reqIndex, c)\n itemTaken.add(c.itemKey)\n }\n\n const requirements: RequirementCheck[] = []\n for (let i = 0; i < gold.requirements.length; i++) {\n const req = gold.requirements[i]!\n const match = assigned.get(i)\n const evidence: string[] = []\n let correct: boolean \| null = null\n let unmeasuredReason: string \| undefined\n\n if (match) {\n evidence.push(match.evidence)\n if (match.content !== null) {\n try {\n const r = await checkCorrectness(req, match.content)\n correct = r.correct\n evidence.push(`correctness: ${r.correct ? 'pass' : 'fail'} — ${r.reason}`)\n } catch (err) {\n // The CHECKER failed, not the requirement. Recording this as a\n // zero would fabricate a model failure out of an infrastructure\n // one; the requirement is unmeasured and leaves the denominator.\n unmeasuredReason =\n err instanceof JudgeParseError\n ? `checker response unparseable after retry: ${err.raw.slice(0, 200)}`\n : `checker call failed: ${err instanceof Error ? err.message : String(err)}`\n evidence.push(`correctness: UNMEASURED — ${unmeasuredReason}`)\n }\n } else {\n evidence.push('correctness: not assessed — matched item carries no content')\n }\n } else {\n const by = req.satisfiedBy ?? 'any'\n const kind = by === 'any' ? 'artifact/proposal/tool-call' : by\n evidence.push(`no produced ${kind} matched this requirement`)\n }\n\n const structurallyPresent = match !== undefined\n const unmeasured = unmeasuredReason !== undefined\n const satisfied = structurallyPresent && !unmeasured && correct !== false\n requirements.push({\n reqId: req.reqId,\n title: req.title,\n structurallyPresent,\n correct,\n satisfied,\n ...(unmeasured ? { unmeasured: true as const, unmeasuredReason } : {}),\n evidence,\n })\n }\n\n // The verdict blends the deterministic structural stage with the injected\n // correctness stage, so the certification names the checker's own member\n // and carries the structural stage's lexical nature as an assumption. A\n // checker without an attestation yields an uncertified verdict.\n const attestation = checkCorrectness.attestation\n const certification: VerdictCertification \| undefined = attestation\n ? {\n strategy: attestation.strategy,\n checker: attestation.checker,\n assumptions: [\n 'structural matching is lexical token recall over produced items — the correctness stage only sees items it matched',\n ...attestation.assumptions,\n ],\n evidenceDigest: certificationEvidenceDigest({ taskId: gold.taskId, requirements }),\n }\n : undefined\n\n return completionVerdict({\n taskId: gold.taskId,\n requirements,\n ...(certification === undefined ? {} : { certification }),\n })\n}\n\nexport interface LlmCorrectnessCheckerOpts {\n model?: string\n /* Optional ledger for direct use. /\n costLedger?: CostLedgerHandle\n costPhase?: string\n costTags?: Record<string, string>\n signal?: AbortSignal\n /* Max chars of artifact content sent to the checker. /\n maxContentChars?: number\n /\n Checker LLM calls per requirement before giving up (parse failures and\n * call errors both consume attempts). The failure then surfaces as an\n * `unmeasured` requirement, never a zero.\n /\n maxAttempts?: number\n /\n Forensic capture of every checker request/response/error — without it a\n * checker failure is unauditable (the agent-turn raws never contain the\n * checker's own calls). Same sink contract as `LlmClient`.\n /\n rawSink?: RawProviderSink\n}\n\n/\n Parse the correctness checker's model response. Tolerates a response\n * truncated mid-JSON (max_tokens cap) by auto-closing the prefix — the\n * verdict boolean usually lands in the first few tokens, so a recovered\n * prefix with a boolean `correct` is a real measurement, not a guess.\n * Fails loud (JudgeParseError) when no boolean verdict is recoverable.\n /\nexport function parseCorrectnessResponse(raw: string): { correct: boolean; reason: string } {\n const readVerdict = (candidate: unknown): { correct: boolean; reason: string } \| null => {\n if (candidate === null \|\| typeof candidate !== 'object') return null\n const { correct, reason } = candidate as { correct?: unknown; reason?: unknown }\n if (typeof correct !== 'boolean') return null\n return { correct, reason: typeof reason === 'string' ? reason : '' }\n }\n\n const match = raw.match(/\\{[\\s\\S]\\}/)\n if (match) {\n try {\n const strict = readVerdict(JSON.parse(match[0]))\n if (strict) return strict\n } catch {\n // fall through to truncation recovery\n }\n }\n // The strict path needs a closing `}`; a cap-hit response has none. Take\n // everything from the first `{` and auto-close it.\n const start = raw.indexOf('{')\n if (start !== -1) {\n const recovered = readVerdict(recoverTruncatedJson(raw.slice(start)))\n if (recovered) return recovered\n }\n throw new JudgeParseError('correctness-checker', raw)\n}\n\n/*\n Production `CorrectnessChecker` — one LLM call per matched artifact,\n * deterministic (temperature 0), structured JSON out. Judges fulfilment\n * only: a plan, a gesture, or a description of what should be done does not\n * fulfil a requirement — the artifact must BE the deliverable.\n /\nexport function createLlmCorrectnessChecker(\n chat: ChatClient,\n opts: LlmCorrectnessCheckerOpts = {},\n): CorrectnessChecker {\n const model = opts.model ?? 'claude-sonnet-4-6'\n const maxContentChars = opts.maxContentChars ?? 8000\n const maxAttempts = opts.maxAttempts ?? 2\n const costLedger = opts.costLedger ?? new CostLedger()\n const sink = opts.rawSink\n const record = async (event: RawProviderEvent): Promise<void> => {\n // Forensic capture is best-effort; the verdict is the system of record.\n try {\n await sink?.record(event)\n } catch {\n // Intentionally swallowed.\n }\n }\n const checker: CorrectnessChecker = async (requirement, content) => {\n const request = {\n model,\n messages: [\n {\n role: 'system' as const,\n content:\n 'You verify whether a produced work artifact actually fulfils a stated requirement. Judge fulfilment only — is the deliverable substantively present and on-point — not polish. A plan to do it later, a vague gesture, or a description of what should be done does NOT fulfil a requirement; the artifact must BE the deliverable. Respond with a single JSON object: {\"correct\": boolean, \"reason\": string (<= 30 words)}.',\n },\n {\n role: 'user' as const,\n content: `Requirement: ${requirement.title}\\n${\n requirement.category ? `Category: ${requirement.category}\\n` : ''\n }\\nProduced artifact:\\n${content.slice(0, maxContentChars)}`,\n },\n ],\n temperature: 0,\n maxTokens: 200,\n } satisfies ChatRequest\n let lastErr: unknown\n for (let attempt = 0; attempt < maxAttempts; attempt++) {\n const started = Date.now()\n await record({\n eventId: randomUUID(),\n provider: chat.transport,\n model,\n endpoint: '/chat',\n baseUrl: '',\n attemptIndex: attempt,\n direction: 'request',\n timestamp: started,\n requestBody: request,\n redactedFields: [],\n })\n try {\n const paid = await costLedger.runPaidCall({\n channel: 'verifier',\n phase: opts.costPhase ?? 'completion.correctness',\n actor: 'correctness-checker',\n model,\n maximumCharge:\n chat.maximumAttempts === undefined\n ? undefined\n : maximumChargeForLlmRequest(request, { maximumAttempts: chat.maximumAttempts }),\n tags: {\n ...opts.costTags,\n requirementId: requirement.reqId,\n attempt: String(attempt),\n },\n signal: opts.signal,\n execute: (signal, callId) => chat.chat(request, { signal, idempotencyKey: callId }),\n receipt: costReceiptFromLlm,\n receiptFromError: costReceiptFromLlmError,\n })\n if (!paid.succeeded) throw paid.error\n const resp = paid.value\n // Hold the transport to its own word: a verdict produced by a model\n // other than the one requested is not this model's verdict. A\n // transport that echoes no id cannot be made to prove identity here —\n // callers needing that proof enable it at the client\n // (`LlmClientOptions.assertServedModel`) or gate on `assertModelsServed`.\n assertServedModel(model, resp.servedModel, {\n allowUnreported: true,\n context: `correctness checker for requirement ${requirement.reqId}`,\n })\n const raw = resp.content\n await record({\n eventId: randomUUID(),\n provider: chat.transport,\n model,\n endpoint: '/chat',\n baseUrl: '',\n attemptIndex: attempt,\n direction: 'response',\n timestamp: Date.now(),\n durationMs: Date.now() - started,\n responseBody: resp,\n redactedFields: [],\n })\n return parseCorrectnessResponse(raw)\n } catch (err) {\n lastErr = err\n await record({\n eventId: randomUUID(),\n provider: chat.transport,\n model,\n endpoint: '/chat',\n baseUrl: '',\n attemptIndex: attempt,\n direction: 'error',\n timestamp: Date.now(),\n durationMs: Date.now() - started,\n errorMessage: err instanceof Error ? err.message : String(err),\n redactedFields: [],\n })\n // Substitution is a verdict about identity, not a transient fault:\n // another attempt reaches the same wrong model and spends more tokens.\n if (err instanceof ModelSubstitutionError) throw err\n }\n }\n throw lastErr instanceof Error ? lastErr : new Error(String(lastErr))\n }\n checker.attestation = {\n strategy: 'judge',\n checker: { name: 'llm-correctness-checker', version: model },\n assumptions: [\n 'served-model identity is accepted when the transport reports none (assertServedModel allowUnreported)',\n ],\n }\n return checker\n}\n\n/* Stopwords for requirement-title tokenization — drops the imperative verbs\n * ('review', 'update', …) common to deliverable titles so recall keys on the\n * substantive nouns, not the boilerplate ask. /\nconst TITLE_STOPWORDS = new Set([\n 'the',\n 'a',\n 'an',\n 'and',\n 'or',\n 'for',\n 'to',\n 'of',\n 'in',\n 'on',\n 'with',\n 'review',\n 'update',\n 'new',\n 'proposed',\n])\n\n/\n Deterministic `CorrectnessChecker` — the no-LLM counterpart to\n * `createLlmCorrectnessChecker`. A produced item fulfils a requirement when its\n * content is substantive (≥ `minContentLength` chars) AND recalls ≥ `minRecall`\n * of the requirement title's significant tokens. No network.\n \n Polarity-blind: token recall credits a negation that contains the\n * requirement's tokens (\"I will NOT produce the comparison\" recalls every token\n * of \"produce the comparison\"). The structural match stage is ALSO lexical, so\n * pairing the two collapses to a single gameable gate. Use this only as an\n * opt-in structural pre-filter or for tasks whose requirements have no polarity\n * to invert; for produced-state grading the correctness checker MUST be semantic\n * (`createLlmCorrectnessChecker`). See the anti-game fixtures in the test suite.\n /\nexport function createTokenRecallChecker(\n opts: { minRecall?: number; minContentLength?: number } = {},\n): CorrectnessChecker {\n const minRecall = opts.minRecall ?? 0.5\n const minLen = opts.minContentLength ?? 120\n const checker: CorrectnessChecker = async (requirement, content) => {\n const body = content.trim()\n if (body.length < minLen)\n return {\n correct: false,\n reason: `content too thin (${body.length} chars) to be the deliverable`,\n }\n const titleTokens = requirement.title\n .toLowerCase()\n .split(/[^a-z0-9]+/)\n .filter((t) => t.length > 2 && !TITLE_STOPWORDS.has(t))\n if (titleTokens.length === 0)\n return {\n correct: true,\n reason: 'requirement title has no significant tokens — structural match accepted',\n }\n const lower = body.toLowerCase()\n const hits = titleTokens.filter((t) => lower.includes(t)).length\n const recall = hits / titleTokens.length\n return recall >= minRecall\n ? {\n correct: true,\n reason: `content recalls ${hits}/${titleTokens.length} requirement tokens`,\n }\n : {\n correct: false,\n reason: `content recalls only ${hits}/${titleTokens.length} requirement tokens`,\n }\n }\n // 'schema': token recall checks shape, not meaning — the member whose\n // documented failure mode (\"a well-formed wrong answer passes\") is\n // exactly this checker's polarity blindness.\n checker.attestation = {\n strategy: 'schema',\n checker: { name: 'token-recall-checker', version: packageVersion() },\n assumptions: ['polarity-blind: a negation that recalls the requirement tokens passes'],\n }\n return checker\n}\n","/\n Produced-state extraction — normalize a run's runtime event stream into the\n * typed `ProducedState` the completion oracle consumes.\n \n `ProducedState` answers \"what did the agent actually produce\" — vault\n * artifacts, proposals, tool calls. The runtime emits these as a stream of\n * events; this module is the single normalization point from that stream to\n * the shape `verifyCompletion` expects.\n \n Input is structurally typed (`RuntimeEventLike`) so this module does not\n * depend on agent-runtime — agent-runtime's `RuntimeStreamEvent` satisfies it\n * structurally. The `content` on `ArtifactEventLike` and the whole\n * `proposal_created` variant are the runtime-side enrichments this contract\n * requires; the runtime emits them, this module consumes them.\n /\n\nimport type { Artifact } from './artifact-validator'\nimport type { ProducedProposal, ProducedState } from './completion-verifier'\n\n/* A tool the agent invoked. /\nexport interface ToolCallEventLike {\n type: 'tool_call'\n toolName: string\n}\n\n/\n An artifact the agent produced. `content` is the enriched field — the\n * runtime's base `artifact` event carries only metadata; the completion\n * oracle needs the body to verify the deliverable, so the runtime emits it.\n /\nexport interface ArtifactEventLike {\n type: 'artifact'\n artifactId: string\n name?: string\n mimeType?: string\n uri?: string\n content?: string\n}\n\n/* A proposal / filing the agent created. /\nexport interface ProposalEventLike {\n type: 'proposal_created'\n proposalId: string\n title: string\n status?: 'pending' \| 'approved' \| 'rejected'\n // body of the proposal (e.g. a submit_proposal `description`). When present,\n // the completion oracle correctness-checks it like artifact content; absent,\n // the proposal is graded presence-only.\n content?: string\n}\n\n/\n The subset of runtime stream events `extractProducedState` consumes.\n * agent-runtime's full `RuntimeStreamEvent` union satisfies this structurally;\n * the `{ type: string }` catch-all keeps the input permissive so callers can\n * pass the whole unfiltered telemetry stream — unrecognized events are skipped.\n /\nexport type RuntimeEventLike =\n \| ToolCallEventLike\n \| ArtifactEventLike\n \| ProposalEventLike\n \| { type: string }\n\nfunction artifactKind(mimeType: string \| undefined): string {\n if (!mimeType) return 'file'\n if (mimeType.includes('json')) return 'json'\n if (mimeType.startsWith('text/')) return 'text'\n return 'file'\n}\n\n/\n Normalize an emission-ordered stream into its latest observed produced state.\n * Artifacts are keyed by their exact output path (name, then URI, then id);\n * proposals by id. Later observations replace earlier ones, including missing\n * content or a rejected status. An obsolete version cannot prove completion.\n * Distinct paths remain distinct; no path normalization or aliasing is inferred.\n * Results retain first-seen identity order. Tool names describe invocation,\n * not success, and are deduplicated in first-seen order.\n \n An artifact without observed content yields empty content, which the\n * completion oracle rejects. This projection does not verify external effects,\n * recover dropped events, or establish freshness beyond the supplied stream.\n */\nexport function extractProducedState(events: readonly RuntimeEventLike[]): ProducedState {\n const artifacts = new Map<string, Artifact>()\n const proposals = new Map<string, ProducedProposal>()\n const toolCalls: string[] = []\n const seenTools = new Set<string>()\n\n for (const ev of events) {\n if (ev.type === 'tool_call') {\n const name = (ev as ToolCallEventLike).toolName\n if (name && !seenTools.has(name)) {\n seenTools.add(name)\n toolCalls.push(name)\n }\n } else if (ev.type === 'artifact') {\n const a = ev as ArtifactEventLike\n const path = a.name ?? a.uri ?? a.artifactId\n artifacts.set(path, {\n kind: artifactKind(a.mimeType),\n path,\n content: a.content ?? '',\n })\n } else if (ev.type === 'proposal_created') {\n const p = ev as ProposalEventLike\n proposals.set(p.proposalId, {\n id: p.proposalId,\n title: p.title,\n status: p.status ?? 'pending',\n ...(p.content !== undefined ? { content: p.content } : {}),\n })\n }\n }\n\n return { artifacts: [...artifacts.values()], proposals: [...proposals.values()], toolCalls }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAeA,MAAa,mBAA2C;CACtD;CACA;CACA;CACA;AACF;;;;;;AAwBA,MAAa,uBAAuB;;;;;;;;;;;;;;;;;;;AAoBpC,SAAgB,kBAAkB,MAAuC;CACvE,MAAM,YAAY,KAAK,aAAa;CACpC,IAAI,UAAU,WAAW,GAAG,MAAM,IAAI,gBAAgB,0CAA0C;CAChG,MAAM,YAAY,KAAK,KAAK,OAAO;CACnC,MAAM,SAAS,KAAK,WAAW,YAAY,CAAC,SAAS,IAAI,CAAC;CAC1D,IAAI,OAAO,WAAW,GACpB,MAAM,IAAI,gBACR,kGACF;CAEF,MAAM,MAAsB,CAAC;CAC7B,MAAM,uBAAO,IAAI,IAAY;CAC7B,KAAK,MAAM,WAAW,WAAW;EAS/B,MAAM,YAAY,KAAK,mBACnB,SACA,OAAO,QAAQ,UAAU,qBAAqB,SAAS,KAAK,CAAC;EACjE,MAAM,YAAY,UAAU,SAAS,IAAI,YAAY,CAAC,oBAAoB;EAC1E,KAAK,MAAM,SAAS,WAAW;GAC7B,MAAM,UAAwB;IAC5B,GAAG,KAAK;IACR;IACA,MAAM,GAAG,KAAK,KAAK,QAAQ,QAAQ,GAAG,QAAQ,GAAG;IACjD,OAAO;KAAE,GAAG,KAAK,KAAK;KAAO,SAAS;IAAM;IAC5C,UAAU;KAAE,GAAI,KAAK,KAAK,YAAY,CAAC;KAAI;KAAS,cAAc;IAAM;GAC1E;GACA,MAAM,KAAK,eAAe,OAAO;GACjC,IAAI,KAAK,IAAI,EAAE,GAAG;GAClB,KAAK,IAAI,EAAE;GACX,IAAI,KAAK,OAAO;EAClB;CACF;CACA,IAAI,IAAI,WAAW,GAGjB,MAAM,IAAI,gBACR,uDAAuD,UAAU,KAAK,IAAI,EAAE,aAAa,OAAO,KAAK,IAAI,EAAE,IAC7G;CAEF,OAAO;AACT;;;;;;;;AASA,SAAgB,cACd,SACqD;CACrD,MAAM,IAAI,QAAQ;CAClB,MAAM,UAAU,GAAG;CACnB,MAAM,QAAQ,GAAG;CACjB,IAAI,OAAO,YAAY,YAAY,OAAO,UAAU,UAClD,OAAO;EAAW;EAAwB;CAAM;AAGpD;;;;;;;;AASA,SAAgB,eAAe,SAA+B;CAE5D,OAAO,GADO,qBAAqB,yBAAyB,OAAO,CAAC,KAAK,UACzD,GAAG,iBAAiB,OAAO,CAAC,CAAC,MAAM,GAAG,EAAE;AAC1D;;;;;AAMA,SAAgB,oBAAoB,SAA+B;CACjE,MAAM,QAAQ,QAAQ,OAAO,SAAS,KAAK;CAC3C,IAAI,CAAC,OAAO;EACV,MAAM,QAAQ,yBAAyB,OAAO,KAAK;EACnD,MAAM,IAAI,gBACR,iBAAiB,MAAM,gDACzB;CACF;CACA,OAAO;AACT;AAEA,SAAS,yBAAyB,SAA2C;CAC3E,OAAO,QAAQ,MAAM,KAAK,KAAK,QAAQ,SAAS,KAAK,KAAK,KAAA;AAC5D;AAEA,SAAS,qBAAqB,OAA+C;CAM3E,OALa,OACT,KAAK,CAAC,CACP,QAAQ,qBAAqB,GAAG,CAAC,CACjC,QAAQ,OAAO,GAAG,CAAC,CACnB,QAAQ,UAAU,EAAE,KACR,KAAA;AACjB;AAEA,SAAS,QAA2C,OAAsB;CACxE,MAAM,MAA+B,CAAC;CACtC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,KAAK,GAC7C,IAAI,UAAU,KAAA,GAAW,IAAI,OAAO;CAEtC,OAAO;AACT;;;;;;;;;;;AAYA,SAAgB,iBAAiB,SAA+B;CAC9D,MAAM,QAAQ,oBAAoB,OAAO;CACzC,MAAM,YAAY,QAAQ;EACxB,GAAG;EACH,MAAM,KAAA;EACN,aAAa,KAAA;EACb,MAAM,QAAQ,OAAO,CAAC,GAAG,QAAQ,IAAI,CAAC,CAAC,KAAK,IAAI,KAAA;EAChD,OAAO,QAAQ;GAAE,GAAG,QAAQ;GAAO,SAAS;EAAM,CAAC;CACrD,CAAC;CACD,OAAO,WAAW,QAAQ,CAAC,CAAC,OAAO,gBAAgB,SAAS,CAAC,CAAC,CAAC,OAAO,KAAK;AAC7E;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACtEA,SAAgB,kBAAkB,OAMZ;CACpB,IAAI,MAAM,aAAa,WAAW,GAChC,MAAM,IAAI,MACR,4BAA4B,MAAM,OAAO,+DAC3C;CAEF,MAAM,aAAa,MAAM,aAAa,QAAQ,MAAM,CAAC,EAAE,UAAU;CACjE,MAAM,kBAAkB,MAAM,aAAa,SAAS,WAAW;CAC/D,IAAI,WAAW,WAAW,GAGxB,MAAM,IAAI,MACR,4BAA4B,MAAM,OAAO,yCAAyC,MAAM,aAAa,OAAO,8BAA8B,MAAM,aAAa,EAAE,EAAE,oBAAoB,iBAAiB,EACxM;CAEF,MAAM,iBAAiB,WAAW,QAAQ,MAAM,EAAE,SAAS,CAAC,CAAC;CAC7D,MAAM,iBAAiB,iBAAiB,WAAW;CAGnD,MAAM,gBAAgB,oBAAoB,KAAK,mBAAmB,WAAW;CAC7E,OAAO;EACL,QAAQ,MAAM;EACd,cAAc,MAAM;EACpB;EACA;EACA;EACA,OAAO;EACP,OAAO;EACP,GAAI,MAAM,kBAAkB,KAAA,IAAY,CAAC,IAAI,EAAE,eAAe,MAAM,cAAc;CACpF;AACF;AA+BA,MAAM,4BAAY,IAAI,IAAI;CACxB;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAUD,MAAM,6CAA6B,IAAI,IAAI;CACzC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAED,MAAM,kBAAkB;AACxB,MAAM,oBAAoB;AAE1B,SAAS,OAAO,GAAW,WAAsC;CAC/D,OAAO,IAAI,IACT,EACG,YAAY,CAAC,CACb,MAAM,YAAY,CAAC,CACnB,QAAQ,MAAM,EAAE,SAAS,KAAK,CAAC,UAAU,IAAI,CAAC,KAAK,CAAC,WAAW,IAAI,CAAC,CAAC,CAC1E;AACF;;;;;;;AAQA,SAAS,YAAY,iBAAyB,eAA+B;CAC3E,MAAM,MAAM,OAAO,iBAAiB,0BAA0B;CAC9D,IAAI,IAAI,SAAS,GAAG,OAAO;CAC3B,MAAM,OAAO,OAAO,aAAa;CACjC,IAAI,MAAM;CACV,KAAK,MAAM,KAAK,KAAK,IAAI,KAAK,IAAI,CAAC,GAAG;CACtC,OAAO,MAAM,IAAI;AACnB;AAYA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,UAAU,GAAG,IAAI,MAAM,GAAG,IAAI,YAAY;CAChD,MAAM,MAAmB,CAAC;CAC1B,UAAU,SAAS,GAAG,MAAM;EAC1B,KAAK,EAAE,WAAW,GAAA,CAAI,KAAK,CAAC,CAAC,SAAS,mBAAmB;EAKzD,IAAI,QAAQ,YACV,SACA,GAAG,EAAE,QAAQ,GAAG,GAAG,EAAE,KAAK,IAAI,EAAE,WAAW,GAAA,CAAI,MAAM,GAAG,GAAI,GAC9D;EACA,IAAI,IAAI,YAAY,EAAE,QAAQ,IAAI,SAAS,YAAY,MAAM,EAAE,KAAK,YAAY,GAC9E,QAAQ,KAAK,IAAI,OAAO,CAAC;EAE3B,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,YAAY;GACrB;GACA,UAAU,aAAa,EAAE,QAAQ,EAAE,KAAK,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACnF,SAAS,EAAE,WAAW;EACxB,CAAC;CACH,CAAC;CACD,OAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,UAAU,GAAG,IAAI,MAAM,GAAG,IAAI,YAAY;CAChD,MAAM,MAAmB,CAAC;CAC1B,KAAK,MAAM,KAAK,WAAW;EAEzB,IAAI,EAAE,WAAW,YAAY;EAM7B,MAAM,QAAQ,EAAE,WAAW,GAAA,CAAI,KAAK;EACpC,IAAI,KAAK,SAAS,mBAAmB;EASrC,MAAM,QAAQ,YAAY,SAAS,GAAG,EAAE,MAAM,GAAG,MAAM;EACvD,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,YAAY,EAAE;GACvB;GACA,UAAU,sBAAsB,EAAE,MAAM,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACnF,SAAS;EACX,CAAC;CACH;CACA,OAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;CACb,MAAM,MAAmB,CAAC;CAC1B,UAAU,SAAS,MAAM,MAAM;EAC7B,MAAM,QAAQ,YAAY,IAAI,OAAO,IAAI;EACzC,IAAI,QAAQ,iBAAiB;EAC7B,IAAI,KAAK;GACP;GACA,SAAS,QAAQ;GACjB;GACA,UAAU,cAAc,KAAK,0BAA0B,MAAM,QAAQ,CAAC,EAAE;GACxE,SAAS;EACX,CAAC;CACH,CAAC;CACD,OAAO;AACT;;;;;;;;AASA,eAAsB,iBACpB,MACA,OACA,kBAC4B;CAC5B,IAAI,KAAK,aAAa,WAAW,GAC/B,MAAM,IAAI,MACR,2BAA2B,KAAK,OAAO,4CACzC;CAMF,MAAM,aAA0B,CAAC;CACjC,KAAK,aAAa,SAAS,KAAK,MAAM;EACpC,MAAM,KAAK,IAAI,eAAe;EAC9B,IAAI,OAAO,cAAc,OAAO,OAC9B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;EAEhE,IAAI,OAAO,cAAc,OAAO,OAC9B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;EAEhE,IAAI,OAAO,eAAe,OAAO,OAC/B,WAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;CAElE,CAAC;CACD,WAAW,MAAM,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;CAE3C,MAAM,2BAAW,IAAI,IAAuB;CAC5C,MAAM,4BAAY,IAAI,IAAY;CAClC,KAAK,MAAM,KAAK,YAAY;EAC1B,IAAI,SAAS,IAAI,EAAE,QAAQ,KAAK,UAAU,IAAI,EAAE,OAAO,GAAG;EAC1D,SAAS,IAAI,EAAE,UAAU,CAAC;EAC1B,UAAU,IAAI,EAAE,OAAO;CACzB;CAEA,MAAM,eAAmC,CAAC;CAC1C,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,aAAa,QAAQ,KAAK;EACjD,MAAM,MAAM,KAAK,aAAa;EAC9B,MAAM,QAAQ,SAAS,IAAI,CAAC;EAC5B,MAAM,WAAqB,CAAC;EAC5B,IAAI,UAA0B;EAC9B,IAAI;EAEJ,IAAI,OAAO;GACT,SAAS,KAAK,MAAM,QAAQ;GAC5B,IAAI,MAAM,YAAY,MACpB,IAAI;IACF,MAAM,IAAI,MAAM,iBAAiB,KAAK,MAAM,OAAO;IACnD,UAAU,EAAE;IACZ,SAAS,KAAK,gBAAgB,EAAE,UAAU,SAAS,OAAO,KAAK,EAAE,QAAQ;GAC3E,SAAS,KAAK;IAIZ,mBACE,eAAe,kBACX,6CAA6C,IAAI,IAAI,MAAM,GAAG,GAAG,MACjE,wBAAwB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IAC7E,SAAS,KAAK,6BAA6B,kBAAkB;GAC/D;QAEA,SAAS,KAAK,6DAA6D;EAE/E,OAAO;GACL,MAAM,KAAK,IAAI,eAAe;GAC9B,MAAM,OAAO,OAAO,QAAQ,gCAAgC;GAC5D,SAAS,KAAK,eAAe,KAAK,0BAA0B;EAC9D;EAEA,MAAM,sBAAsB,UAAU,KAAA;EACtC,MAAM,aAAa,qBAAqB,KAAA;EACxC,MAAM,YAAY,uBAAuB,CAAC,cAAc,YAAY;EACpE,aAAa,KAAK;GAChB,OAAO,IAAI;GACX,OAAO,IAAI;GACX;GACA;GACA;GACA,GAAI,aAAa;IAAE,YAAY;IAAe;GAAiB,IAAI,CAAC;GACpE;EACF,CAAC;CACH;CAMA,MAAM,cAAc,iBAAiB;CACrC,MAAM,gBAAkD,cACpD;EACE,UAAU,YAAY;EACtB,SAAS,YAAY;EACrB,aAAa,CACX,sHACA,GAAG,YAAY,WACjB;EACA,gBAAgB,4BAA4B;GAAE,QAAQ,KAAK;GAAQ;EAAa,CAAC;CACnF,IACA,KAAA;CAEJ,OAAO,kBAAkB;EACvB,QAAQ,KAAK;EACb;EACA,GAAI,kBAAkB,KAAA,IAAY,CAAC,IAAI,EAAE,cAAc;CACzD,CAAC;AACH;;;;;;;;AAgCA,SAAgB,yBAAyB,KAAmD;CAC1F,MAAM,eAAe,cAAoE;EACvF,IAAI,cAAc,QAAQ,OAAO,cAAc,UAAU,OAAO;EAChE,MAAM,EAAE,SAAS,WAAW;EAC5B,IAAI,OAAO,YAAY,WAAW,OAAO;EACzC,OAAO;GAAE;GAAS,QAAQ,OAAO,WAAW,WAAW,SAAS;EAAG;CACrE;CAEA,MAAM,QAAQ,IAAI,MAAM,aAAa;CACrC,IAAI,OACF,IAAI;EACF,MAAM,SAAS,YAAY,KAAK,MAAM,MAAM,EAAE,CAAC;EAC/C,IAAI,QAAQ,OAAO;CACrB,QAAQ,CAER;CAIF,MAAM,QAAQ,IAAI,QAAQ,GAAG;CAC7B,IAAI,UAAU,IAAI;EAChB,MAAM,YAAY,YAAY,qBAAqB,IAAI,MAAM,KAAK,CAAC,CAAC;EACpE,IAAI,WAAW,OAAO;CACxB;CACA,MAAM,IAAI,gBAAgB,uBAAuB,GAAG;AACtD;;;;;;;AAQA,SAAgB,4BACd,MACA,OAAkC,CAAC,GACf;CACpB,MAAM,QAAQ,KAAK,SAAS;CAC5B,MAAM,kBAAkB,KAAK,mBAAmB;CAChD,MAAM,cAAc,KAAK,eAAe;CACxC,MAAM,aAAa,KAAK,cAAc,IAAI,WAAW;CACrD,MAAM,OAAO,KAAK;CAClB,MAAM,SAAS,OAAO,UAA2C;EAE/D,IAAI;GACF,MAAM,MAAM,OAAO,KAAK;EAC1B,QAAQ,CAER;CACF;CACA,MAAM,UAA8B,OAAO,aAAa,YAAY;EAClE,MAAM,UAAU;GACd;GACA,UAAU,CACR;IACE,MAAM;IACN,SACE;GACJ,GACA;IACE,MAAM;IACN,SAAS,gBAAgB,YAAY,MAAM,IACzC,YAAY,WAAW,aAAa,YAAY,SAAS,MAAM,GAChE,wBAAwB,QAAQ,MAAM,GAAG,eAAe;GAC3D,CACF;GACA,aAAa;GACb,WAAW;EACb;EACA,IAAI;EACJ,KAAK,IAAI,UAAU,GAAG,UAAU,aAAa,WAAW;GACtD,MAAM,UAAU,KAAK,IAAI;GACzB,MAAM,OAAO;IACX,SAAS,WAAW;IACpB,UAAU,KAAK;IACf;IACA,UAAU;IACV,SAAS;IACT,cAAc;IACd,WAAW;IACX,WAAW;IACX,aAAa;IACb,gBAAgB,CAAC;GACnB,CAAC;GACD,IAAI;IACF,MAAM,OAAO,MAAM,WAAW,YAAY;KACxC,SAAS;KACT,OAAO,KAAK,aAAa;KACzB,OAAO;KACP;KACA,eACE,KAAK,oBAAoB,KAAA,IACrB,KAAA,IACA,2BAA2B,SAAS,EAAE,iBAAiB,KAAK,gBAAgB,CAAC;KACnF,MAAM;MACJ,GAAG,KAAK;MACR,eAAe,YAAY;MAC3B,SAAS,OAAO,OAAO;KACzB;KACA,QAAQ,KAAK;KACb,UAAU,QAAQ,WAAW,KAAK,KAAK,SAAS;MAAE;MAAQ,gBAAgB;KAAO,CAAC;KAClF,SAAS;KACT,kBAAkB;IACpB,CAAC;IACD,IAAI,CAAC,KAAK,WAAW,MAAM,KAAK;IAChC,MAAM,OAAO,KAAK;IAMlB,kBAAkB,OAAO,KAAK,aAAa;KACzC,iBAAiB;KACjB,SAAS,uCAAuC,YAAY;IAC9D,CAAC;IACD,MAAM,MAAM,KAAK;IACjB,MAAM,OAAO;KACX,SAAS,WAAW;KACpB,UAAU,KAAK;KACf;KACA,UAAU;KACV,SAAS;KACT,cAAc;KACd,WAAW;KACX,WAAW,KAAK,IAAI;KACpB,YAAY,KAAK,IAAI,IAAI;KACzB,cAAc;KACd,gBAAgB,CAAC;IACnB,CAAC;IACD,OAAO,yBAAyB,GAAG;GACrC,SAAS,KAAK;IACZ,UAAU;IACV,MAAM,OAAO;KACX,SAAS,WAAW;KACpB,UAAU,KAAK;KACf;KACA,UAAU;KACV,SAAS;KACT,cAAc;KACd,WAAW;KACX,WAAW,KAAK,IAAI;KACpB,YAAY,KAAK,IAAI,IAAI;KACzB,cAAc,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;KAC7D,gBAAgB,CAAC;IACnB,CAAC;IAGD,IAAI,eAAe,wBAAwB,MAAM;GACnD;EACF;EACA,MAAM,mBAAmB,QAAQ,UAAU,IAAI,MAAM,OAAO,OAAO,CAAC;CACtE;CACA,QAAQ,cAAc;EACpB,UAAU;EACV,SAAS;GAAE,MAAM;GAA2B,SAAS;EAAM;EAC3D,aAAa,CACX,uGACF;CACF;CACA,OAAO;AACT;;;ACpmBA,SAAS,aAAa,UAAsC;CAC1D,IAAI,CAAC,UAAU,OAAO;CACtB,IAAI,SAAS,SAAS,MAAM,GAAG,OAAO;CACtC,IAAI,SAAS,WAAW,OAAO,GAAG,OAAO;CACzC,OAAO;AACT;;;;;;;;;;;;;;AAeA,SAAgB,qBAAqB,QAAoD;CACvF,MAAM,4BAAY,IAAI,IAAsB;CAC5C,MAAM,4BAAY,IAAI,IAA8B;CACpD,MAAM,YAAsB,CAAC;CAC7B,MAAM,4BAAY,IAAI,IAAY;CAElC,KAAK,MAAM,MAAM,QACf,IAAI,GAAG,SAAS,aAAa;EAC3B,MAAM,OAAQ,GAAyB;EACvC,IAAI,QAAQ,CAAC,UAAU,IAAI,IAAI,GAAG;GAChC,UAAU,IAAI,IAAI;GAClB,UAAU,KAAK,IAAI;EACrB;CACF,OAAO,IAAI,GAAG,SAAS,YAAY;EACjC,MAAM,IAAI;EACV,MAAM,OAAO,EAAE,QAAQ,EAAE,OAAO,EAAE;EAClC,UAAU,IAAI,MAAM;GAClB,MAAM,aAAa,EAAE,QAAQ;GAC7B;GACA,SAAS,EAAE,WAAW;EACxB,CAAC;CACH,OAAO,IAAI,GAAG,SAAS,oBAAoB;EACzC,MAAM,IAAI;EACV,UAAU,IAAI,EAAE,YAAY;GAC1B,IAAI,EAAE;GACN,OAAO,EAAE;GACT,QAAQ,EAAE,UAAU;GACpB,GAAI,EAAE,YAAY,KAAA,IAAY,EAAE,SAAS,EAAE,QAAQ,IAAI,CAAC;EAC1D,CAAC;CACH;CAGF,OAAO;EAAE,WAAW,CAAC,GAAG,UAAU,OAAO,CAAC;EAAG,WAAW,CAAC,GAAG,UAAU,OAAO,CAAC;EAAG;CAAU;AAC7F"}

package/dist/{abort-signal-CtzAM_sJ.js → proposal-findings-CXblsskH.js} RENAMED Viewed

@@ -55,15 +55,6 @@ function findingLabel(finding, index) {
 	return typeof id === "string" && id.length > 0 ? id : `index ${index}`;
 }
 //#endregion
-//#region src/abort-signal.ts
-/** Combine active cancellation sources without wrapping a single source. */
-function combineAbortSignals(...signals) {
-	const active = [...new Set(signals.filter((signal) => signal !== void 0))];
-	if (active.length === 0) return void 0;
-	if (active.length === 1) return active[0];
-	return AbortSignal.any(active);
-}
-//#endregion
-export { assertProposalFindings as n, isProposalFinding as r, combineAbortSignals as t };
+export { isProposalFinding as n, assertProposalFindings as t };
-//# sourceMappingURL=abort-signal-CtzAM_sJ.js.map
+//# sourceMappingURL=proposal-findings-CXblsskH.js.map

package/dist/proposal-findings-CXblsskH.js.map ADDED Viewed

@@ -0,0 +1 @@

+ {"version":3,"file":"proposal-findings-CXblsskH.js","names":[],"sources":["../src/analyst/proposal-findings.ts"],"sourcesContent":["import { z } from 'zod'\nimport type { ProposalFinding } from './types'\n\nconst ProposalFindingSchema = z\n .object({\n schema_version: z.literal('1.0.0'),\n finding_id: z.string().min(1),\n analyst_id: z.string().min(1),\n produced_at: z.string().min(1),\n severity: z.enum(['critical', 'high', 'medium', 'low', 'info']),\n area: z.string().min(1),\n claim: z.string().min(1),\n rationale: z.string().optional(),\n evidence_refs: z.array(\n z\n .object({\n kind: z.enum(['span', 'event', 'artifact', 'finding', 'metric']),\n uri: z.string().min(1),\n excerpt: z.string().optional(),\n })\n .strict(),\n ),\n recommended_action: z.string().optional(),\n validation_plan: z.string().optional(),\n confidence: z.number().min(0).max(1),\n subject: z.string().optional(),\n derived_from_judge: z.boolean().optional(),\n metadata: z.record(z.string(), z.unknown()).optional(),\n proposal_origin: z.enum(['search', 'production']),\n })\n .strict() satisfies z.ZodType<ProposalFinding>\n\n/** True when a finding names a source candidate generation may learn from. */\nexport function isProposalFinding(finding: unknown): finding is ProposalFinding {\n return ProposalFindingSchema.safeParse(finding).success\n}\n\n/**\n * Reject findings whose source has not been explicitly admitted for candidate\n * generation. Search feedback and observed production behavior are allowed;\n * final evaluation data has no allowed origin.\n */\nexport function assertProposalFindings(\n findings: unknown,\n context = 'proposal findings',\n): ReadonlyArray<ProposalFinding> {\n if (!Array.isArray(findings)) {\n throw new TypeError(`${context}: expected an array`)\n }\n const rejected = findings.flatMap((finding, index) =>\n isProposalFinding(finding) ? [] : [findingLabel(finding, index)],\n )\n if (rejected.length > 0) {\n throw new Error(\n `${context}: every finding must match AnalystFinding and declare ` +\n `proposal_origin as search or production. ` +\n `Rejected findings: [${rejected.join(', ')}].`,\n )\n }\n return findings as ReadonlyArray<ProposalFinding>\n}\n\nfunction findingLabel(finding: unknown, index: number): string {\n if (typeof finding !== 'object' || finding === null) return `index ${index}`\n const id = (finding as { finding_id?: unknown }).finding_id\n return typeof id === 'string' && id.length > 0 ? id : `index ${index}`\n}\n"],"mappings":";;AAGA,MAAM,wBAAwB,EAC3B,OAAO;CACN,gBAAgB,EAAE,QAAQ,OAAO;CACjC,YAAY,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC;CAC5B,YAAY,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC;CAC5B,aAAa,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC;CAC7B,UAAU,EAAE,KAAK;EAAC;EAAY;EAAQ;EAAU;EAAO;CAAM,CAAC;CAC9D,MAAM,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC;CACtB,OAAO,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC;CACvB,WAAW,EAAE,OAAO,CAAC,CAAC,SAAS;CAC/B,eAAe,EAAE,MACf,EACG,OAAO;EACN,MAAM,EAAE,KAAK;GAAC;GAAQ;GAAS;GAAY;GAAW;EAAQ,CAAC;EAC/D,KAAK,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC;EACrB,SAAS,EAAE,OAAO,CAAC,CAAC,SAAS;CAC/B,CAAC,CAAC,CACD,OAAO,CACZ;CACA,oBAAoB,EAAE,OAAO,CAAC,CAAC,SAAS;CACxC,iBAAiB,EAAE,OAAO,CAAC,CAAC,SAAS;CACrC,YAAY,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC;CACnC,SAAS,EAAE,OAAO,CAAC,CAAC,SAAS;CAC7B,oBAAoB,EAAE,QAAQ,CAAC,CAAC,SAAS;CACzC,UAAU,EAAE,OAAO,EAAE,OAAO,GAAG,EAAE,QAAQ,CAAC,CAAC,CAAC,SAAS;CACrD,iBAAiB,EAAE,KAAK,CAAC,UAAU,YAAY,CAAC;AAClD,CAAC,CAAC,CACD,OAAO;;AAGV,SAAgB,kBAAkB,SAA8C;CAC9E,OAAO,sBAAsB,UAAU,OAAO,CAAC,CAAC;AAClD;;;;;;AAOA,SAAgB,uBACd,UACA,UAAU,qBACsB;CAChC,IAAI,CAAC,MAAM,QAAQ,QAAQ,GACzB,MAAM,IAAI,UAAU,GAAG,QAAQ,oBAAoB;CAErD,MAAM,WAAW,SAAS,SAAS,SAAS,UAC1C,kBAAkB,OAAO,IAAI,CAAC,IAAI,CAAC,aAAa,SAAS,KAAK,CAAC,CACjE;CACA,IAAI,SAAS,SAAS,GACpB,MAAM,IAAI,MACR,GAAG,QAAQ,qHAEc,SAAS,KAAK,IAAI,EAAE,GAC/C;CAEF,OAAO;AACT;AAEA,SAAS,aAAa,SAAkB,OAAuB;CAC7D,IAAI,OAAO,YAAY,YAAY,YAAY,MAAM,OAAO,SAAS;CACrE,MAAM,KAAM,QAAqC;CACjD,OAAO,OAAO,OAAO,YAAY,GAAG,SAAS,IAAI,KAAK,SAAS;AACjE"}

package/dist/{registry-Bu3lNole.d.ts → registry-CT-hMSkR.d.ts} RENAMED Viewed

@@ -1,7 +1,7 @@
 import { c as CostLedgerHandle } from "./cost-ledger-DbQdN3nO.js";
-import { c as AnalystRunInputs, i as AnalystFinding, l as AnalystRunResult, n as AnalystContext, s as AnalystRunEvent, t as Analyst, u as AnalystRunSummary } from "./types-CeDfCRXp.js";
-import { x as ChatClient } from "./types-C0bWoSKC.js";
-import { i as ExactAnalystRunEvent, o as ExactAnalystRunResult, u as ExactExecutionComponentIdentity } from "./exact-types-KGxBWH4b.js";
+import { c as AnalystRunInputs, i as AnalystFinding, l as AnalystRunResult, n as AnalystContext, s as AnalystRunEvent, t as Analyst, u as AnalystRunSummary } from "./types-6VEehmEg.js";
+import { x as ChatClient } from "./types-DLxe96rp.js";
+import { i as ExactAnalystRunEvent, o as ExactAnalystRunResult, u as ExactExecutionComponentIdentity } from "./exact-types-BBMMYk3g.js";
 //#region src/analyst/registry.d.ts
 interface AnalystHooks {
   /** Legacy runs may mutate ctx; exact runs provide a frozen observational context. */
@@ -173,4 +173,4 @@ declare class AnalystRegistry {
 }
 //#endregion
 export { ExactAnalystBudgetPolicy as a, RegistryRunOpts as c, BudgetPolicy as i, AnalystRegistry as n, ExactAnalystRunExecutionError as o, AnalystRegistryOptions as r, ExactRegistryRunOpts as s, AnalystHooks as t };
-//# sourceMappingURL=registry-Bu3lNole.d.ts.map
+//# sourceMappingURL=registry-CT-hMSkR.d.ts.map

package/dist/{registry-Bu3lNole.d.ts.map → registry-CT-hMSkR.d.ts.map} RENAMED Viewed

	@@ -1 +1 @@
1	- {"version":3,"file":"registry-~~Bu3lNole~~.d.ts","names":[],"sources":["../src/analyst/registry.ts"],"mappings":";;;;;UAqDiB;;EAEf,iBAAiB;IACf,SAAS;IACT,KAAK;IACL;aACS;;EAEX,gBAAgB;IACd,SAAS;IACT,SAAS;IACT,UAAU;IACV;aACS;;;;;;EAMX,SAAS;IACP,SAAS;IACT,OAAO;IACP;MACE,+BAA+B,QAAQ;;EAE3C,YAAY;IAAQ,QAAQ;aAA4B;;UAGzC;;EAEf;;EAEA,UAAU;;;;;;;EAOV,YAAY;IACV,SAAS;IACT;IACA;IACA;;;UAIa;;EAEf,OAAO;;EAEP,OAAO,aAAa,SAAS;;EAE7B,QAAQ;;EAER,gBAAgB;;EAEhB,gBAAgB;;EAEhB,eAAe;;UAGA;;EAEf;;EAEA;;EAEA,SAAS;;EAET;;EAEA,SAAS;;EAET,aAAa;;EAEb;;EAEA,OAAO;;;;;;;;;EASP,gBAAgB,cAAc,kBAAkB,eAAe,cAAc;;;;;;EAM7E;;;KAIU;WAEG;WACA;;WAGA;WACA;;WAEA,SAAS,SAAS;;;;;;;;;;UAWhB;WACN;WACA,QAAQ;WACR;WACA,QAAQ;WACR,YAAY;WACZ,oBAAoB;WACpB;WACA,MAAM,SAAS;WACf,eACL,cAAc,kBACd,SAAS,eAAe,cAAc;WAEjC;WACA;WACA;WACA;;;cAqCE,sCAAsC;WACxC;WACA,QAAQ;EAEjB,YAAY,iBAAiB,QAAQ,uBAAuB,UAAU;;cAU3D;mBACM;mBACA;EAEjB,YAAY,UAAS;EAIrB,SAAS,SAAS;EAsBlB,QAAQ;IACN;IACA;IACA;IACA,MAAM;;EAUF,IACJ,eACA,QAAQ,kBACR,UAAS,kBACR,QAAQ;;EAUL,SACJ,eACA,QAAQ,kBACR,SAAS,uBACR,QAAQ;;EAQJ,eACL,eACA,QAAQ,kBACR,SAAS,uBACR,eAAe;;;;;;;;;;;;EAmBX,UACL,eACA,QAAQ,kBACR,UAAS,kBACR,eAAe;UAIV;UA8BA;UAwFO;UA2aP;UAaA;UAUA"}
1	+ {"version":3,"file":"registry-CT-hMSkR.d.ts","names":[],"sources":["../src/analyst/registry.ts"],"mappings":";;;;;UAqDiB;;EAEf,iBAAiB;IACf,SAAS;IACT,KAAK;IACL;aACS;;EAEX,gBAAgB;IACd,SAAS;IACT,SAAS;IACT,UAAU;IACV;aACS;;;;;;EAMX,SAAS;IACP,SAAS;IACT,OAAO;IACP;MACE,+BAA+B,QAAQ;;EAE3C,YAAY;IAAQ,QAAQ;aAA4B;;UAGzC;;EAEf;;EAEA,UAAU;;;;;;;EAOV,YAAY;IACV,SAAS;IACT;IACA;IACA;;;UAIa;;EAEf,OAAO;;EAEP,OAAO,aAAa,SAAS;;EAE7B,QAAQ;;EAER,gBAAgB;;EAEhB,gBAAgB;;EAEhB,eAAe;;UAGA;;EAEf;;EAEA;;EAEA,SAAS;;EAET;;EAEA,SAAS;;EAET,aAAa;;EAEb;;EAEA,OAAO;;;;;;;;;EASP,gBAAgB,cAAc,kBAAkB,eAAe,cAAc;;;;;;EAM7E;;;KAIU;WAEG;WACA;;WAGA;WACA;;WAEA,SAAS,SAAS;;;;;;;;;;UAWhB;WACN;WACA,QAAQ;WACR;WACA,QAAQ;WACR,YAAY;WACZ,oBAAoB;WACpB;WACA,MAAM,SAAS;WACf,eACL,cAAc,kBACd,SAAS,eAAe,cAAc;WAEjC;WACA;WACA;WACA;;;cAqCE,sCAAsC;WACxC;WACA,QAAQ;EAEjB,YAAY,iBAAiB,QAAQ,uBAAuB,UAAU;;cAU3D;mBACM;mBACA;EAEjB,YAAY,UAAS;EAIrB,SAAS,SAAS;EAsBlB,QAAQ;IACN;IACA;IACA;IACA,MAAM;;EAUF,IACJ,eACA,QAAQ,kBACR,UAAS,kBACR,QAAQ;;EAUL,SACJ,eACA,QAAQ,kBACR,SAAS,uBACR,QAAQ;;EAQJ,eACL,eACA,QAAQ,kBACR,SAAS,uBACR,eAAe;;;;;;;;;;;;EAmBX,UACL,eACA,QAAQ,kBACR,UAAS,kBACR,eAAe;UAIV;UA8BA;UAwFO;UA2aP;UAaA;UAUA"}

package/dist/{report-command-DzPpFgjB.js → report-command-D0XL2Cvi.js} RENAMED Viewed

@@ -1,3 +1,4 @@
+import { a as hashCanonical } from "./canonical-BsZVgBhW.js";
 import { o as summarizeNumberSeries } from "./descriptive-D9LUFk1-.js";
 import { a as parseJson, c as workerSourceKey, d as SUPERVISOR_RUN_SCHEMA, f as isUnavailable, i as asRecord, l as NO_SOURCE_LIMITS, m as unavailable, o as parseJsonl, p as showMeasured, r as readTerminalRecord, s as parseSupervisorTree, u as SUPERVISOR_RUN_ROLLUP_SCHEMA } from "./terminal-record-BHocwFw1.js";
 import { basename, join, resolve } from "node:path";
@@ -322,7 +323,18 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
 			unknownNodes: closeUsdUnreported && rootId !== null ? [rootId] : []
 		}
 	};
+	const noteCaptureGaps = (owner, captures) => {
+		if (captures === null || captures?.length === 0) gaps.push(`${owner} workspaceCapture: no retained output capture evidence`);
+		else if (captures !== void 0) for (const [index, capture] of captures.entries()) {
+			const source = `${owner} workspaceCapture[${index + 1}]`;
+			if (capture.projectionGap !== null) gaps.push(`${source}: ${capture.projectionGap}`);
+			if (capture.coverageComplete === false) gaps.push(`${source}: ${capture.incompleteReason ?? "provider reported incomplete coverage"}`);
+			if (capture.coverageComplete === null && capture.projectionGap === null) gaps.push(`${source}: capture coverage was not recorded`);
+		}
+	};
+	noteCaptureGaps(`root ${rootId ?? src.runRef}`, src.rootWorkspaceCaptures);
 	const perWorker = (src.workers ?? []).map((w) => {
+		noteCaptureGaps(`worker ${w.workerId ?? w.label}`, w.workspaceCaptures);
 		const f = tree.workerLogs.get(workerSourceKey(w));
 		const matchingSpawns = spawnsForSource(w);
 		const spawn = spawnForSource(w);
@@ -347,7 +359,8 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
 			usd: usdLimit !== null ? null : w.workerId === void 0 ? workerUsdByLabel.get(w.label) ?? null : workerUsdById.get(w.workerId) ?? null,
 			patchBytes: w.patchBytes ?? f?.finishedPatchBytes ?? null,
 			passed,
-			score: close?.score ?? f?.score ?? null
+			score: close?.score ?? f?.score ?? null,
+			workspaceCaptures: w.workspaceCaptures ?? null
 		};
 	});
 	const wallDistribution = summarizeNumberSeries(perWorker.map((w) => w.wallMs).filter((w) => w !== null));
@@ -408,6 +421,7 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
 		decision,
 		economics,
 		outcome,
+		rootWorkspaceCaptures: src.rootWorkspaceCaptures ?? null,
 		gaps,
 		traceCommand: src.traceCommand ?? "npx --yes @tangle-network/traces@latest analyze --harness opencode --cwd <worker-clone-cwd>"
 	};
@@ -750,7 +764,7 @@ function renderSupervisorRollupMarkdown(rollup, title = "Round rollup") {
 *
 * Transcripts come from Runtime's own records too: `root-stream.jsonl` is the root's, each
 * worker's turn outputs and native-session receipt are read from the journal, and every blob
-* a record names is checked on disk (see `workerTranscripts`).
+* a record names is checked on disk (see `nodeTranscripts`).
 *
 * The run's terminal record is Runtime's own: `result.json` is the
 * `SupervisedResult` that `supervise()` returned, verbatim, and its `kind`
@@ -1043,13 +1057,87 @@ function runtimeBeginState(root, startedAt) {
 function blobFile(runDir, ref) {
 	return typeof ref === "string" && /^sha256:[0-9a-f]{64}$/u.test(ref) ? join(runDir, BLOB_DIR, `${ref.replace(":", "-")}.json`) : null;
 }
+async function workspaceCapturesFromOutput(runDir, ref) {
+	const outRef = typeof ref === "string" ? ref : null;
+	const blobPath = blobFile(runDir, ref);
+	const missing = (projectionGap) => ({
+		outRef,
+		blobPath,
+		receiptPointer: null,
+		receipt: null,
+		attempts: null,
+		coverageComplete: null,
+		incompleteReason: null,
+		projectionGap
+	});
+	if (blobPath === null) return [missing("Runtime output has no valid content address")];
+	if (!await isFile(blobPath)) return [missing("Runtime output blob is missing or not a file")];
+	const raw = await readMaybe(blobPath);
+	if (raw === null) return [missing("Runtime output blob is missing")];
+	let output;
+	try {
+		output = JSON.parse(raw);
+	} catch {
+		return [missing("Runtime output blob is not JSON")];
+	}
+	let addressGap = null;
+	try {
+		if (hashCanonical(output) !== outRef) addressGap = "Runtime output content address does not match output blob";
+	} catch {
+		addressGap = "Runtime output cannot be canonically hashed";
+	}
+	const project = (rawReceipt, receiptPointer) => {
+		const receipt = record(rawReceipt);
+		if (receipt === null) return {
+			...missing(`workspaceCapture receipt is absent or malformed${addressGap === null ? "" : `; ${addressGap}`}`),
+			receiptPointer,
+			receipt: rawReceipt
+		};
+		const provenance = record(receipt.provenance);
+		const attempts = Array.isArray(provenance?.attempts) ? provenance.attempts : null;
+		const coverageComplete = typeof receipt.coverageComplete === "boolean" ? receipt.coverageComplete : null;
+		const incompleteReason = typeof receipt.incompleteReason === "string" ? receipt.incompleteReason : null;
+		const gaps = [];
+		if (addressGap !== null) gaps.push(addressGap);
+		if (coverageComplete === null) gaps.push("workspaceCapture.coverageComplete is absent");
+		if (record(receipt.snapshot) === null) gaps.push("workspaceCapture.snapshot is absent");
+		if (provenance === null) gaps.push("workspaceCapture.provenance is absent");
+		if (attempts === null) gaps.push("workspaceCapture.provenance.attempts is absent");
+		else if (attempts.length === 0 || attempts.some((attempt) => record(attempt) === null)) gaps.push("workspaceCapture.provenance.attempts is empty or malformed");
+		return {
+			outRef,
+			blobPath,
+			receiptPointer,
+			receipt,
+			attempts,
+			coverageComplete,
+			incompleteReason,
+			projectionGap: gaps.length > 0 ? gaps.join("; ") : null
+		};
+	};
+	const result = record(output);
+	const captures = [];
+	for (const key of Object.keys(result ?? {})) if (key === "workspaceCapture") captures.push(project(result?.[key], "/workspaceCapture"));
+	else if (key === "workspaceCaptures") {
+		const receipts = result?.[key];
+		if (Array.isArray(receipts) && receipts.length > 0) for (const [index, receipt] of receipts.entries()) captures.push(project(receipt, `/workspaceCaptures/${index}`));
+		else captures.push({
+			...missing(`workspaceCaptures is empty or malformed${addressGap === null ? "" : `; ${addressGap}`}`),
+			receiptPointer: "/workspaceCaptures",
+			receipt: receipts ?? null
+		});
+	}
+	return captures.length > 0 ? captures : [project(null, null)];
+}
 /**
-* What Runtime retained of each worker's decisions, from its own records.
+* What Runtime retained of each node's decisions, from its own records.
 *
 * - A turn is one `execution-admitted` event in the `dispatched` phase. Its `execution-result`
 *   names an output blob holding that turn's provider event stream (reasoning, tool calls and
 *   results as the harness reported them). A dispatched turn with no retained output lost its
 *   record, which is how a worker that went down mid-turn shows up.
+* - Every event's `outRef` enters the same collector, once per node and address. Terminal
+*   records retain ordinary and steerable results; they are not counted as dispatched turns.
 * - The terminal event's `harnessTranscript` receipt covers the harness's native session files.
 *   Those are the only record of the harness's own subagents, so an unavailable receipt is a
 *   separate gap from a missing turn, and its reason is Runtime's, verbatim.
@@ -1057,14 +1145,15 @@ function blobFile(runDir, ref) {
 * Blob files are checked on disk: a receipt that names a blob the directory no longer holds is
 * not a retained transcript.
 */
-async function workerTranscripts(runDir, events, workerIds) {
+async function nodeTranscripts(runDir, events, nodeIds) {
 	const facts = /* @__PURE__ */ new Map();
 	const entry = (id) => {
 		let found = facts.get(id);
 		if (found === void 0) {
 			found = {
 				dispatched: 0,
-				outputFiles: [],
+				outputRefs: [],
+				outputPositions: /* @__PURE__ */ new Map(),
 				receipt: null
 			};
 			facts.set(id, found);
@@ -1073,12 +1162,23 @@ async function workerTranscripts(runDir, events, workerIds) {
 	};
 	for (const event of events) {
 		const id = nonEmptyString(event.id);
-		if (id === null || !workerIds.has(id)) continue;
+		if (id === null || !nodeIds.has(id)) continue;
+		if (Object.hasOwn(event, "outRef")) {
+			const fact = entry(id);
+			const ref = event.outRef;
+			const position = typeof ref === "string" ? fact.outputPositions.get(ref) : void 0;
+			if (position !== void 0) {
+				if (event.kind === "execution-result") fact.outputRefs[position].turn = true;
+			} else {
+				if (typeof ref === "string") fact.outputPositions.set(ref, fact.outputRefs.length);
+				fact.outputRefs.push({
+					value: ref,
+					turn: event.kind === "execution-result"
+				});
+			}
+		}
 		if (event.kind === "execution-admitted" && record(event.admission)?.phase === "dispatched") entry(id).dispatched += 1;
-		else if (event.kind === "execution-result") {
-			const file = blobFile(runDir, event.outRef);
-			if (file !== null) entry(id).outputFiles.push(file);
-		} else if (event.kind === "settled" || event.kind === "cancelled") {
+		else if (event.kind === "settled" || event.kind === "cancelled") {
 			const receipt = record(event.harnessTranscript);
 			if (receipt !== null) entry(id).receipt = receipt;
 		}
@@ -1086,7 +1186,16 @@ async function workerTranscripts(runDir, events, workerIds) {
 	const out = /* @__PURE__ */ new Map();
 	for (const [id, fact] of facts) {
 		const retained = [];
-		for (const file of fact.outputFiles) if (await isFile(file)) retained.push(file);
+		let retainedTurns = 0;
+		const workspaceCaptures = [];
+		for (const { value: ref, turn } of fact.outputRefs) {
+			const file = blobFile(runDir, ref);
+			if (file !== null && await isFile(file)) {
+				retained.push(file);
+				if (turn) retainedTurns += 1;
+			}
+			workspaceCaptures.push(...await workspaceCapturesFromOutput(runDir, ref));
+		}
 		let nativeSession = null;
 		if (fact.receipt !== null) {
 			if (fact.receipt.status === "available") {
@@ -1107,9 +1216,10 @@ async function workerTranscripts(runDir, events, workerIds) {
 			transcriptRef: nativeSession?.status === "available" ? nativeSession.ref : retained.at(-1) ?? null,
 			turns: {
 				dispatched: fact.dispatched,
-				retained: retained.length
+				retained: retainedTurns
 			},
-			nativeSession
+			nativeSession,
+			workspaceCaptures: workspaceCaptures.length > 0 ? workspaceCaptures : null
 		});
 	}
 	return out;
@@ -1153,6 +1263,7 @@ function absentRuntimeSupervisorRun(runDir, resultText, failureText, instanceId)
 		harnessMissingReason: "Runtime FileRunContext has no external worker-token join",
 		limits: NO_SOURCE_LIMITS,
 		rootTranscriptRef: null,
+		rootWorkspaceCaptures: null,
 		traceCommand: "unavailable — Runtime FileRunContext records no provider-session trace identity"
 	};
 }
@@ -1188,7 +1299,7 @@ async function readRuntimeSupervisorRun(runDir, opts = {}) {
 	assertResultMatchesJournal(normalized.root, result, resultPath);
 	const childSpawns = normalized.events.filter((event) => event.kind === "spawned" && nonEmptyString(event.id) !== null).filter((event) => event.id !== normalized.root);
 	const workerIds = new Set(childSpawns.map((event) => nonEmptyString(event.id)).filter((id) => id !== null));
-	const transcripts = await workerTranscripts(runDir, normalized.events, workerIds);
+	const transcripts = await nodeTranscripts(runDir, normalized.events, /* @__PURE__ */ new Set([normalized.root, ...workerIds]));
 	const workers = childSpawns.map((event) => {
 		const workerId = nonEmptyString(event.id);
 		const transcript = transcripts.get(workerId);
@@ -1204,6 +1315,7 @@ async function readRuntimeSupervisorRun(runDir, opts = {}) {
 				retained: 0
 			},
 			nativeSession: transcript?.nativeSession ?? null,
+			workspaceCaptures: transcript?.workspaceCaptures ?? null,
 			patchPath: null
 		};
 	});
@@ -1230,6 +1342,7 @@ async function readRuntimeSupervisorRun(runDir, opts = {}) {
 		harnessMissingReason: "Runtime FileRunContext has no external worker-token join",
 		limits: sourceLimits(normalized.root, normalized.events, workerIds),
 		rootTranscriptRef: await isFile(rootStream) ? rootStream : null,
+		rootWorkspaceCaptures: transcripts.get(normalized.root)?.workspaceCaptures ?? null,
 		traceCommand: "unavailable — Runtime FileRunContext records no provider-session trace identity"
 	};
 }
@@ -1484,4 +1597,4 @@ async function runSupervisorRunCommand(argv, io = PROCESS_IO) {
 //#endregion
 export { writeSupervisorRunReport as a, readRuntimeSupervisorRun as c, renderSupervisorRunHeadline as d, renderSupervisorRunMarkdown as f, rollupSupervisorRuns as h, reportSupervisorRound as i, runtimeSupervisorRunReader as l, parsePatch as m, analyzeSupervisorRun as n, writeSupervisorRunReportSafe as o, analyzeSupervisorRunSources as p, findSupervisorRunDirs as r, isRuntimeSupervisorRunDir as s, runSupervisorRunCommand as t, renderSupervisorRollupMarkdown as u };
-//# sourceMappingURL=report-command-DzPpFgjB.js.map
+//# sourceMappingURL=report-command-D0XL2Cvi.js.map