npm - @tangle-network/agent-eval - Versions diffs - 0.208.2 → 0.210.0 - Mend

@tangle-network/agent-eval 0.208.2 → 0.210.0

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 (237) hide show

package/dist/paired-binary-b2L3WKqr.js.map DELETED Viewed

@@ -1 +0,0 @@

- {"version":3,"file":"paired-binary-b2L3WKqr.js","names":[],"sources":["../src/statistics/paired-binary.ts"],"sourcesContent":["import { regularizedIncompleteBeta } from '../math/special-functions'\nimport { binomialSignTwoSided, zQuantile } from './internal'\n\n// ── Binomial proportion + paired-binary + coding-eval estimators ─────\n//\n// The paired family above (pairedBootstrap/pairedTTest/wilcoxonSignedRank)\n// operates on continuous scores. Pass/fail A/B comparisons — \"does treatment\n// X raise the success RATE vs control\" — are binary and paired, so they need\n// their own correct estimators: McNemar for significance (only the discordant\n// pairs carry signal), the paired risk difference for effect size, Wilson for\n// a single-arm proportion CI, and pass@k for the standard k-sample coding-eval\n// metric. The normal approximation is wrong for proportions near 0/1 and for\n// the small discordant counts typical of eval runs, so these are exact /\n// Wilson-based, not Wald.\n\n/** A binomial proportion estimate with a confidence interval. */\nexport interface ProportionInterval {\n /** Point estimate successes / n (0 when n = 0). */\n estimate: number\n /** Lower bound, clamped to [0, 1]. */\n lower: number\n /** Upper bound, clamped to [0, 1]. */\n upper: number\n}\n\n/**\n * Wilson score interval for a binomial proportion. Correct at small n and near\n * 0/1, where the normal (Wald) approximation produces bounds outside [0, 1] and\n * understates coverage. Use this for any pass-rate / hit-rate / realness-rate\n * CI — the continuous `confidenceInterval` assumes the wrong distribution for a\n * proportion. Counts must be finite nonnegative integers, successes cannot exceed\n * n, and confidence must lie strictly between zero and one. `n = 0 ⇒ {0, 0, 0}`.\n */\nexport function wilson(successes: number, n: number, confidence = 0.95): ProportionInterval {\n if (!Number.isSafeInteger(n) || n < 0) {\n throw new Error('wilson: n must be a finite nonnegative integer')\n }\n if (!Number.isSafeInteger(successes) || successes < 0 || successes > n) {\n throw new Error(`wilson: successes (${successes}) must be in [0, ${n}]`)\n }\n if (!Number.isFinite(confidence) || confidence <= 0 || confidence >= 1) {\n throw new Error('wilson: confidence must be between zero and one')\n }\n if (n === 0) return { estimate: 0, lower: 0, upper: 0 }\n const z = zQuantile(1 - (1 - confidence) / 2)\n const p = successes / n\n const z2 = z * z\n const denom = 1 + z2 / n\n const center = (p + z2 / (2 * n)) / denom\n const half = (z * Math.sqrt((p * (1 - p) + z2 / (4 * n)) / n)) / denom\n return {\n estimate: p,\n lower: Math.max(0, center - half),\n upper: Math.min(1, center + half),\n }\n}\n\n/**\n * Are these per-item outcomes binary (every value exactly 0 or 1)?\n *\n * The discriminator a promotion gate needs before choosing a paired statistic.\n * On binary outcomes the paired delta vector lives in {-1, 0, +1} and is\n * normally dominated by zeros (both arms solve, or both arms miss, most items),\n * so its MEDIAN is pinned at exactly 0 no matter how large the real shift in\n * success rate is — and a bootstrap CI on that median collapses to [0, 0].\n * A gate keying on `ci.low > threshold` is then structurally unable to see\n * either a gain or a regression. Detect this shape and switch to the\n * paired-binary estimators ({@link mcnemar}, {@link pairedRiskDifference})\n * instead of silently answering \"no\" forever.\n *\n * Empty input is NOT binary: there is no evidence of the outcome's shape, and\n * defaulting an empty vector into the binary branch would pick a statistic on\n * no data at all.\n *\n * NOT the right discriminator for a gate. It recognises the literal {0, 1}\n * encoding and nothing else, so a pass/fail dimension emitted on 0-100 — which\n * judges in this codebase do routinely — reads as non-binary, and a single\n * partial-credit score in an otherwise pass/fail vector flips it to false while\n * leaving the median just as blind. Gates want {@link pairedBinaryScale} (any\n * two-point encoding). This predicate remains for callers that specifically\n * mean \"literally 0/1\".\n */\nexport function isBinaryOutcomeVector(values: ArrayLike<number>): boolean {\n if (values.length === 0) return false\n for (let i = 0; i < values.length; i++) {\n const v = values[i]!\n if (v !== 0 && v !== 1) return false\n }\n return true\n}\n\n/** Result of a McNemar paired-binary significance test. */\nexport interface McNemarResult {\n /** Total paired observations. */\n n: number\n /** Discordant pairs (b + c) — the only ones that carry signal. */\n nDiscordant: number\n /** Pairs where treatment succeeded and control failed (\"newly correct\"). */\n b: number\n /** Pairs where control succeeded and treatment failed (\"newly wrong\"). */\n c: number\n /** Continuity-corrected chi-square statistic (reference; exact p drives the call). */\n statistic: number\n /** Two-sided p-value. Exact (binomial sign test on discordant pairs). */\n pValue: number\n}\n\n/**\n * McNemar's test for paired binary outcomes — the correct significance test for\n * \"does treatment change the success rate vs control on the SAME items\". Only\n * discordant pairs (one arm right, the other wrong) carry information; concordant\n * pairs are uninformative, so a paired t-test / two-proportion z-test on the raw\n * rates is wrong here. The p-value is exact: under H0 the b \"treatment-wins\" are\n * Binomial(b + c, 0.5), so the two-sided p is the doubled binomial tail — correct\n * at the small discordant counts typical of eval runs (no continuity-corrected\n * chi-square approximation needed, though it is returned as `statistic` for\n * reference). Inputs are paired 0/1 (or boolean) arrays, control first to match\n * the module's (before, after) convention. Throws on unequal lengths.\n */\nexport function mcnemar(\n control: ArrayLike<number | boolean>,\n treatment: ArrayLike<number | boolean>,\n): McNemarResult {\n if (control.length !== treatment.length) {\n throw new Error(`mcnemar: unequal sample sizes (${control.length} vs ${treatment.length})`)\n }\n const n = control.length\n let b = 0 // treatment 1, control 0\n let c = 0 // treatment 0, control 1\n for (let i = 0; i < n; i++) {\n const ctrl = control[i] ? 1 : 0\n const treat = treatment[i] ? 1 : 0\n if (treat === 1 && ctrl === 0) b++\n else if (treat === 0 && ctrl === 1) c++\n }\n const nDiscordant = b + c\n const statistic = nDiscordant === 0 ? 0 : (Math.abs(b - c) - 1) ** 2 / nDiscordant\n return { n, nDiscordant, b, c, statistic, pValue: binomialSignTwoSided(b, c) }\n}\n\n/** A paired binary effect size (treatment rate − control rate) with a CI. */\nexport interface RiskDifferenceResult {\n /** Total paired observations. */\n n: number\n /** Discordant pairs: treatment-win count. */\n b: number\n /** Discordant pairs: control-win count. */\n c: number\n /** Paired risk difference p(treatment) − p(control) = (b − c) / n. */\n riskDifference: number\n /** Lower bound of the CI, clamped to [-1, 1]. */\n lower: number\n /** Upper bound of the CI, clamped to [-1, 1]. */\n upper: number\n /** Confidence level used. */\n confidence: number\n}\n\n/**\n * Paired risk difference (the effect-size companion to {@link mcnemar}): the\n * change in success rate p(treatment) − p(control) on matched items, which for\n * paired binary data equals (b − c) / n. The CI uses the paired variance from\n * the discordant counts, not the independent-samples formula (which overstates\n * the interval by ignoring the pairing). Inputs are paired 0/1 (or boolean)\n * arrays, control first. Throws on unequal lengths.\n *\n * REPORTING ONLY — do NOT decide a promotion on this interval. The CI is a Wald\n * normal approximation, which badly UNDERCOVERS when only a handful of pairs are\n * discordant: at n = 3 with b = 2, c = 0 it returns [0.133, 1.000], excluding 0,\n * while McNemar's exact test on the same data gives p = 0.50. A gate keying on\n * `lower > 0` would promote noise. Use {@link pairedRiskDifferenceExact}, whose\n * interval is dual to the exact test by construction, for any decision.\n */\nexport function pairedRiskDifference(\n control: ArrayLike<number | boolean>,\n treatment: ArrayLike<number | boolean>,\n confidence = 0.95,\n): RiskDifferenceResult {\n if (control.length !== treatment.length) {\n throw new Error(\n `pairedRiskDifference: unequal sample sizes (${control.length} vs ${treatment.length})`,\n )\n }\n const n = control.length\n if (n === 0) return { n: 0, b: 0, c: 0, riskDifference: 0, lower: 0, upper: 0, confidence }\n let b = 0\n let c = 0\n for (let i = 0; i < n; i++) {\n const ctrl = control[i] ? 1 : 0\n const treat = treatment[i] ? 1 : 0\n if (treat === 1 && ctrl === 0) b++\n else if (treat === 0 && ctrl === 1) c++\n }\n const rd = (b - c) / n\n const variance = (b + c - (b - c) ** 2 / n) / (n * n)\n const z = zQuantile(1 - (1 - confidence) / 2)\n const half = z * Math.sqrt(Math.max(0, variance))\n return {\n n,\n b,\n c,\n riskDifference: rd,\n lower: Math.max(-1, rd - half),\n upper: Math.min(1, rd + half),\n confidence,\n }\n}\n\n/** A paired binary effect size with an EXACT interval and the exact test that\n * bounds it — one object so a caller cannot read the estimate without the\n * significance it is entitled to. */\nexport interface ExactRiskDifferenceResult {\n /** Total paired observations. */\n n: number\n /** Discordant pairs: treatment-win count. */\n b: number\n /** Discordant pairs: control-win count. */\n c: number\n /** Discordant pairs (b + c) — the only ones carrying information. */\n nDiscordant: number\n /** Paired risk difference p(treatment) − p(control) = (b − c) / n. */\n riskDifference: number\n /** Exact conditional CI lower bound. 0 when there are no discordant pairs. */\n lower: number\n /** Exact conditional CI upper bound. 0 when there are no discordant pairs. */\n upper: number\n /** Confidence level used. */\n confidence: number\n /** McNemar's exact two-sided p-value on the same discordant counts. */\n pValue: number\n}\n\n/**\n * Paired risk difference with the EXACT CONDITIONAL interval — the estimator a\n * promotion gate may decide on.\n *\n * Conditional on the number of discordant pairs m = b + c, the treatment-win\n * count b is Binomial(m, π) with π = P(treatment wins | discordant), and the\n * risk difference is an exact reparameterisation: RD = (2π − 1)·m/n. So a\n * Clopper-Pearson exact interval for π maps straight onto RD. This buys the\n * property the Wald interval in {@link pairedRiskDifference} does not have:\n *\n * **`lower > 0` ⟺ McNemar's exact test rejects at α = 1 − confidence.**\n *\n * Clopper-Pearson excludes π = 0.5 exactly when the two-sided exact binomial\n * test of π = 0.5 rejects, and that test IS {@link mcnemar}'s p-value — so the\n * interval and the test can never disagree, and a gate keyed on `lower` cannot\n * promote what the exact test refuses. The exact p is returned in the same\n * object so the two are impossible to compute apart.\n *\n * The interval is conservative (exact intervals over-cover; conditioning on m\n * discards the concordant pairs' information about m itself). That is the\n * correct direction for a promotion gate: it refuses more often, never less.\n *\n * With m = 0 there are no discordant pairs and π is not identified: the result\n * is the degenerate [0, 0] with p = 1. That is NOT evidence of equivalence —\n * callers must treat a zero-width interval as \"cannot decide\", not as \"no\n * difference\". Inputs are paired 0/1 (or boolean) arrays, control first.\n * Throws on unequal lengths.\n */\nexport function pairedRiskDifferenceExact(\n control: ArrayLike<number | boolean>,\n treatment: ArrayLike<number | boolean>,\n confidence = 0.95,\n): ExactRiskDifferenceResult {\n if (control.length !== treatment.length) {\n throw new Error(\n `pairedRiskDifferenceExact: unequal sample sizes (${control.length} vs ${treatment.length})`,\n )\n }\n if (confidence <= 0 || confidence >= 1) {\n throw new Error(`pairedRiskDifferenceExact: confidence must be in (0,1), got ${confidence}`)\n }\n const n = control.length\n if (n === 0) {\n return {\n n: 0,\n b: 0,\n c: 0,\n nDiscordant: 0,\n riskDifference: 0,\n lower: 0,\n upper: 0,\n confidence,\n pValue: 1,\n }\n }\n let b = 0\n let c = 0\n for (let i = 0; i < n; i++) {\n const ctrl = control[i] ? 1 : 0\n const treat = treatment[i] ? 1 : 0\n if (treat === 1 && ctrl === 0) b++\n else if (treat === 0 && ctrl === 1) c++\n }\n const m = b + c\n const riskDifference = (b - c) / n\n const pValue = binomialSignTwoSided(b, c)\n if (m === 0) {\n return { n, b, c, nDiscordant: 0, riskDifference: 0, lower: 0, upper: 0, confidence, pValue }\n }\n const alpha = 1 - confidence\n const piLow = b === 0 ? 0 : betaQuantile(alpha / 2, b, m - b + 1)\n const piHigh = b === m ? 1 : betaQuantile(1 - alpha / 2, b + 1, m - b)\n const scale = m / n\n return {\n n,\n b,\n c,\n nDiscordant: m,\n riskDifference,\n lower: Math.max(-1, (2 * piLow - 1) * scale),\n upper: Math.min(1, (2 * piHigh - 1) * scale),\n confidence,\n pValue,\n }\n}\n\n/** Inverse regularized incomplete beta by bisection on\n * {@link regularizedIncompleteBeta}, which is monotone increasing in x. 80\n * halvings of [0,1] resolve to ~8e-25, far past the continued fraction's own\n * 3e-15 tolerance, so the quantile is as exact as the CDF it inverts. */\nfunction betaQuantile(p: number, a: number, b: number): number {\n if (p <= 0) return 0\n if (p >= 1) return 1\n let lo = 0\n let hi = 1\n for (let i = 0; i < 80; i++) {\n const mid = (lo + hi) / 2\n if (regularizedIncompleteBeta(mid, a, b) < p) lo = mid\n else hi = mid\n }\n return (lo + hi) / 2\n}\n\n/** A paired binary effect size with an interval that is valid at a NONZERO\n * margin — the estimator a noninferiority decision may be made on. */\nexport interface ScoreRiskDifferenceResult {\n /** Total paired observations. */\n n: number\n /** Discordant pairs: treatment-win count. */\n b: number\n /** Discordant pairs: control-win count. */\n c: number\n /** Discordant pairs (b + c). */\n nDiscordant: number\n /** Paired risk difference p(treatment) − p(control) = (b − c) / n. */\n riskDifference: number\n /** Score-interval lower bound on the population risk difference. */\n lower: number\n /** Score-interval upper bound on the population risk difference. */\n upper: number\n /** Confidence level used. */\n confidence: number\n}\n\n/**\n * Constrained MLE of q = P(treatment loses) under the hypothesis RD = `delta`.\n *\n * Profiling the two concordant cells out of the multinomial leaves\n * `L(q) = b·log(q+delta) + c·log(q) + e·log(1 − 2q − delta)` with `e = n − b − c`,\n * whose stationary point is the positive root of\n * `2n·q² − [(b + c) − delta·(b + 3c + 2e)]·q − c·delta·(1 − delta) = 0`.\n * At `delta = 0` this returns `(b + c) / 2n`, the familiar null.\n */\nfunction constrainedLossRate(b: number, c: number, n: number, delta: number): number {\n const e = n - b - c\n const quadratic = 2 * n\n const linear = -(b + c - delta * (b + 3 * c + 2 * e))\n const constant = -c * delta * (1 - delta)\n const discriminant = linear * linear - 4 * quadratic * constant\n const root = discriminant > 0 ? Math.sqrt(discriminant) : 0\n const q = (-linear + root) / (2 * quadratic)\n // Clamp into the region where all four cell probabilities stay non-negative.\n return Math.min(Math.max(q, Math.max(0, -delta)), Math.max(0, (1 - delta) / 2))\n}\n\n/** Tango's score statistic for H0: RD = `delta`. `Var(b − c) = n·(2q + delta −\n * delta²)` under that hypothesis, evaluated at the constrained MLE of q. */\nfunction tangoScore(b: number, c: number, n: number, delta: number): number {\n const numerator = b - c - n * delta\n const q = constrainedLossRate(b, c, n, delta)\n const variance = n * (2 * q + delta * (1 - delta))\n if (!(variance > 0)) {\n if (numerator === 0) return 0\n return numerator > 0 ? Number.POSITIVE_INFINITY : Number.NEGATIVE_INFINITY\n }\n return numerator / Math.sqrt(variance)\n}\n\n/**\n * Paired risk difference with TANGO'S (1998) SCORE INTERVAL — the estimator a\n * promotion gate may decide on **at a nonzero margin**.\n *\n * {@link pairedRiskDifferenceExact} conditions on the observed discordant count\n * `m = b + c`, builds a Clopper-Pearson interval for the win share among those\n * `m` pairs, and multiplies by the observed `m/n`. That is exact for testing\n * RD = 0 — it is dual to McNemar — but it is NOT a confidence interval for the\n * population risk difference at a nonzero margin, because the sampling\n * variability of `m/n` itself is discarded. The gap is not academic: with the\n * production caller's `pairedDeltaThreshold: -0.05`, a process whose true risk\n * difference sits exactly on that margin clears a nominal-95 % `lower > margin`\n * check 24.75 % of the time at n = 40 and 43.95 % at n = 76 (2000 replicates\n * each) when the conditional interval decides.\n *\n * Tango's interval inverts the score test of RD = delta, which estimates the\n * nuisance loss rate under each hypothesised delta instead of fixing it at the\n * observed value, so `m` contributes its own uncertainty. It is the method\n * `ratesci::scorepairci` uses for paired risk-difference noninferiority, and it\n * is not conditional, so it stays valid as the margin moves away from zero.\n *\n * The bounds are found by bisecting `tangoScore(delta) = ±z` — the score is\n * monotone decreasing in delta, so each crossing is unique. Inputs are paired\n * 0/1 (or boolean) arrays, control first. Throws on unequal lengths.\n */\nexport function pairedRiskDifferenceScore(\n control: ArrayLike<number | boolean>,\n treatment: ArrayLike<number | boolean>,\n confidence = 0.95,\n): ScoreRiskDifferenceResult {\n if (control.length !== treatment.length) {\n throw new Error(\n `pairedRiskDifferenceScore: unequal sample sizes (${control.length} vs ${treatment.length})`,\n )\n }\n if (confidence <= 0 || confidence >= 1) {\n throw new Error(`pairedRiskDifferenceScore: confidence must be in (0,1), got ${confidence}`)\n }\n const n = control.length\n if (n === 0) {\n return { n: 0, b: 0, c: 0, nDiscordant: 0, riskDifference: 0, lower: -1, upper: 1, confidence }\n }\n let b = 0\n let c = 0\n for (let i = 0; i < n; i++) {\n const ctrl = control[i] ? 1 : 0\n const treat = treatment[i] ? 1 : 0\n if (treat === 1 && ctrl === 0) b++\n else if (treat === 0 && ctrl === 1) c++\n }\n const riskDifference = (b - c) / n\n const z = zQuantile(1 - (1 - confidence) / 2)\n\n // Lower bound: the smallest delta still inside the interval, i.e. the root of\n // score(delta) = +z on [-1, riskDifference]. score(riskDifference) = 0 < z, so\n // the right endpoint is always inside and the bisection is well posed.\n let lo = -1\n let hi = riskDifference\n // 64 halvings resolve the [-1, 1] search interval below 6e-20, past the\n // precision available to the score calculation. More iterations only\n // repeat identical floating-point values while multiplying every gate\n // decision's cost.\n for (let i = 0; i < 64; i++) {\n const mid = (lo + hi) / 2\n if (tangoScore(b, c, n, mid) > z) lo = mid\n else hi = mid\n }\n const lower = (lo + hi) / 2\n\n // Upper bound: root of score(delta) = -z on [riskDifference, 1].\n let ulo = riskDifference\n let uhi = 1\n for (let i = 0; i < 64; i++) {\n const mid = (ulo + uhi) / 2\n if (tangoScore(b, c, n, mid) > -z) ulo = mid\n else uhi = mid\n }\n const upper = (ulo + uhi) / 2\n\n return {\n n,\n b,\n c,\n nDiscordant: b + c,\n riskDifference,\n lower: Math.max(-1, lower),\n upper: Math.min(1, upper),\n confidence,\n }\n}\n\n/**\n * The common positive level `s` such that EVERY value across both paired arms is\n * exactly 0 or `s` — i.e. the outcome is pass/fail, whatever encoding it arrived\n * in. Returns null when the outcomes are not two-point, when the two arms use\n * different levels, or when no positive value was observed at all (all-zero\n * arms: the level is not identified, and there is nothing to decide anyway).\n *\n * This is the scale-aware successor to {@link isBinaryOutcomeVector}, which only\n * recognises literal {0, 1}. Judges in this codebase emit dimensions on 0-100 as\n * well as [0,1] (see `detectScale` in `campaign/gates/statistical-heldout.ts`),\n * so a pass/fail dimension routinely arrives as {0, 100} and a {0,1}-only test\n * silently sends it down the median path that cannot see it. Any positive level\n * is accepted, not just 1 and 100: for a two-point {0, s} outcome the mean paired\n * delta is exactly s·(b − c)/n, so the binary estimators apply after dividing by\n * s and rescaling the result back into the caller's native units.\n *\n * Non-finite values ⇒ null: an unusable outcome must not be classified as a\n * clean pass/fail shape.\n */\nexport function pairedBinaryScale(\n before: ArrayLike<number>,\n after: ArrayLike<number>,\n): number | null {\n let level: number | null = null\n for (const arm of [before, after]) {\n for (let i = 0; i < arm.length; i++) {\n const v = arm[i]!\n if (!Number.isFinite(v)) return null\n if (v === 0) continue\n if (v < 0) return null\n if (level === null) level = v\n else if (v !== level) return null\n }\n }\n return level\n}\n\n/**\n * Unbiased pass@k for code generation (Chen et al. 2021, \"Evaluating Large\n * Language Models Trained on Code\"). Given `n` independent samples for one\n * problem of which `c` pass, the probability that at least one of a random k of\n * them passes is 1 − C(n−c, k) / C(n, k). Estimating pass@k as \"did any of the\n * first k pass\" is biased high at small n; this is the variance-reduced estimator\n * averaged implicitly over all k-subsets. Average the per-problem values across\n * the suite for the corpus pass@k. Computed in the numerically stable product\n * form. Requires 1 ≤ k ≤ n and 0 ≤ c ≤ n.\n */\nexport function passAtK(n: number, c: number, k: number): number {\n if (!Number.isInteger(n) || !Number.isInteger(c) || !Number.isInteger(k)) {\n throw new Error(`passAtK: n, c, k must be integers (got n=${n}, c=${c}, k=${k})`)\n }\n if (k < 1 || k > n || c < 0 || c > n) {\n throw new Error(`passAtK: require 1 ≤ k ≤ n and 0 ≤ c ≤ n (got n=${n}, c=${c}, k=${k})`)\n }\n if (n - c < k) return 1\n let prob = 1\n for (let i = n - c + 1; i <= n; i++) prob *= 1 - k / i\n return 1 - prob\n}\n"],"mappings":";;;;;;;;;;AAiCA,SAAgB,OAAO,WAAmB,GAAW,aAAa,KAA0B;CAC1F,IAAI,CAAC,OAAO,cAAc,CAAC,KAAK,IAAI,GAClC,MAAM,IAAI,MAAM,gDAAgD;CAElE,IAAI,CAAC,OAAO,cAAc,SAAS,KAAK,YAAY,KAAK,YAAY,GACnE,MAAM,IAAI,MAAM,sBAAsB,UAAU,mBAAmB,EAAE,EAAE;CAEzE,IAAI,CAAC,OAAO,SAAS,UAAU,KAAK,cAAc,KAAK,cAAc,GACnE,MAAM,IAAI,MAAM,iDAAiD;CAEnE,IAAI,MAAM,GAAG,OAAO;EAAE,UAAU;EAAG,OAAO;EAAG,OAAO;CAAE;CACtD,MAAM,IAAI,UAAU,KAAK,IAAI,cAAc,CAAC;CAC5C,MAAM,IAAI,YAAY;CACtB,MAAM,KAAK,IAAI;CACf,MAAM,QAAQ,IAAI,KAAK;CACvB,MAAM,UAAU,IAAI,MAAM,IAAI,MAAM;CACpC,MAAM,OAAQ,IAAI,KAAK,MAAM,KAAK,IAAI,KAAK,MAAM,IAAI,MAAM,CAAC,IAAK;CACjE,OAAO;EACL,UAAU;EACV,OAAO,KAAK,IAAI,GAAG,SAAS,IAAI;EAChC,OAAO,KAAK,IAAI,GAAG,SAAS,IAAI;CAClC;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BA,SAAgB,sBAAsB,QAAoC;CACxE,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;EACtC,MAAM,IAAI,OAAO;EACjB,IAAI,MAAM,KAAK,MAAM,GAAG,OAAO;CACjC;CACA,OAAO;AACT;;;;;;;;;;;;;AA8BA,SAAgB,QACd,SACA,WACe;CACf,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MAAM,kCAAkC,QAAQ,OAAO,MAAM,UAAU,OAAO,EAAE;CAE5F,MAAM,IAAI,QAAQ;CAClB,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,cAAc,IAAI;CACxB,MAAM,YAAY,gBAAgB,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,IAAI,MAAM,IAAI;CACvE,OAAO;EAAE;EAAG;EAAa;EAAG;EAAG;EAAW,QAAQ,qBAAqB,GAAG,CAAC;CAAE;AAC/E;;;;;;;;;;;;;;;;AAmCA,SAAgB,qBACd,SACA,WACA,aAAa,KACS;CACtB,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MACR,+CAA+C,QAAQ,OAAO,MAAM,UAAU,OAAO,EACvF;CAEF,MAAM,IAAI,QAAQ;CAClB,IAAI,MAAM,GAAG,OAAO;EAAE,GAAG;EAAG,GAAG;EAAG,GAAG;EAAG,gBAAgB;EAAG,OAAO;EAAG,OAAO;EAAG;CAAW;CAC1F,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,MAAM,IAAI,KAAK;CACrB,MAAM,YAAY,IAAI,KAAK,IAAI,MAAM,IAAI,MAAM,IAAI;CAEnD,MAAM,OADI,UAAU,KAAK,IAAI,cAAc,CAC9B,IAAI,KAAK,KAAK,KAAK,IAAI,GAAG,QAAQ,CAAC;CAChD,OAAO;EACL;EACA;EACA;EACA,gBAAgB;EAChB,OAAO,KAAK,IAAI,IAAI,KAAK,IAAI;EAC7B,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI;EAC5B;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsDA,SAAgB,0BACd,SACA,WACA,aAAa,KACc;CAC3B,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MACR,oDAAoD,QAAQ,OAAO,MAAM,UAAU,OAAO,EAC5F;CAEF,IAAI,cAAc,KAAK,cAAc,GACnC,MAAM,IAAI,MAAM,+DAA+D,YAAY;CAE7F,MAAM,IAAI,QAAQ;CAClB,IAAI,MAAM,GACR,OAAO;EACL,GAAG;EACH,GAAG;EACH,GAAG;EACH,aAAa;EACb,gBAAgB;EAChB,OAAO;EACP,OAAO;EACP;EACA,QAAQ;CACV;CAEF,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,IAAI,IAAI;CACd,MAAM,kBAAkB,IAAI,KAAK;CACjC,MAAM,SAAS,qBAAqB,GAAG,CAAC;CACxC,IAAI,MAAM,GACR,OAAO;EAAE;EAAG;EAAG;EAAG,aAAa;EAAG,gBAAgB;EAAG,OAAO;EAAG,OAAO;EAAG;EAAY;CAAO;CAE9F,MAAM,QAAQ,IAAI;CAClB,MAAM,QAAQ,MAAM,IAAI,IAAI,aAAa,QAAQ,GAAG,GAAG,IAAI,IAAI,CAAC;CAChE,MAAM,SAAS,MAAM,IAAI,IAAI,aAAa,IAAI,QAAQ,GAAG,IAAI,GAAG,IAAI,CAAC;CACrE,MAAM,QAAQ,IAAI;CAClB,OAAO;EACL;EACA;EACA;EACA,aAAa;EACb;EACA,OAAO,KAAK,IAAI,KAAK,IAAI,QAAQ,KAAK,KAAK;EAC3C,OAAO,KAAK,IAAI,IAAI,IAAI,SAAS,KAAK,KAAK;EAC3C;EACA;CACF;AACF;;;;;AAMA,SAAS,aAAa,GAAW,GAAW,GAAmB;CAC7D,IAAI,KAAK,GAAG,OAAO;CACnB,IAAI,KAAK,GAAG,OAAO;CACnB,IAAI,KAAK;CACT,IAAI,KAAK;CACT,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK;EAC3B,MAAM,OAAO,KAAK,MAAM;EACxB,IAAI,0BAA0B,KAAK,GAAG,CAAC,IAAI,GAAG,KAAK;OAC9C,KAAK;CACZ;CACA,QAAQ,KAAK,MAAM;AACrB;;;;;;;;;;AAgCA,SAAS,oBAAoB,GAAW,GAAW,GAAW,OAAuB;CACnF,MAAM,IAAI,IAAI,IAAI;CAClB,MAAM,YAAY,IAAI;CACtB,MAAM,SAAS,EAAE,IAAI,IAAI,SAAS,IAAI,IAAI,IAAI,IAAI;CAClD,MAAM,WAAW,CAAC,IAAI,SAAS,IAAI;CACnC,MAAM,eAAe,SAAS,SAAS,IAAI,YAAY;CACvD,MAAM,OAAO,eAAe,IAAI,KAAK,KAAK,YAAY,IAAI;CAC1D,MAAM,KAAK,CAAC,SAAS,SAAS,IAAI;CAElC,OAAO,KAAK,IAAI,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,KAAK,CAAC,GAAG,KAAK,IAAI,IAAI,IAAI,SAAS,CAAC,CAAC;AAChF;;;AAIA,SAAS,WAAW,GAAW,GAAW,GAAW,OAAuB;CAC1E,MAAM,YAAY,IAAI,IAAI,IAAI;CAE9B,MAAM,WAAW,KAAK,IADZ,oBAAoB,GAAG,GAAG,GAAG,KACb,IAAI,SAAS,IAAI;CAC3C,IAAI,EAAE,WAAW,IAAI;EACnB,IAAI,cAAc,GAAG,OAAO;EAC5B,OAAO,YAAY,IAAI,OAAO,oBAAoB,OAAO;CAC3D;CACA,OAAO,YAAY,KAAK,KAAK,QAAQ;AACvC;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BA,SAAgB,0BACd,SACA,WACA,aAAa,KACc;CAC3B,IAAI,QAAQ,WAAW,UAAU,QAC/B,MAAM,IAAI,MACR,oDAAoD,QAAQ,OAAO,MAAM,UAAU,OAAO,EAC5F;CAEF,IAAI,cAAc,KAAK,cAAc,GACnC,MAAM,IAAI,MAAM,+DAA+D,YAAY;CAE7F,MAAM,IAAI,QAAQ;CAClB,IAAI,MAAM,GACR,OAAO;EAAE,GAAG;EAAG,GAAG;EAAG,GAAG;EAAG,aAAa;EAAG,gBAAgB;EAAG,OAAO;EAAI,OAAO;EAAG;CAAW;CAEhG,IAAI,IAAI;CACR,IAAI,IAAI;CACR,KAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;EAC1B,MAAM,OAAO,QAAQ,KAAK,IAAI;EAC9B,MAAM,QAAQ,UAAU,KAAK,IAAI;EACjC,IAAI,UAAU,KAAK,SAAS,GAAG;OAC1B,IAAI,UAAU,KAAK,SAAS,GAAG;CACtC;CACA,MAAM,kBAAkB,IAAI,KAAK;CACjC,MAAM,IAAI,UAAU,KAAK,IAAI,cAAc,CAAC;CAK5C,IAAI,KAAK;CACT,IAAI,KAAK;CAKT,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK;EAC3B,MAAM,OAAO,KAAK,MAAM;EACxB,IAAI,WAAW,GAAG,GAAG,GAAG,GAAG,IAAI,GAAG,KAAK;OAClC,KAAK;CACZ;CACA,MAAM,SAAS,KAAK,MAAM;CAG1B,IAAI,MAAM;CACV,IAAI,MAAM;CACV,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK;EAC3B,MAAM,OAAO,MAAM,OAAO;EAC1B,IAAI,WAAW,GAAG,GAAG,GAAG,GAAG,IAAI,CAAC,GAAG,MAAM;OACpC,MAAM;CACb;CACA,MAAM,SAAS,MAAM,OAAO;CAE5B,OAAO;EACL;EACA;EACA;EACA,aAAa,IAAI;EACjB;EACA,OAAO,KAAK,IAAI,IAAI,KAAK;EACzB,OAAO,KAAK,IAAI,GAAG,KAAK;EACxB;CACF;AACF;;;;;;;;;;;;;;;;;;;;AAqBA,SAAgB,kBACd,QACA,OACe;CACf,IAAI,QAAuB;CAC3B,KAAK,MAAM,OAAO,CAAC,QAAQ,KAAK,GAC9B,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;EACnC,MAAM,IAAI,IAAI;EACd,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;EAChC,IAAI,MAAM,GAAG;EACb,IAAI,IAAI,GAAG,OAAO;EAClB,IAAI,UAAU,MAAM,QAAQ;OACvB,IAAI,MAAM,OAAO,OAAO;CAC/B;CAEF,OAAO;AACT;;;;;;;;;;;AAYA,SAAgB,QAAQ,GAAW,GAAW,GAAmB;CAC/D,IAAI,CAAC,OAAO,UAAU,CAAC,KAAK,CAAC,OAAO,UAAU,CAAC,KAAK,CAAC,OAAO,UAAU,CAAC,GACrE,MAAM,IAAI,MAAM,4CAA4C,EAAE,MAAM,EAAE,MAAM,EAAE,EAAE;CAElF,IAAI,IAAI,KAAK,IAAI,KAAK,IAAI,KAAK,IAAI,GACjC,MAAM,IAAI,MAAM,mDAAmD,EAAE,MAAM,EAAE,MAAM,EAAE,EAAE;CAEzF,IAAI,IAAI,IAAI,GAAG,OAAO;CACtB,IAAI,OAAO;CACX,KAAK,IAAI,IAAI,IAAI,IAAI,GAAG,KAAK,GAAG,KAAK,QAAQ,IAAI,IAAI;CACrD,OAAO,IAAI;AACb"}

package/dist/produced-state-Bzlh09oj.js.map DELETED Viewed

@@ -1 +0,0 @@

- {"version":3,"file":"produced-state-Bzlh09oj.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/release-confidence-CkoXFx9I.d.ts.map DELETED Viewed

@@ -1 +0,0 @@

- {"version":3,"file":"release-confidence-CkoXFx9I.d.ts","names":[],"sources":["../src/statistics/rank-tests.ts","../src/promotion-gate.ts","../src/release-confidence.ts"],"mappings":";;;;;;KAiCY;;;;;KAMA;UAEK;;EAEf,SAAS;;EAET;;;EAGA;;;cAIW;;cAEA;;cAEA;;cAEA;UAEI;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER;;;;;;;;;;;;;;iBAec,aACd,aACA,aACA,OAAM,kBACL;UAuFc;;;EAGf;;EAEA;;EAEA,QAAQ;;EAER;;EAEA;;;;;;;;;;;;;;;iBAgBc,mBACd,kBACA,iBACA,OAAM,kBACL;;;;;;;;;;;;;;;;;;;;;;;;;;;KCpLS;UAEK;EACf;EACA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,SAAS;;UAGM;;EAEf;;EAEA;;;;;;EAMA;;;EAGA;;;;;;;;;iBAUc,YACd,oBACA,qBACA,UAAS,mBACR;;;;;;;;;;;;;;;;UA0Gc,oBAAoB;EACnC,iBAAiB;EACjB,kBAAkB;;EAElB,QAAQ,QAAQ,YAAY;EAC5B;EACA;;EAEA;;EAEA;;;;;iBAMoB,gBAAgB,SACpC,MAAM,oBAAoB,WACzB,QAAQ;EAAoB;EAAyB;;;;;KC7K5C;;UAGK;;EAEf;;EAEA;EACA,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;EACA,WAAW;;KAGD;KACA;UAQK;EACf;EACA;EACA,QAAQ;EACR;EACA;EACA;EACA;EACA;;EAEA,eAAe;EACf,MAAM;EACN,WAAW;;UAGI;;EAEf;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,UAAU;EACV,qBAAqB;EACrB,gBAAgB;EAChB,kBAAkB;EAClB,eAAe;EACf,aAAa;;UAGE;EACf,MAAM;EACN,QAAQ;EACR;EACA;;UAGe;EACf,MAAM;EACN;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,aAAa,OAAO;EACpB,cAAc;EACd,oBAAoB,QAAQ,OAAO;EACnC,0BAA0B;;;;;;;EAO1B;;UAGe;EACf;EACA;EACA;EACA,QAAQ;EACR;EACA,MAAM;EACN,QAAQ;EACR,SAAS;EACT,SAAS;EACT,cAAc;EACd;;iBAkBc,0BACd,OAAO,yBACN;iBA4Ha,wBAAwB,OAAO,yBAAyB"}

package/dist/rollout-l9w3Gdvh.js.map DELETED Viewed

@@ -1 +0,0 @@

- {"version":3,"file":"rollout-l9w3Gdvh.js","names":["isRecord"],"sources":["../src/rollout/interchange/harbor.ts","../src/rollout/readers/opencode-sqlite.ts"],"sourcesContent":["/**\n * Harbor ATIF-v1.7 interchange — `tangle.rollout.v1` ⇄ Agent Trajectory\n * Interchange Format.\n *\n * ATIF is the portability format (spec:\n * https://www.harborframework.com/docs/agents/trajectory-format, normative\n * RFC: harbor-framework/harbor `rfcs/0001-trajectory-format.md`). It sits\n * BELOW the waist of the rollout hourglass in both directions — export reads\n * `RolloutLine[]`, import writes `RolloutLine[]` — and it is never a source\n * of training labels:\n *\n * ATIF models NO reward, NO judge verdict, NO task/split coordinates.\n *\n * Consequences, both deliberate:\n * - EXPORT drops `outcome.reward`, `outcome.reward_source` and\n * `outcome.verdict` entirely. They are not smuggled into `extra`: a\n * third-party reading our ATIF file must not be able to mistake an\n * agent-eval judge score for something ATIF sanctioned.\n * - IMPORT therefore mints UNLABELED lines: `reward: null` (the existing\n * \"null reward is a labeled gap, never 0\" semantics), `verdict: null`,\n * and a `provenance.gap` naming the missing label. An imported\n * trajectory is not a training example until a judge scores it.\n *\n * Everything else we own that ATIF has no field for travels in a namespaced\n * escrow at `extra.tangle.*`, so our own round-trip is exact while a foreign\n * reader can ignore it. Fields that neither ATIF nor the escrow can carry\n * come back explicitly null / fail-closed, never invented.\n *\n * THE ESCROW IS NAMESPACED, NOT AUTHENTICATED. Anyone can write\n * `extra.tangle.*` into a file. So the escrow may restore what a value IS, but\n * never what a line is ALLOWED to do: `task.split` is forced to `holdout` on\n * every import regardless of what the document claims, and promoting an\n * imported trajectory to a trainable split is an explicit, greppable act\n * (`relabelImportedSplit`) rather than a property of the file. The document\n * keeps its claim — the claim just is not authority.\n *\n * Multi-agent shape differs on purpose. ATIF EMBEDS children in\n * `subagent_trajectories`; we keep a flat ledger with a normalized\n * `parent_rollout_id` edge. Export assembles the tree, import flattens it.\n * `session_id` is RUN-scoped in ATIF, so it carries `run_id` — the coordinate\n * that is shared by every invocation of one run — not `rollout_id`, which\n * identifies a single invocation and would split one run across session ids.\n *\n * ROUND-TRIPPING IS IDEMPOTENT: `import(export(import(export(x))))` is\n * byte-identical to `import(export(x))`. Import composes `provenance.gap` as a\n * de-duplicated ordered set rather than appending, and it emits every\n * `ChatMessage` with keys in the canonical schema order (role, content,\n * reasoning_content, tool_calls, tool_call_id, name, is_copied_context), so a\n * ledger hashed on serialized bytes sees no diff across further passes. The\n * FIRST import may re-order a producer's keys — that is the canonicalization.\n *\n * NOT building a Letta converter. Letta's trajectory-v1 is a strict subset of\n * what we need from ATIF here — no per-step or aggregate cost, no\n * multi-agent/subagent structure, no token-id or logprob channel — so a Letta\n * sink would carry less than this one and add a second format to keep\n * correct. Decision recorded in docs/rollout.md; do not re-litigate without a\n * concrete consumer that reads Letta and cannot read ATIF.\n */\n\nimport {\n assertRolloutLine,\n type ChatMessage,\n type ChatRole,\n type ChatToolCall,\n type GatedEvidence,\n ROLLOUT_CAPTURES,\n ROLLOUT_ROLES,\n ROLLOUT_SCHEMA,\n ROLLOUT_SPLITS,\n type RolloutArtifacts,\n type RolloutCapture,\n type RolloutCostBlock,\n type RolloutLine,\n type RolloutPolicy,\n type RolloutRole,\n type RolloutSplit,\n type RolloutStep,\n type RolloutTask,\n type ToolDef,\n} from '../schema'\n\nexport const ATIF_SCHEMA_VERSION = 'ATIF-v1.7'\n\n/** Gap note on every imported line — ATIF carries no verdict, so nothing is scored. */\nexport const HARBOR_IMPORT_GAP = 'imported from Harbor ATIF; no verdict'\n\n/** Namespaced escrow key for everything ATIF does not model but we must not lose. */\nconst ESCROW = 'tangle'\n\n/**\n * Marks a step we synthesized purely to hold tool results that answered no\n * assistant turn IN THIS DOCUMENT. ATIF has no `tool` source, so such results\n * need a carrier step; the marker lets import drop the carrier instead of\n * inventing a turn.\n *\n * A carrier's results carry NO `source_call_id`. RFC 0001 rule 2 requires every\n * `source_call_id` to match a `tool_call_id` in the same step's `tool_calls`,\n * and a carrier declares no calls (a `system` step cannot legally make one), so\n * emitting the id there would produce an invalid document. The id is escrowed\n * instead and restored verbatim on import — the link survives without the\n * document lying about who made the call.\n */\nconst TOOL_RESULTS_ONLY = 'tool-results-only'\n\n// ---------------------------------------------------------------------------\n// ATIF-v1.7 wire types (RFC 0001). Optional fields are optional here too.\n// ---------------------------------------------------------------------------\n\nexport type HarborStepSource = 'system' | 'user' | 'agent'\n\nexport interface HarborImageSource {\n media_type: string\n path: string\n}\n\nexport interface HarborContentPart {\n type: 'text' | 'image'\n text?: string\n source?: HarborImageSource\n}\n\nexport interface HarborToolCall {\n tool_call_id: string\n function_name: string\n /** ATIF requires a decoded JSON object here, unlike our raw argument string. */\n arguments: Record<string, unknown>\n extra?: Record<string, unknown>\n}\n\nexport interface HarborSubagentTrajectoryRef {\n trajectory_id?: string\n trajectory_path?: string\n /** Informational only since v1.7 — never a resolution key. */\n session_id?: string\n extra?: Record<string, unknown>\n}\n\nexport interface HarborObservationResult {\n source_call_id?: string\n content?: string | HarborContentPart[]\n subagent_trajectory_ref?: HarborSubagentTrajectoryRef[]\n extra?: Record<string, unknown>\n}\n\nexport interface HarborObservation {\n results: HarborObservationResult[]\n}\n\nexport interface HarborMetrics {\n prompt_tokens?: number\n completion_tokens?: number\n cached_tokens?: number\n cost_usd?: number\n prompt_token_ids?: number[]\n completion_token_ids?: number[]\n logprobs?: number[]\n extra?: Record<string, unknown>\n}\n\nexport interface HarborStep {\n /** Ordinal, sequential from 1. */\n step_id: number\n timestamp?: string\n source: HarborStepSource\n model_name?: string\n reasoning_effort?: string | number\n message: string | HarborContentPart[]\n reasoning_content?: string\n tool_calls?: HarborToolCall[]\n observation?: HarborObservation\n metrics?: HarborMetrics\n llm_call_count?: number\n is_copied_context?: boolean\n extra?: Record<string, unknown>\n}\n\nexport interface HarborAgent {\n name: string\n version: string\n model_name?: string\n /** OpenAI function-calling schema — byte-identical to our `ToolDef`. */\n tool_definitions?: ToolDef[]\n extra?: Record<string, unknown>\n}\n\nexport interface HarborFinalMetrics {\n total_prompt_tokens?: number\n total_completion_tokens?: number\n total_cached_tokens?: number\n total_cost_usd?: number\n total_steps?: number\n extra?: Record<string, unknown>\n}\n\nexport interface HarborTrajectory {\n schema_version: string\n session_id?: string\n /** Required on embedded subagents; we always set it so lines stay joinable. */\n trajectory_id?: string\n agent: HarborAgent\n steps: HarborStep[]\n notes?: string\n final_metrics?: HarborFinalMetrics\n continued_trajectory_ref?: string\n subagent_trajectories?: HarborTrajectory[]\n extra?: Record<string, unknown>\n}\n\n// ---------------------------------------------------------------------------\n// Shared helpers\n// ---------------------------------------------------------------------------\n\nconst isRecord = (v: unknown): v is Record<string, unknown> =>\n typeof v === 'object' && v !== null && !Array.isArray(v)\n\nconst asString = (v: unknown): string | undefined => (typeof v === 'string' ? v : undefined)\n\nconst asNumberOrNull = (v: unknown): number | null =>\n typeof v === 'number' && Number.isFinite(v) ? v : null\n\nconst asIntegerOrUndefined = (v: unknown): number | undefined =>\n Number.isInteger(v) ? (v as number) : undefined\n\nconst asNumberArray = (v: unknown): number[] | undefined =>\n Array.isArray(v) && v.every((n) => typeof n === 'number' && Number.isFinite(n))\n ? (v as number[])\n : undefined\n\nconst asIntegerArray = (v: unknown): number[] | undefined =>\n Array.isArray(v) && v.every((n) => Number.isInteger(n)) ? (v as number[]) : undefined\n\n/**\n * Builds a `ChatMessage` with keys in the schema's declaration order.\n *\n * Object key order is insertion order in JS, so a message assembled\n * conditionally field-by-field serializes differently depending on which\n * optional fields were present — which makes a byte-hashed ledger see a diff\n * across an import that changed nothing. One builder, one order, stable bytes.\n */\nfunction canonicalChatMessage(parts: {\n role: ChatRole\n content: string | null\n reasoning_content?: string\n tool_calls?: ChatToolCall[]\n tool_call_id?: string\n name?: string\n is_copied_context?: boolean\n}): ChatMessage {\n const message: ChatMessage = { role: parts.role, content: parts.content }\n if (parts.reasoning_content !== undefined) message.reasoning_content = parts.reasoning_content\n if (parts.tool_calls !== undefined) message.tool_calls = parts.tool_calls\n if (parts.tool_call_id !== undefined) message.tool_call_id = parts.tool_call_id\n if (parts.name !== undefined) message.name = parts.name\n if (parts.is_copied_context === true) message.is_copied_context = true\n return message\n}\n\n/** Reads the `extra.tangle` escrow, tolerating a foreign file that has none. */\nfunction escrowOf(extra: Record<string, unknown> | undefined): Record<string, unknown> | undefined {\n if (!isRecord(extra)) return undefined\n const value = extra[ESCROW]\n return isRecord(value) ? value : undefined\n}\n\nfunction escrowSection(\n escrow: Record<string, unknown> | undefined,\n key: string,\n): Record<string, unknown> | undefined {\n if (escrow === undefined) return undefined\n const value = escrow[key]\n return isRecord(value) ? value : undefined\n}\n\n// ---------------------------------------------------------------------------\n// Export: RolloutLine[] → ATIF trajectory tree\n// ---------------------------------------------------------------------------\n\nfunction toHarborToolCall(call: ChatToolCall): HarborToolCall {\n // ATIF requires `arguments` to be a JSON object; ours is the raw\n // JSON-encoded string the model emitted, and models really do emit\n // malformed ones. Keep the exact bytes in `extra` so import restores the\n // string verbatim instead of re-serializing a normalized parse.\n let parsed: Record<string, unknown> | undefined\n try {\n const decoded: unknown = JSON.parse(call.function.arguments)\n if (isRecord(decoded)) parsed = decoded\n } catch {\n parsed = undefined\n }\n return {\n tool_call_id: call.id,\n function_name: call.function.name,\n arguments: parsed ?? {},\n extra: { [ESCROW]: { arguments_raw: call.function.arguments } },\n }\n}\n\n/**\n * `linked` = this result is attached to the step that actually declared the\n * call, so `source_call_id` is legal (RFC 0001 rule 2). When it is not, the id\n * goes to escrow instead of onto the wire.\n */\nfunction toObservationResult(message: ChatMessage, linked: boolean): HarborObservationResult {\n const result: HarborObservationResult = {}\n if (linked && message.tool_call_id !== undefined) result.source_call_id = message.tool_call_id\n if (message.content !== null) result.content = message.content\n const escrow: Record<string, unknown> = {}\n if (!linked && message.tool_call_id !== undefined) escrow.source_call_id = message.tool_call_id\n if (message.name !== undefined) escrow.name = message.name\n // `content: null` on a tool turn is meaningful (a tool that returned\n // nothing) and is not the same as the empty string ATIF would round-trip it to.\n if (message.content === null) escrow.content_null = true\n if (message.reasoning_content !== undefined) escrow.reasoning_content = message.reasoning_content\n if (message.is_copied_context === true) escrow.is_copied_context = true\n if (Object.keys(escrow).length > 0) result.extra = { [ESCROW]: escrow }\n return result\n}\n\n/** True when `step` is the agent turn that declared `callId` — the rule-2 test. */\nfunction declaresCall(step: HarborStep, callId: string | undefined): boolean {\n if (callId === undefined) return false\n return step.tool_calls?.some((call) => call.tool_call_id === callId) === true\n}\n\n/**\n * Attaches the four RL fields lifted from ATIF onto the agent step the span\n * describes. They are ALSO escrowed with the whole span under `extra.tangle`,\n * on purpose: the escrow is how our own round-trip stays exact, and these\n * native fields are how a foreign consumer — which never reads our escrow —\n * gets the logprobs and token ids at all. A field carried only in the escrow is\n * not an interchange field.\n */\nfunction applySpanMetrics(step: HarborStep, span: RolloutStep | undefined): void {\n if (span === undefined) return\n if (span.llm_call_count !== undefined) step.llm_call_count = span.llm_call_count\n const metrics: HarborMetrics = {}\n if (span.prompt_token_ids !== undefined) metrics.prompt_token_ids = span.prompt_token_ids\n if (span.completion_token_ids !== undefined) {\n metrics.completion_token_ids = span.completion_token_ids\n }\n if (span.logprobs !== undefined) metrics.logprobs = span.logprobs\n if (Object.keys(metrics).length > 0) step.metrics = metrics\n}\n\n/**\n * Coalescing fold, not a 1:1 map: ATIF has no `tool` source, so a tool result\n * becomes an `observation.results[]` entry on the agent step THAT DECLARED ITS\n * `tool_call_id`.\n *\n * Linking by id, not by adjacency. Adjacency emitted invalid documents in three\n * shapes the RFC forbids — an assistant turn with no `tool_calls` swallowing the\n * results that followed it, a result whose id matched none of the preceding\n * step's calls, and an unanswered result riding a `system` step that carried a\n * `source_call_id` it could never declare. Attaching only to the step that owns\n * the id also preserves message order exactly: a result either joins the step\n * immediately before it or becomes its own carrier at its own position.\n *\n * `line.steps` span projections are matched to agent steps in order (k-th llm\n * span → k-th agent step) purely to fill ATIF's native per-step `metrics`;\n * extra spans on either side are simply not matched, never invented.\n */\nfunction messagesToSteps(line: RolloutLine): HarborStep[] {\n const steps: HarborStep[] = []\n const llmSpans = (line.steps ?? []).filter((span) => span.kind === 'llm')\n let nextSpan = 0\n for (const message of line.messages) {\n if (message.role === 'tool') {\n const previous = steps[steps.length - 1]\n if (previous !== undefined && declaresCall(previous, message.tool_call_id)) {\n if (previous.observation === undefined) previous.observation = { results: [] }\n previous.observation.results.push(toObservationResult(message, true))\n continue\n }\n // A tool result answering no assistant turn in this document (truncated or\n // reconstructed transcript). It still has to survive, so it rides a marked\n // carrier step whose observation states no call id.\n const carrier: HarborStep = {\n step_id: steps.length + 1,\n source: 'system',\n message: '',\n observation: { results: [toObservationResult(message, false)] },\n extra: { [ESCROW]: { synthetic: TOOL_RESULTS_ONLY } },\n }\n if (message.is_copied_context === true) carrier.is_copied_context = true\n steps.push(carrier)\n continue\n }\n const step: HarborStep = {\n step_id: steps.length + 1,\n source: message.role === 'assistant' ? 'agent' : message.role,\n message: message.content ?? '',\n }\n const escrow: Record<string, unknown> = {}\n if (message.content === null) escrow.content_null = true\n if (message.name !== undefined) escrow.name = message.name\n if (message.role === 'assistant') {\n if (line.policy.model !== null) step.model_name = line.policy.model\n if (message.reasoning_content !== undefined) {\n step.reasoning_content = message.reasoning_content\n }\n if (message.tool_calls !== undefined && message.tool_calls.length > 0) {\n step.tool_calls = message.tool_calls.map(toHarborToolCall)\n }\n applySpanMetrics(step, llmSpans[nextSpan])\n nextSpan += 1\n } else if (message.reasoning_content !== undefined) {\n // ATIF confines `reasoning_content` to agent steps; ours is not confined.\n escrow.reasoning_content = message.reasoning_content\n }\n // RFC 0001 rule 7: a copied-context turn was not authored by this agent, and\n // an SFT pipeline MUST exclude it. Native ATIF field, both directions.\n if (message.is_copied_context === true) step.is_copied_context = true\n if (Object.keys(escrow).length > 0) step.extra = { [ESCROW]: escrow }\n steps.push(step)\n }\n return steps\n}\n\nfunction finalMetricsOf(line: RolloutLine, stepCount: number): HarborFinalMetrics {\n const metrics: HarborFinalMetrics = {}\n if (line.cost.tokens_in !== null) metrics.total_prompt_tokens = line.cost.tokens_in\n if (line.cost.tokens_out !== null) metrics.total_completion_tokens = line.cost.tokens_out\n if (line.cost.cache_read !== null) metrics.total_cached_tokens = line.cost.cache_read\n // Omitted, never 0, when cost was not captured — a fake 0 is a lie about spend.\n if (line.cost.usd !== null) metrics.total_cost_usd = line.cost.usd\n metrics.total_steps = stepCount\n // ATIF has no aggregate field for these three; `reasoning_tokens` under\n // `extra` is the key the RFC's own worked example uses.\n const extra: Record<string, unknown> = {}\n if (line.cost.tokens_reasoning !== null) extra.reasoning_tokens = line.cost.tokens_reasoning\n if (line.cost.cache_write !== null) extra.cache_write_tokens = line.cost.cache_write\n if (line.cost.wall_s !== null) extra.wall_s = line.cost.wall_s\n if (line.cost.llm_call_count !== undefined && line.cost.llm_call_count !== null) {\n extra.llm_call_count = line.cost.llm_call_count\n }\n if (Object.keys(extra).length > 0) metrics.extra = extra\n return metrics\n}\n\nfunction notesOf(line: RolloutLine): string | undefined {\n const parts: string[] = []\n if (line.provenance.gap !== undefined) parts.push(line.provenance.gap)\n if (line.outcome.realness_gated === true) {\n // ATIF has no gate field. A third-party consumer that ignores `extra`\n // would otherwise see a gamed trajectory with nothing marking it, so the\n // flag is stated in prose as well. This export carries no reward at all,\n // so it is not a training-data door — but it is an audit record.\n parts.push('realness-gated: this run faked its success signal (anti-Goodhart gate fired)')\n }\n return parts.length > 0 ? parts.join(' | ') : undefined\n}\n\nfunction toTrajectoryNode(line: RolloutLine): HarborTrajectory {\n const steps = messagesToSteps(line)\n const trajectory: HarborTrajectory = {\n schema_version: ATIF_SCHEMA_VERSION,\n // RUN-scoped, per the spec — so it is `run_id`, the coordinate every\n // invocation of one run shares. `rollout_id` identifies a single\n // invocation: using it gave two roots of the same run different session\n // ids, and foreign tooling that groups by session_id would split the run.\n session_id: line.run_id,\n trajectory_id: line.rollout_id,\n agent: {\n // ATIF requires both; ours are nullable, so a null is substituted here\n // and restored on import from the escrowed policy — never guessed back.\n name: line.policy.harness ?? 'unknown',\n version: line.policy.harness_version ?? '0.0.0',\n ...(line.policy.model !== null ? { model_name: line.policy.model } : {}),\n ...(line.tool_defs.length > 0 ? { tool_definitions: line.tool_defs } : {}),\n },\n steps,\n final_metrics: finalMetricsOf(line, steps.length),\n extra: {\n [ESCROW]: {\n schema: line.schema,\n rollout_id: line.rollout_id,\n parent_rollout_id: line.parent_rollout_id,\n run_id: line.run_id,\n ...(line.experiment_id !== undefined ? { experiment_id: line.experiment_id } : {}),\n ...(line.candidate_id !== undefined ? { candidate_id: line.candidate_id } : {}),\n generation: line.generation,\n candidate_index: line.candidate_index,\n role: line.role,\n task: line.task,\n policy: line.policy,\n // Span projections overlap the same turns as `steps[]`; folding them\n // into ATIF steps would double-count the run, so they stay escrowed.\n ...(line.steps !== undefined ? { spans: line.steps } : {}),\n outcome: {\n // reward / reward_source / verdict are deliberately absent — see the\n // module header. Only the non-scalar outcome fields travel.\n metrics: line.outcome.metrics,\n is_completed: line.outcome.is_completed,\n is_truncated: line.outcome.is_truncated,\n error: line.outcome.error,\n ...(line.outcome.realness_gated !== undefined\n ? { realness_gated: line.outcome.realness_gated }\n : {}),\n },\n artifacts: line.artifacts,\n provenance: line.provenance,\n },\n },\n }\n const notes = notesOf(line)\n if (notes !== undefined) trajectory.notes = notes\n return trajectory\n}\n\ninterface Forest {\n roots: RolloutLine[]\n childrenOf: Map<string, RolloutLine[]>\n}\n\nfunction buildForest(lines: RolloutLine[]): Forest {\n const byId = new Map<string, RolloutLine>()\n for (const line of lines) {\n if (byId.has(line.rollout_id)) {\n throw new Error(`duplicate rollout_id in input: ${line.rollout_id}`)\n }\n byId.set(line.rollout_id, line)\n }\n const roots: RolloutLine[] = []\n const childrenOf = new Map<string, RolloutLine[]>()\n for (const line of lines) {\n const parent = line.parent_rollout_id\n // A parent outside this set is a root OF WHAT WE HAVE: the edge is kept in\n // escrow so a later import restores the dangling pointer instead of nulling it.\n if (parent === null || !byId.has(parent)) {\n roots.push(line)\n continue\n }\n const siblings = childrenOf.get(parent)\n if (siblings === undefined) childrenOf.set(parent, [line])\n else siblings.push(line)\n }\n // Every line has a parent inside the set: the edges form a cycle with no\n // root. Returning zero documents would drop the whole episode silently.\n if (lines.length > 0 && roots.length === 0) {\n throw new Error(\n `parent_rollout_id cycle: no root among ${lines.length} lines (${lines.map((l) => l.rollout_id).join(', ')})`,\n )\n }\n return { roots, childrenOf }\n}\n\nfunction assemble(line: RolloutLine, forest: Forest, onPath: Set<string>): HarborTrajectory {\n if (onPath.has(line.rollout_id)) {\n throw new Error(`parent_rollout_id cycle through rollout_id ${line.rollout_id}`)\n }\n onPath.add(line.rollout_id)\n const node = toTrajectoryNode(line)\n const children = forest.childrenOf.get(line.rollout_id)\n if (children !== undefined && children.length > 0) {\n node.subagent_trajectories = children.map((child) => assemble(child, forest, onPath))\n }\n onPath.delete(line.rollout_id)\n return node\n}\n\n/**\n * Assemble one episode's flat lines into a single ATIF trajectory tree,\n * linked by `parent_rollout_id`.\n *\n * Reward, verdict and split are NOT emitted (ATIF models none of them); the\n * split and the rest of the task coordinates survive only in `extra.tangle`.\n *\n * We deliberately do NOT synthesize an `observation.subagent_trajectory_ref`\n * pointing at each child: our ledger records WHICH invocation spawned a\n * worker, not which STEP did, and attaching the ref to a guessed step would\n * fabricate a causal claim. Children are embedded in `subagent_trajectories`\n * (each with the `trajectory_id` the spec requires) and the edge is stated in\n * the child's escrowed `parent_rollout_id`.\n *\n * Throws when the lines are not one tree — use `toHarborTrajectories` for a forest.\n */\nexport function toHarborTrajectory(lines: RolloutLine[]): HarborTrajectory {\n const trees = toHarborTrajectories(lines)\n if (trees.length === 0) throw new Error('toHarborTrajectory: no lines')\n if (trees.length > 1) {\n const ids = trees.map((t) => t.trajectory_id ?? '?').join(', ')\n throw new Error(\n `toHarborTrajectory: ${trees.length} roots (${ids}) — ATIF is one tree per document; use toHarborTrajectories`,\n )\n }\n return trees[0]!\n}\n\n/** Every independent tree in the input, one ATIF document each. */\nexport function toHarborTrajectories(lines: RolloutLine[]): HarborTrajectory[] {\n const forest = buildForest(lines)\n return forest.roots.map((root) => assemble(root, forest, new Set()))\n}\n\n// ---------------------------------------------------------------------------\n// Import: ATIF trajectory tree → RolloutLine[]\n// ---------------------------------------------------------------------------\n\nfunction contentToString(message: string | HarborContentPart[]): string {\n if (typeof message === 'string') return message\n return message\n .map((part) =>\n part.type === 'image'\n ? // Our chat content is text-only. Describing the image is honest about\n // what the source held; dropping it silently would not be.\n `[image ${part.source?.media_type ?? 'unknown'} ${part.source?.path ?? ''}]`.trim()\n : (part.text ?? ''),\n )\n .join('\\n')\n}\n\nfunction fromHarborToolCall(call: HarborToolCall): ChatToolCall {\n const raw = asString(escrowOf(call.extra)?.arguments_raw)\n return {\n id: call.tool_call_id,\n type: 'function',\n function: { name: call.function_name, arguments: raw ?? JSON.stringify(call.arguments ?? {}) },\n }\n}\n\nfunction fromObservationResult(\n result: HarborObservationResult,\n callId: string | undefined,\n): ChatMessage {\n const escrow = escrowOf(result.extra)\n return canonicalChatMessage({\n role: 'tool',\n content:\n escrow?.content_null === true\n ? null\n : result.content === undefined\n ? ''\n : contentToString(result.content),\n reasoning_content: asString(escrow?.reasoning_content),\n tool_call_id: callId,\n name: asString(escrow?.name),\n is_copied_context: escrow?.is_copied_context === true,\n })\n}\n\nfunction stepsToMessages(steps: HarborStep[]): ChatMessage[] {\n const messages: ChatMessage[] = []\n for (const step of steps) {\n const escrow = escrowOf(step.extra)\n const carrierOnly = escrow?.synthetic === TOOL_RESULTS_ONLY\n if (!carrierOnly) {\n messages.push(\n canonicalChatMessage({\n role: step.source === 'agent' ? 'assistant' : step.source,\n content: escrow?.content_null === true ? null : contentToString(step.message),\n reasoning_content: step.reasoning_content ?? asString(escrow?.reasoning_content),\n tool_calls:\n step.tool_calls !== undefined && step.tool_calls.length > 0\n ? step.tool_calls.map(fromHarborToolCall)\n : undefined,\n name: asString(escrow?.name),\n is_copied_context: step.is_copied_context === true,\n }),\n )\n }\n for (const result of step.observation?.results ?? []) {\n const resultEscrow = escrowOf(result.extra)\n const callId = result.source_call_id ?? asString(resultEscrow?.source_call_id)\n if (carrierOnly) {\n // Our own carrier: every result on it was a tool turn, and its call id\n // (if the transcript had one) is in escrow, not on the wire.\n messages.push(fromObservationResult(result, callId))\n continue\n }\n if (typeof callId === 'string') {\n messages.push(fromObservationResult(result, callId))\n continue\n }\n // ATIF allows a result from a non-tool-calling action. Our chat schema\n // requires `tool_call_id` on a tool turn, so inventing one would forge a\n // link; the text is preserved as a system observation instead.\n if (result.content !== undefined) {\n messages.push(\n canonicalChatMessage({ role: 'system', content: contentToString(result.content) }),\n )\n }\n }\n }\n return messages\n}\n\nfunction taskFrom(escrow: Record<string, unknown> | undefined, fallbackId: string): RolloutTask {\n const task = escrowSection(escrow, 'task')\n return {\n suite: asString(task?.suite) ?? 'harbor-atif-import',\n instance_id: asString(task?.instance_id) ?? fallbackId,\n // ALWAYS holdout — the escrowed claim is read for nothing here.\n //\n // `extra.tangle.task.split` is a namespaced key, not an authenticated one:\n // a hand-written or third-party document can set `split: 'search'` as\n // easily as our own exporter can, and honouring it made \"this file says so\"\n // sufficient to walk into a training export. Trainability is a decision\n // about a file, so it is made by an operator through\n // `relabelImportedSplit`, never by the file about itself. The claim is not\n // destroyed — it stays readable in the source document.\n split: 'holdout',\n seed: asNumberOrNull(task?.seed),\n rep: asIntegerOrUndefined(task?.rep) ?? 0,\n }\n}\n\nfunction policyFrom(\n escrow: Record<string, unknown> | undefined,\n agent: HarborAgent,\n): RolloutPolicy {\n const escrowed = escrowSection(escrow, 'policy')\n if (escrowed !== undefined) {\n // Our own export: restore verbatim, including the nulls ATIF forced us to\n // substitute placeholders for in `agent.name` / `agent.version`.\n return {\n harness: asString(escrowed.harness) ?? null,\n harness_version: asString(escrowed.harness_version) ?? null,\n model: asString(escrowed.model) ?? null,\n provider: asString(escrowed.provider) ?? null,\n profile_commit: asString(escrowed.profile_commit) ?? null,\n ...(escrowed.prompt_hash !== undefined\n ? { prompt_hash: asString(escrowed.prompt_hash) ?? null }\n : {}),\n ...(escrowed.config_hash !== undefined\n ? { config_hash: asString(escrowed.config_hash) ?? null }\n : {}),\n ...(escrowed.agent_profile_cell_id !== undefined\n ? { agent_profile_cell_id: asString(escrowed.agent_profile_cell_id) ?? null }\n : {}),\n sampling: isRecord(escrowed.sampling) ? escrowed.sampling : null,\n }\n }\n return {\n harness: agent.name,\n harness_version: agent.version,\n model: agent.model_name ?? null,\n provider: null,\n profile_commit: null,\n sampling: null,\n }\n}\n\nfunction artifactsFrom(escrow: Record<string, unknown> | undefined): RolloutArtifacts {\n const artifacts = escrowSection(escrow, 'artifacts')\n return {\n patch_path: asString(artifacts?.patch_path) ?? null,\n run_dir: asString(artifacts?.run_dir) ?? null,\n transcript_ref: asString(artifacts?.transcript_ref) ?? null,\n }\n}\n\nfunction costFrom(final: HarborFinalMetrics | undefined): RolloutCostBlock {\n const rawExtra = final?.extra\n const extra = isRecord(rawExtra) ? rawExtra : undefined\n const calls = asIntegerOrUndefined(extra?.llm_call_count)\n return {\n usd: asNumberOrNull(final?.total_cost_usd),\n tokens_in: asNumberOrNull(final?.total_prompt_tokens),\n tokens_out: asNumberOrNull(final?.total_completion_tokens),\n tokens_reasoning: asNumberOrNull(extra?.reasoning_tokens),\n cache_read: asNumberOrNull(final?.total_cached_tokens),\n cache_write: asNumberOrNull(extra?.cache_write_tokens),\n wall_s: asNumberOrNull(extra?.wall_s),\n ...(calls !== undefined ? { llm_call_count: calls } : {}),\n }\n}\n\n/**\n * Recovers span projections from ATIF's NATIVE per-step channel, for documents\n * with no `extra.tangle.spans` escrow — i.e. everything a foreign producer\n * writes. Without this the logprobs and token ids a Harbor-native trainer\n * records would be read, validated, and then dropped on the floor.\n *\n * Only steps that actually carry one of the four fields produce a span: an\n * agent step with no metrics means \"not captured\", and inventing an empty span\n * for it would claim the run had a shape we did not observe.\n */\nfunction spansFromSteps(steps: HarborStep[]): RolloutStep[] {\n const spans: RolloutStep[] = []\n for (const step of steps) {\n if (step.source !== 'agent') continue\n const calls = asIntegerOrUndefined(step.llm_call_count)\n const promptIds = asIntegerArray(step.metrics?.prompt_token_ids)\n const completionIds = asIntegerArray(step.metrics?.completion_token_ids)\n const logprobs = asNumberArray(step.metrics?.logprobs)\n if (\n calls === undefined &&\n promptIds === undefined &&\n completionIds === undefined &&\n logprobs === undefined\n ) {\n continue\n }\n spans.push({\n kind: 'llm',\n name: step.model_name ?? 'chat',\n ...(calls !== undefined ? { llm_call_count: calls } : {}),\n ...(promptIds !== undefined ? { prompt_token_ids: promptIds } : {}),\n ...(completionIds !== undefined ? { completion_token_ids: completionIds } : {}),\n ...(logprobs !== undefined ? { logprobs } : {}),\n })\n }\n return spans\n}\n\n/**\n * Composes `provenance.gap` as an ordered SET of reasons.\n *\n * Appending made the note accrete on every pass (\"…no verdict | …no verdict\"),\n * which both grows without bound and breaks byte-level idempotency for a ledger\n * hashed on its serialized lines.\n */\nfunction composeGap(escrowedGap: string | undefined, added: readonly string[]): string {\n const parts: string[] = []\n const seen = new Set<string>()\n for (const raw of [...(escrowedGap?.split(' | ') ?? []), ...added]) {\n const reason = raw.trim()\n if (reason.length === 0 || seen.has(reason)) continue\n seen.add(reason)\n parts.push(reason)\n }\n return parts.join(' | ')\n}\n\nexport interface FromHarborOptions {\n /** Injected clock for deterministic output when the source carries no capture time. */\n now?: () => Date\n}\n\nfunction nodeToLine(\n trajectory: HarborTrajectory,\n parentId: string | null,\n capturedAt: string,\n): RolloutLine {\n const escrow = escrowOf(trajectory.extra)\n const rolloutId =\n asString(escrow?.rollout_id) ?? trajectory.trajectory_id ?? trajectory.session_id\n if (rolloutId === undefined || rolloutId.length === 0) {\n throw new Error(\n 'fromHarborTrajectory: trajectory has no trajectory_id or session_id — cannot mint a joinable rollout_id without inventing one',\n )\n }\n const outcome = escrowSection(escrow, 'outcome')\n const provenance = escrowSection(escrow, 'provenance')\n const capture = provenance?.capture\n const escrowedGap = asString(provenance?.gap)\n const role = escrow?.role\n const escrowedSpans = escrow?.spans\n // The escrow wins when it exists (our own export: exact, including an empty\n // array, which claims \"captured, none\" rather than \"not captured\"), and the\n // native per-step channel is the fallback (a foreign export: still signal).\n const nativeSpans = spansFromSteps(trajectory.steps)\n const spans: RolloutStep[] | undefined = Array.isArray(escrowedSpans)\n ? (escrowedSpans as RolloutStep[])\n : nativeSpans.length > 0\n ? nativeSpans\n : undefined\n const messages = stepsToMessages(trajectory.steps)\n const line: RolloutLine = {\n schema: ROLLOUT_SCHEMA,\n rollout_id: rolloutId,\n // The tree edge wins for embedded children; for a root, the escrowed\n // pointer may reference a line outside this document and is preserved.\n parent_rollout_id: parentId ?? asString(escrow?.parent_rollout_id) ?? null,\n run_id: asString(escrow?.run_id) ?? trajectory.session_id ?? rolloutId,\n // Required keys on the wire (0.127.0): a foreign document that never\n // stated them imports as explicit `null`, not as an absent field.\n experiment_id: asString(escrow?.experiment_id) ?? null,\n candidate_id: asString(escrow?.candidate_id) ?? null,\n generation: asIntegerOrUndefined(escrow?.generation) ?? null,\n candidate_index: asIntegerOrUndefined(escrow?.candidate_index) ?? null,\n role: ROLLOUT_ROLES.includes(role as RolloutRole) ? (role as RolloutRole) : 'agent',\n task: taskFrom(escrow, rolloutId),\n policy: policyFrom(escrow, trajectory.agent),\n messages,\n tool_defs: trajectory.agent.tool_definitions ?? [],\n ...(spans !== undefined ? { steps: spans } : {}),\n outcome: {\n // ATIF carries no label. Null is the labeled gap; a 0 here would be a\n // fabricated failure and a 1 a fabricated success.\n reward: null,\n reward_source: null,\n verdict: null,\n metrics: isRecord(outcome?.metrics) ? outcome.metrics : {},\n is_completed: typeof outcome?.is_completed === 'boolean' ? outcome.is_completed : true,\n is_truncated:\n typeof outcome?.is_truncated === 'boolean'\n ? outcome.is_truncated\n : trajectory.continued_trajectory_ref !== undefined,\n error: asString(outcome?.error) ?? null,\n // The anti-Goodhart flag is restored when the source document stated\n // it; the wire schema requires the field, so a document that never did\n // imports as `false`. That is safe here because the reward is already\n // forced to `null` (not trainable) and `realness_screened` stays\n // absent = unknown rather than claiming a screen ran.\n realness_gated: outcome?.realness_gated === true,\n },\n cost: costFrom(trajectory.final_metrics),\n artifacts: artifactsFrom(escrow),\n provenance: {\n captured_at: asString(provenance?.captured_at) ?? capturedAt,\n capture: ROLLOUT_CAPTURES.includes(capture as RolloutCapture)\n ? (capture as RolloutCapture)\n : 'backfill',\n gap: composeGap(escrowedGap, [HARBOR_IMPORT_GAP]),\n // Restored for the same reason `realness_gated` is, and with the opposite\n // risk profile from the reward: this is the gated run's own measurement\n // bag, moved off `outcome` by `gateGamedOutcome` so no exporter reads it\n // as training input. Dropping it here would silently destroy the audit\n // trail that says WHY the run was flagged and what it claimed, on the one\n // population an auditor most wants to inspect. It cannot be fail-open —\n // nothing projects `provenance` into a training row.\n ...(isRecord(provenance?.gated_evidence)\n ? { gated_evidence: provenance.gated_evidence as GatedEvidence }\n : {}),\n },\n }\n assertRolloutLine(line, `rollout line imported from ATIF trajectory ${rolloutId}`)\n return line\n}\n\n/**\n * Flatten an ATIF trajectory tree back into `tangle.rollout.v1` lines, parent\n * first, each child carrying `parent_rollout_id`.\n *\n * Every line comes back UNLABELED: `reward`, `reward_source` and `verdict` are\n * null and `provenance.gap` says why. ATIF models no verdict, so scoring an\n * imported trajectory is a judge's job, not this function's. Every line lands\n * on `holdout` whatever the document claims — see `relabelImportedSplit`.\n */\nexport function fromHarborTrajectory(\n trajectory: HarborTrajectory,\n options: FromHarborOptions = {},\n): RolloutLine[] {\n const capturedAt = (options.now?.() ?? new Date()).toISOString()\n const lines: RolloutLine[] = []\n const walk = (node: HarborTrajectory, parentId: string | null, onPath: Set<string>): void => {\n const line = nodeToLine(node, parentId, capturedAt)\n if (onPath.has(line.rollout_id)) {\n throw new Error(`subagent_trajectories cycle through trajectory_id ${line.rollout_id}`)\n }\n onPath.add(line.rollout_id)\n lines.push(line)\n for (const child of node.subagent_trajectories ?? []) {\n walk(child, line.rollout_id, onPath)\n }\n onPath.delete(line.rollout_id)\n }\n walk(trajectory, null, new Set())\n return lines\n}\n\n/**\n * THE explicit door out of `holdout` for imported lines.\n *\n * Import forces `holdout` because a document's own claim about its split is not\n * evidence — anyone can write `extra.tangle.task.split`. Promoting a file to a\n * trainable split is an operator's decision about provenance they verified, so\n * it is a separate, greppable call: `grep relabelImportedSplit` enumerates\n * every place foreign data was declared trainable, which is exactly the audit\n * the trusted-escrow version made impossible.\n *\n * Returns plain `RolloutLine`s. They still have to pass `assertMinted` (and its\n * anti-Goodhart check) to reach an exporter — re-labeling a split is not\n * minting a reward.\n */\nexport function relabelImportedSplit(\n lines: readonly RolloutLine[],\n split: RolloutSplit,\n): RolloutLine[] {\n if (!ROLLOUT_SPLITS.includes(split)) {\n throw new Error(`relabelImportedSplit: unknown split ${String(split)}`)\n }\n return lines.map((line) => ({ ...line, task: { ...line.task, split } }))\n}\n","/**\n * Read-only backfill reader over the opencode sqlite store\n * (~/.local/share/opencode/opencode.db) → canonical chat-with-tools messages.\n *\n * Schema consumed (observed, 2026-07): `session` rows carry directory /\n * parent_id / agent / model / cost / tokens_*; `message` rows carry a JSON\n * `data` blob ({role, modelID, providerID, tokens, cost, finish}); `part`\n * rows carry the actual content ({type: text|reasoning|tool|step-start|\n * step-finish|snapshot…}). Tool parts hold {callID, state:{input, output,\n * status}} — both the call and its result, which we split into an assistant\n * tool_call plus a role:\"tool\" result message.\n *\n * The store is mutable and can be corrupt (a `.corrupt-bak` sibling ships\n * next to it in the wild), so `openOpencodeDb` returns null instead of\n * throwing — callers record a gap line, never crash the backfill.\n */\n\nimport { createRequire } from 'node:module'\nimport { homedir } from 'node:os'\nimport { join } from 'node:path'\nimport type { DatabaseSync } from 'node:sqlite'\nimport type { ChatMessage, ChatToolCall } from '../schema'\n\nexport const DEFAULT_OPENCODE_DB = join(homedir(), '.local', 'share', 'opencode', 'opencode.db')\n\nexport interface OpencodeSessionRow {\n id: string\n parentId: string | null\n directory: string\n agent: string | null\n /** Raw session.model JSON: {id, providerID, variant} where present. */\n model: { id?: string; providerID?: string } | null\n costUsd: number | null\n tokensInput: number | null\n tokensOutput: number | null\n tokensReasoning: number | null\n tokensCacheRead: number | null\n tokensCacheWrite: number | null\n timeCreated: number\n timeUpdated: number\n}\n\n// `node:sqlite` is loaded through CommonJS `require`, not `import()`. esbuild\n// (bundling) and Vite (tests) both rewrite a dynamic import and strip the\n// `node:` prefix under an es20xx target, turning this builtin into a bogus\n// \"sqlite\" package lookup; composing the specifier at runtime does not reliably\n// defeat that (it still resolved through Vite's transform in some workers, so\n// the failure moved around as test files were added). A require obtained from\n// `createRequire` is not an analyzable module reference in either tool, so\n// neither can rewrite it.\n/** Open the store read-only; null = unavailable/corrupt (caller records a gap). */\nexport async function openOpencodeDb(\n path: string = DEFAULT_OPENCODE_DB,\n): Promise<DatabaseSync | null> {\n try {\n // Keep Node-only module initialization inside the Node-only operation.\n // Root imports are shared with edge consumers that do not define import.meta.url.\n const nodeRequire = createRequire(import.meta.url)\n const { DatabaseSync } = nodeRequire('node:sqlite') as typeof import('node:sqlite')\n const db = new DatabaseSync(path, { readOnly: true })\n // Probe: a corrupt store can open() fine and fail on first page read.\n db.prepare('SELECT id FROM session LIMIT 1').get()\n return db\n } catch {\n return null\n }\n}\n\nconst isRecord = (v: unknown): v is Record<string, unknown> =>\n typeof v === 'object' && v !== null && !Array.isArray(v)\n\nfunction observedNumber(value: unknown): number | null {\n return typeof value === 'number' && Number.isFinite(value) && value >= 0 ? value : null\n}\n\nfunction parseSessionRow(row: Record<string, unknown>): OpencodeSessionRow {\n let model: OpencodeSessionRow['model'] = null\n if (typeof row.model === 'string' && row.model.length > 0) {\n try {\n const parsed: unknown = JSON.parse(row.model)\n if (isRecord(parsed)) model = parsed as { id?: string; providerID?: string }\n } catch {\n model = null\n }\n }\n return {\n id: String(row.id),\n parentId: row.parent_id === null || row.parent_id === undefined ? null : String(row.parent_id),\n directory: String(row.directory),\n agent: row.agent === null || row.agent === undefined ? null : String(row.agent),\n model,\n costUsd: observedNumber(row.cost),\n tokensInput: observedNumber(row.tokens_input),\n tokensOutput: observedNumber(row.tokens_output),\n tokensReasoning: observedNumber(row.tokens_reasoning),\n tokensCacheRead: observedNumber(row.tokens_cache_read),\n tokensCacheWrite: observedNumber(row.tokens_cache_write),\n timeCreated: Number(row.time_created ?? 0),\n timeUpdated: Number(row.time_updated ?? 0),\n }\n}\n\nconst SESSION_COLUMNS =\n 'id, parent_id, directory, agent, model, cost, tokens_input, tokens_output, tokens_reasoning, tokens_cache_read, tokens_cache_write, time_created, time_updated'\n\n/** Sessions whose cwd is `directory` (the worker-clone join key). */\nexport function findOpencodeSessionsByDirectory(\n db: DatabaseSync,\n directory: string,\n): OpencodeSessionRow[] {\n const rows = db\n .prepare(`SELECT ${SESSION_COLUMNS} FROM session WHERE directory = ? ORDER BY time_created`)\n .all(directory) as Array<Record<string, unknown>>\n return rows.map(parseSessionRow)\n}\n\nconst PROJECTED_PART_TYPES = new Set([\n 'text',\n 'reasoning',\n 'tool',\n 'step-start',\n 'step-finish',\n 'snapshot',\n 'patch',\n])\n\ninterface OpencodePart {\n type?: string\n text?: string\n tool?: string\n callID?: string\n state?: { status?: string; input?: unknown; output?: unknown }\n}\n\nfunction toolResultContent(output: unknown): string {\n if (typeof output === 'string') return output\n if (output === null || output === undefined) return ''\n return JSON.stringify(output)\n}\n\n/**\n * Convert one session's message+part rows into canonical messages.\n * An opencode assistant message row spans several model steps; each step's\n * parts (reasoning → text → tool …) become one assistant message followed by\n * the role:\"tool\" results of its calls, preserving order.\n */\nexport function readOpencodeSessionMessages(db: DatabaseSync, sessionId: string): ChatMessage[] {\n const messageRows = db\n .prepare('SELECT id, data FROM message WHERE session_id = ? ORDER BY time_created, id')\n .all(sessionId) as Array<{ id: string; data: string }>\n const partsStmt = db.prepare('SELECT id, data FROM part WHERE message_id = ? ORDER BY id')\n\n const messages: ChatMessage[] = []\n for (const messageRow of messageRows) {\n let data: Record<string, unknown>\n try {\n const parsed: unknown = JSON.parse(messageRow.data)\n if (!isRecord(parsed)) {\n throw new Error('message payload is not an object')\n }\n data = parsed\n } catch (error) {\n throw new Error(`OpenCode session ${sessionId} message ${messageRow.id}: malformed payload`, {\n cause: error,\n })\n }\n const parts: OpencodePart[] = []\n for (const row of partsStmt.all(messageRow.id) as Array<{ id: string; data: string }>) {\n let parsed: unknown\n try {\n parsed = JSON.parse(row.data)\n } catch (error) {\n throw new Error(`OpenCode session ${sessionId} part ${row.id}: malformed JSON`, {\n cause: error,\n })\n }\n if (!isRecord(parsed)) {\n throw new Error(`OpenCode session ${sessionId} part ${row.id}: payload is not an object`)\n }\n if (typeof parsed.type !== 'string' || !PROJECTED_PART_TYPES.has(parsed.type)) {\n throw new Error(\n `OpenCode session ${sessionId} part ${row.id}: unsupported part type ${String(parsed.type)}`,\n )\n }\n if (\n ((parsed.type === 'text' || parsed.type === 'reasoning') &&\n typeof parsed.text !== 'string') ||\n (parsed.type === 'tool' &&\n (typeof parsed.callID !== 'string' ||\n !parsed.callID ||\n !isRecord(parsed.state) ||\n parsed.state.status !== 'completed' ||\n parsed.state.output === undefined))\n ) {\n throw new Error(\n `OpenCode session ${sessionId} part ${row.id}: incomplete ${parsed.type} part`,\n )\n }\n parts.push(parsed as OpencodePart)\n }\n\n if (data.role === 'user') {\n const text = parts\n .filter((p) => p.type === 'text' && typeof p.text === 'string')\n .map((p) => p.text as string)\n .join('\\n')\n messages.push({ role: 'user', content: text })\n continue\n }\n if (data.role !== 'assistant') {\n throw new Error(\n `OpenCode session ${sessionId} message ${messageRow.id}: unsupported role ${String(data.role)}`,\n )\n }\n\n // Split the row into steps at step-start boundaries; parts before the\n // first step-start (none observed, but tolerated) form an implicit step.\n const steps: OpencodePart[][] = []\n let current: OpencodePart[] = []\n for (const part of parts) {\n if (part.type === 'step-start') {\n if (current.length > 0) steps.push(current)\n current = []\n continue\n }\n if (part.type === 'step-finish' || part.type === 'snapshot' || part.type === 'patch') continue\n current.push(part)\n }\n if (current.length > 0) steps.push(current)\n\n for (const step of steps) {\n const reasoning = step\n .filter((p) => p.type === 'reasoning' && typeof p.text === 'string' && p.text.length > 0)\n .map((p) => p.text as string)\n .join('\\n')\n const text = step\n .filter((p) => p.type === 'text' && typeof p.text === 'string')\n .map((p) => p.text as string)\n .join('\\n')\n const toolParts = step.filter((p) => p.type === 'tool' && typeof p.callID === 'string')\n const toolCalls: ChatToolCall[] = toolParts.map((p) => ({\n id: p.callID as string,\n type: 'function',\n function: {\n name: p.tool ?? 'unknown',\n arguments: JSON.stringify(p.state?.input ?? {}),\n },\n }))\n if (reasoning.length === 0 && text.length === 0 && toolCalls.length === 0) continue\n messages.push({\n role: 'assistant',\n content: text.length > 0 ? text : null,\n ...(reasoning.length > 0 ? { reasoning_content: reasoning } : {}),\n ...(toolCalls.length > 0 ? { tool_calls: toolCalls } : {}),\n })\n for (const p of toolParts) {\n messages.push({\n role: 'tool',\n tool_call_id: p.callID as string,\n name: p.tool ?? 'unknown',\n content: toolResultContent(p.state?.output),\n })\n }\n }\n }\n return messages\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiFA,MAAa,sBAAsB;;AAGnC,MAAa,oBAAoB;;AAGjC,MAAM,SAAS;;;;;;;;;;;;;;AAef,MAAM,oBAAoB;AA8G1B,MAAMA,cAAY,MAChB,OAAO,MAAM,YAAY,MAAM,QAAQ,CAAC,MAAM,QAAQ,CAAC;AAEzD,MAAM,YAAY,MAAoC,OAAO,MAAM,WAAW,IAAI,KAAA;AAElF,MAAM,kBAAkB,MACtB,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAEpD,MAAM,wBAAwB,MAC5B,OAAO,UAAU,CAAC,IAAK,IAAe,KAAA;AAExC,MAAM,iBAAiB,MACrB,MAAM,QAAQ,CAAC,KAAK,EAAE,OAAO,MAAM,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,CAAC,IACzE,IACD,KAAA;AAEN,MAAM,kBAAkB,MACtB,MAAM,QAAQ,CAAC,KAAK,EAAE,OAAO,MAAM,OAAO,UAAU,CAAC,CAAC,IAAK,IAAiB,KAAA;;;;;;;;;AAU9E,SAAS,qBAAqB,OAQd;CACd,MAAM,UAAuB;EAAE,MAAM,MAAM;EAAM,SAAS,MAAM;CAAQ;CACxE,IAAI,MAAM,sBAAsB,KAAA,GAAW,QAAQ,oBAAoB,MAAM;CAC7E,IAAI,MAAM,eAAe,KAAA,GAAW,QAAQ,aAAa,MAAM;CAC/D,IAAI,MAAM,iBAAiB,KAAA,GAAW,QAAQ,eAAe,MAAM;CACnE,IAAI,MAAM,SAAS,KAAA,GAAW,QAAQ,OAAO,MAAM;CACnD,IAAI,MAAM,sBAAsB,MAAM,QAAQ,oBAAoB;CAClE,OAAO;AACT;;AAGA,SAAS,SAAS,OAAiF;CACjG,IAAI,CAACA,WAAS,KAAK,GAAG,OAAO,KAAA;CAC7B,MAAM,QAAQ,MAAM;CACpB,OAAOA,WAAS,KAAK,IAAI,QAAQ,KAAA;AACnC;AAEA,SAAS,cACP,QACA,KACqC;CACrC,IAAI,WAAW,KAAA,GAAW,OAAO,KAAA;CACjC,MAAM,QAAQ,OAAO;CACrB,OAAOA,WAAS,KAAK,IAAI,QAAQ,KAAA;AACnC;AAMA,SAAS,iBAAiB,MAAoC;CAK5D,IAAI;CACJ,IAAI;EACF,MAAM,UAAmB,KAAK,MAAM,KAAK,SAAS,SAAS;EAC3D,IAAIA,WAAS,OAAO,GAAG,SAAS;CAClC,QAAQ;EACN,SAAS,KAAA;CACX;CACA,OAAO;EACL,cAAc,KAAK;EACnB,eAAe,KAAK,SAAS;EAC7B,WAAW,UAAU,CAAC;EACtB,OAAO,GAAG,SAAS,EAAE,eAAe,KAAK,SAAS,UAAU,EAAE;CAChE;AACF;;;;;;AAOA,SAAS,oBAAoB,SAAsB,QAA0C;CAC3F,MAAM,SAAkC,CAAC;CACzC,IAAI,UAAU,QAAQ,iBAAiB,KAAA,GAAW,OAAO,iBAAiB,QAAQ;CAClF,IAAI,QAAQ,YAAY,MAAM,OAAO,UAAU,QAAQ;CACvD,MAAM,SAAkC,CAAC;CACzC,IAAI,CAAC,UAAU,QAAQ,iBAAiB,KAAA,GAAW,OAAO,iBAAiB,QAAQ;CACnF,IAAI,QAAQ,SAAS,KAAA,GAAW,OAAO,OAAO,QAAQ;CAGtD,IAAI,QAAQ,YAAY,MAAM,OAAO,eAAe;CACpD,IAAI,QAAQ,sBAAsB,KAAA,GAAW,OAAO,oBAAoB,QAAQ;CAChF,IAAI,QAAQ,sBAAsB,MAAM,OAAO,oBAAoB;CACnE,IAAI,OAAO,KAAK,MAAM,CAAC,CAAC,SAAS,GAAG,OAAO,QAAQ,GAAG,SAAS,OAAO;CACtE,OAAO;AACT;;AAGA,SAAS,aAAa,MAAkB,QAAqC;CAC3E,IAAI,WAAW,KAAA,GAAW,OAAO;CACjC,OAAO,KAAK,YAAY,MAAM,SAAS,KAAK,iBAAiB,MAAM,MAAM;AAC3E;;;;;;;;;AAUA,SAAS,iBAAiB,MAAkB,MAAqC;CAC/E,IAAI,SAAS,KAAA,GAAW;CACxB,IAAI,KAAK,mBAAmB,KAAA,GAAW,KAAK,iBAAiB,KAAK;CAClE,MAAM,UAAyB,CAAC;CAChC,IAAI,KAAK,qBAAqB,KAAA,GAAW,QAAQ,mBAAmB,KAAK;CACzE,IAAI,KAAK,yBAAyB,KAAA,GAChC,QAAQ,uBAAuB,KAAK;CAEtC,IAAI,KAAK,aAAa,KAAA,GAAW,QAAQ,WAAW,KAAK;CACzD,IAAI,OAAO,KAAK,OAAO,CAAC,CAAC,SAAS,GAAG,KAAK,UAAU;AACtD;;;;;;;;;;;;;;;;;;AAmBA,SAAS,gBAAgB,MAAiC;CACxD,MAAM,QAAsB,CAAC;CAC7B,MAAM,YAAY,KAAK,SAAS,CAAC,EAAA,CAAG,QAAQ,SAAS,KAAK,SAAS,KAAK;CACxE,IAAI,WAAW;CACf,KAAK,MAAM,WAAW,KAAK,UAAU;EACnC,IAAI,QAAQ,SAAS,QAAQ;GAC3B,MAAM,WAAW,MAAM,MAAM,SAAS;GACtC,IAAI,aAAa,KAAA,KAAa,aAAa,UAAU,QAAQ,YAAY,GAAG;IAC1E,IAAI,SAAS,gBAAgB,KAAA,GAAW,SAAS,cAAc,EAAE,SAAS,CAAC,EAAE;IAC7E,SAAS,YAAY,QAAQ,KAAK,oBAAoB,SAAS,IAAI,CAAC;IACpE;GACF;GAIA,MAAM,UAAsB;IAC1B,SAAS,MAAM,SAAS;IACxB,QAAQ;IACR,SAAS;IACT,aAAa,EAAE,SAAS,CAAC,oBAAoB,SAAS,KAAK,CAAC,EAAE;IAC9D,OAAO,GAAG,SAAS,EAAE,WAAW,kBAAkB,EAAE;GACtD;GACA,IAAI,QAAQ,sBAAsB,MAAM,QAAQ,oBAAoB;GACpE,MAAM,KAAK,OAAO;GAClB;EACF;EACA,MAAM,OAAmB;GACvB,SAAS,MAAM,SAAS;GACxB,QAAQ,QAAQ,SAAS,cAAc,UAAU,QAAQ;GACzD,SAAS,QAAQ,WAAW;EAC9B;EACA,MAAM,SAAkC,CAAC;EACzC,IAAI,QAAQ,YAAY,MAAM,OAAO,eAAe;EACpD,IAAI,QAAQ,SAAS,KAAA,GAAW,OAAO,OAAO,QAAQ;EACtD,IAAI,QAAQ,SAAS,aAAa;GAChC,IAAI,KAAK,OAAO,UAAU,MAAM,KAAK,aAAa,KAAK,OAAO;GAC9D,IAAI,QAAQ,sBAAsB,KAAA,GAChC,KAAK,oBAAoB,QAAQ;GAEnC,IAAI,QAAQ,eAAe,KAAA,KAAa,QAAQ,WAAW,SAAS,GAClE,KAAK,aAAa,QAAQ,WAAW,IAAI,gBAAgB;GAE3D,iBAAiB,MAAM,SAAS,SAAS;GACzC,YAAY;EACd,OAAO,IAAI,QAAQ,sBAAsB,KAAA,GAEvC,OAAO,oBAAoB,QAAQ;EAIrC,IAAI,QAAQ,sBAAsB,MAAM,KAAK,oBAAoB;EACjE,IAAI,OAAO,KAAK,MAAM,CAAC,CAAC,SAAS,GAAG,KAAK,QAAQ,GAAG,SAAS,OAAO;EACpE,MAAM,KAAK,IAAI;CACjB;CACA,OAAO;AACT;AAEA,SAAS,eAAe,MAAmB,WAAuC;CAChF,MAAM,UAA8B,CAAC;CACrC,IAAI,KAAK,KAAK,cAAc,MAAM,QAAQ,sBAAsB,KAAK,KAAK;CAC1E,IAAI,KAAK,KAAK,eAAe,MAAM,QAAQ,0BAA0B,KAAK,KAAK;CAC/E,IAAI,KAAK,KAAK,eAAe,MAAM,QAAQ,sBAAsB,KAAK,KAAK;CAE3E,IAAI,KAAK,KAAK,QAAQ,MAAM,QAAQ,iBAAiB,KAAK,KAAK;CAC/D,QAAQ,cAAc;CAGtB,MAAM,QAAiC,CAAC;CACxC,IAAI,KAAK,KAAK,qBAAqB,MAAM,MAAM,mBAAmB,KAAK,KAAK;CAC5E,IAAI,KAAK,KAAK,gBAAgB,MAAM,MAAM,qBAAqB,KAAK,KAAK;CACzE,IAAI,KAAK,KAAK,WAAW,MAAM,MAAM,SAAS,KAAK,KAAK;CACxD,IAAI,KAAK,KAAK,mBAAmB,KAAA,KAAa,KAAK,KAAK,mBAAmB,MACzE,MAAM,iBAAiB,KAAK,KAAK;CAEnC,IAAI,OAAO,KAAK,KAAK,CAAC,CAAC,SAAS,GAAG,QAAQ,QAAQ;CACnD,OAAO;AACT;AAEA,SAAS,QAAQ,MAAuC;CACtD,MAAM,QAAkB,CAAC;CACzB,IAAI,KAAK,WAAW,QAAQ,KAAA,GAAW,MAAM,KAAK,KAAK,WAAW,GAAG;CACrE,IAAI,KAAK,QAAQ,mBAAmB,MAKlC,MAAM,KAAK,8EAA8E;CAE3F,OAAO,MAAM,SAAS,IAAI,MAAM,KAAK,KAAK,IAAI,KAAA;AAChD;AAEA,SAAS,iBAAiB,MAAqC;CAC7D,MAAM,QAAQ,gBAAgB,IAAI;CAClC,MAAM,aAA+B;EACnC,gBAAgB;EAKhB,YAAY,KAAK;EACjB,eAAe,KAAK;EACpB,OAAO;GAGL,MAAM,KAAK,OAAO,WAAW;GAC7B,SAAS,KAAK,OAAO,mBAAmB;GACxC,GAAI,KAAK,OAAO,UAAU,OAAO,EAAE,YAAY,KAAK,OAAO,MAAM,IAAI,CAAC;GACtE,GAAI,KAAK,UAAU,SAAS,IAAI,EAAE,kBAAkB,KAAK,UAAU,IAAI,CAAC;EAC1E;EACA;EACA,eAAe,eAAe,MAAM,MAAM,MAAM;EAChD,OAAO,GACJ,SAAS;GACR,QAAQ,KAAK;GACb,YAAY,KAAK;GACjB,mBAAmB,KAAK;GACxB,QAAQ,KAAK;GACb,GAAI,KAAK,kBAAkB,KAAA,IAAY,EAAE,eAAe,KAAK,cAAc,IAAI,CAAC;GAChF,GAAI,KAAK,iBAAiB,KAAA,IAAY,EAAE,cAAc,KAAK,aAAa,IAAI,CAAC;GAC7E,YAAY,KAAK;GACjB,iBAAiB,KAAK;GACtB,MAAM,KAAK;GACX,MAAM,KAAK;GACX,QAAQ,KAAK;GAGb,GAAI,KAAK,UAAU,KAAA,IAAY,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;GACxD,SAAS;IAGP,SAAS,KAAK,QAAQ;IACtB,cAAc,KAAK,QAAQ;IAC3B,cAAc,KAAK,QAAQ;IAC3B,OAAO,KAAK,QAAQ;IACpB,GAAI,KAAK,QAAQ,mBAAmB,KAAA,IAChC,EAAE,gBAAgB,KAAK,QAAQ,eAAe,IAC9C,CAAC;GACP;GACA,WAAW,KAAK;GAChB,YAAY,KAAK;EACnB,EACF;CACF;CACA,MAAM,QAAQ,QAAQ,IAAI;CAC1B,IAAI,UAAU,KAAA,GAAW,WAAW,QAAQ;CAC5C,OAAO;AACT;AAOA,SAAS,YAAY,OAA8B;CACjD,MAAM,uBAAO,IAAI,IAAyB;CAC1C,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,KAAK,IAAI,KAAK,UAAU,GAC1B,MAAM,IAAI,MAAM,kCAAkC,KAAK,YAAY;EAErE,KAAK,IAAI,KAAK,YAAY,IAAI;CAChC;CACA,MAAM,QAAuB,CAAC;CAC9B,MAAM,6BAAa,IAAI,IAA2B;CAClD,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,SAAS,KAAK;EAGpB,IAAI,WAAW,QAAQ,CAAC,KAAK,IAAI,MAAM,GAAG;GACxC,MAAM,KAAK,IAAI;GACf;EACF;EACA,MAAM,WAAW,WAAW,IAAI,MAAM;EACtC,IAAI,aAAa,KAAA,GAAW,WAAW,IAAI,QAAQ,CAAC,IAAI,CAAC;OACpD,SAAS,KAAK,IAAI;CACzB;CAGA,IAAI,MAAM,SAAS,KAAK,MAAM,WAAW,GACvC,MAAM,IAAI,MACR,0CAA0C,MAAM,OAAO,UAAU,MAAM,KAAK,MAAM,EAAE,UAAU,CAAC,CAAC,KAAK,IAAI,EAAE,EAC7G;CAEF,OAAO;EAAE;EAAO;CAAW;AAC7B;AAEA,SAAS,SAAS,MAAmB,QAAgB,QAAuC;CAC1F,IAAI,OAAO,IAAI,KAAK,UAAU,GAC5B,MAAM,IAAI,MAAM,8CAA8C,KAAK,YAAY;CAEjF,OAAO,IAAI,KAAK,UAAU;CAC1B,MAAM,OAAO,iBAAiB,IAAI;CAClC,MAAM,WAAW,OAAO,WAAW,IAAI,KAAK,UAAU;CACtD,IAAI,aAAa,KAAA,KAAa,SAAS,SAAS,GAC9C,KAAK,wBAAwB,SAAS,KAAK,UAAU,SAAS,OAAO,QAAQ,MAAM,CAAC;CAEtF,OAAO,OAAO,KAAK,UAAU;CAC7B,OAAO;AACT;;;;;;;;;;;;;;;;;AAkBA,SAAgB,mBAAmB,OAAwC;CACzE,MAAM,QAAQ,qBAAqB,KAAK;CACxC,IAAI,MAAM,WAAW,GAAG,MAAM,IAAI,MAAM,8BAA8B;CACtE,IAAI,MAAM,SAAS,GAAG;EACpB,MAAM,MAAM,MAAM,KAAK,MAAM,EAAE,iBAAiB,GAAG,CAAC,CAAC,KAAK,IAAI;EAC9D,MAAM,IAAI,MACR,uBAAuB,MAAM,OAAO,UAAU,IAAI,4DACpD;CACF;CACA,OAAO,MAAM;AACf;;AAGA,SAAgB,qBAAqB,OAA0C;CAC7E,MAAM,SAAS,YAAY,KAAK;CAChC,OAAO,OAAO,MAAM,KAAK,SAAS,SAAS,MAAM,wBAAQ,IAAI,IAAI,CAAC,CAAC;AACrE;AAMA,SAAS,gBAAgB,SAA+C;CACtE,IAAI,OAAO,YAAY,UAAU,OAAO;CACxC,OAAO,QACJ,KAAK,SACJ,KAAK,SAAS,UAGV,UAAU,KAAK,QAAQ,cAAc,UAAU,GAAG,KAAK,QAAQ,QAAQ,GAAG,GAAG,KAAK,IACjF,KAAK,QAAQ,EACpB,CAAC,CACA,KAAK,IAAI;AACd;AAEA,SAAS,mBAAmB,MAAoC;CAC9D,MAAM,MAAM,SAAS,SAAS,KAAK,KAAK,CAAC,EAAE,aAAa;CACxD,OAAO;EACL,IAAI,KAAK;EACT,MAAM;EACN,UAAU;GAAE,MAAM,KAAK;GAAe,WAAW,OAAO,KAAK,UAAU,KAAK,aAAa,CAAC,CAAC;EAAE;CAC/F;AACF;AAEA,SAAS,sBACP,QACA,QACa;CACb,MAAM,SAAS,SAAS,OAAO,KAAK;CACpC,OAAO,qBAAqB;EAC1B,MAAM;EACN,SACE,QAAQ,iBAAiB,OACrB,OACA,OAAO,YAAY,KAAA,IACjB,KACA,gBAAgB,OAAO,OAAO;EACtC,mBAAmB,SAAS,QAAQ,iBAAiB;EACrD,cAAc;EACd,MAAM,SAAS,QAAQ,IAAI;EAC3B,mBAAmB,QAAQ,sBAAsB;CACnD,CAAC;AACH;AAEA,SAAS,gBAAgB,OAAoC;CAC3D,MAAM,WAA0B,CAAC;CACjC,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,SAAS,SAAS,KAAK,KAAK;EAClC,MAAM,cAAc,QAAQ,cAAc;EAC1C,IAAI,CAAC,aACH,SAAS,KACP,qBAAqB;GACnB,MAAM,KAAK,WAAW,UAAU,cAAc,KAAK;GACnD,SAAS,QAAQ,iBAAiB,OAAO,OAAO,gBAAgB,KAAK,OAAO;GAC5E,mBAAmB,KAAK,qBAAqB,SAAS,QAAQ,iBAAiB;GAC/E,YACE,KAAK,eAAe,KAAA,KAAa,KAAK,WAAW,SAAS,IACtD,KAAK,WAAW,IAAI,kBAAkB,IACtC,KAAA;GACN,MAAM,SAAS,QAAQ,IAAI;GAC3B,mBAAmB,KAAK,sBAAsB;EAChD,CAAC,CACH;EAEF,KAAK,MAAM,UAAU,KAAK,aAAa,WAAW,CAAC,GAAG;GACpD,MAAM,eAAe,SAAS,OAAO,KAAK;GAC1C,MAAM,SAAS,OAAO,kBAAkB,SAAS,cAAc,cAAc;GAC7E,IAAI,aAAa;IAGf,SAAS,KAAK,sBAAsB,QAAQ,MAAM,CAAC;IACnD;GACF;GACA,IAAI,OAAO,WAAW,UAAU;IAC9B,SAAS,KAAK,sBAAsB,QAAQ,MAAM,CAAC;IACnD;GACF;GAIA,IAAI,OAAO,YAAY,KAAA,GACrB,SAAS,KACP,qBAAqB;IAAE,MAAM;IAAU,SAAS,gBAAgB,OAAO,OAAO;GAAE,CAAC,CACnF;EAEJ;CACF;CACA,OAAO;AACT;AAEA,SAAS,SAAS,QAA6C,YAAiC;CAC9F,MAAM,OAAO,cAAc,QAAQ,MAAM;CACzC,OAAO;EACL,OAAO,SAAS,MAAM,KAAK,KAAK;EAChC,aAAa,SAAS,MAAM,WAAW,KAAK;EAU5C,OAAO;EACP,MAAM,eAAe,MAAM,IAAI;EAC/B,KAAK,qBAAqB,MAAM,GAAG,KAAK;CAC1C;AACF;AAEA,SAAS,WACP,QACA,OACe;CACf,MAAM,WAAW,cAAc,QAAQ,QAAQ;CAC/C,IAAI,aAAa,KAAA,GAGf,OAAO;EACL,SAAS,SAAS,SAAS,OAAO,KAAK;EACvC,iBAAiB,SAAS,SAAS,eAAe,KAAK;EACvD,OAAO,SAAS,SAAS,KAAK,KAAK;EACnC,UAAU,SAAS,SAAS,QAAQ,KAAK;EACzC,gBAAgB,SAAS,SAAS,cAAc,KAAK;EACrD,GAAI,SAAS,gBAAgB,KAAA,IACzB,EAAE,aAAa,SAAS,SAAS,WAAW,KAAK,KAAK,IACtD,CAAC;EACL,GAAI,SAAS,gBAAgB,KAAA,IACzB,EAAE,aAAa,SAAS,SAAS,WAAW,KAAK,KAAK,IACtD,CAAC;EACL,GAAI,SAAS,0BAA0B,KAAA,IACnC,EAAE,uBAAuB,SAAS,SAAS,qBAAqB,KAAK,KAAK,IAC1E,CAAC;EACL,UAAUA,WAAS,SAAS,QAAQ,IAAI,SAAS,WAAW;CAC9D;CAEF,OAAO;EACL,SAAS,MAAM;EACf,iBAAiB,MAAM;EACvB,OAAO,MAAM,cAAc;EAC3B,UAAU;EACV,gBAAgB;EAChB,UAAU;CACZ;AACF;AAEA,SAAS,cAAc,QAA+D;CACpF,MAAM,YAAY,cAAc,QAAQ,WAAW;CACnD,OAAO;EACL,YAAY,SAAS,WAAW,UAAU,KAAK;EAC/C,SAAS,SAAS,WAAW,OAAO,KAAK;EACzC,gBAAgB,SAAS,WAAW,cAAc,KAAK;CACzD;AACF;AAEA,SAAS,SAAS,OAAyD;CACzE,MAAM,WAAW,OAAO;CACxB,MAAM,QAAQA,WAAS,QAAQ,IAAI,WAAW,KAAA;CAC9C,MAAM,QAAQ,qBAAqB,OAAO,cAAc;CACxD,OAAO;EACL,KAAK,eAAe,OAAO,cAAc;EACzC,WAAW,eAAe,OAAO,mBAAmB;EACpD,YAAY,eAAe,OAAO,uBAAuB;EACzD,kBAAkB,eAAe,OAAO,gBAAgB;EACxD,YAAY,eAAe,OAAO,mBAAmB;EACrD,aAAa,eAAe,OAAO,kBAAkB;EACrD,QAAQ,eAAe,OAAO,MAAM;EACpC,GAAI,UAAU,KAAA,IAAY,EAAE,gBAAgB,MAAM,IAAI,CAAC;CACzD;AACF;;;;;;;;;;;AAYA,SAAS,eAAe,OAAoC;CAC1D,MAAM,QAAuB,CAAC;CAC9B,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,KAAK,WAAW,SAAS;EAC7B,MAAM,QAAQ,qBAAqB,KAAK,cAAc;EACtD,MAAM,YAAY,eAAe,KAAK,SAAS,gBAAgB;EAC/D,MAAM,gBAAgB,eAAe,KAAK,SAAS,oBAAoB;EACvE,MAAM,WAAW,cAAc,KAAK,SAAS,QAAQ;EACrD,IACE,UAAU,KAAA,KACV,cAAc,KAAA,KACd,kBAAkB,KAAA,KAClB,aAAa,KAAA,GAEb;EAEF,MAAM,KAAK;GACT,MAAM;GACN,MAAM,KAAK,cAAc;GACzB,GAAI,UAAU,KAAA,IAAY,EAAE,gBAAgB,MAAM,IAAI,CAAC;GACvD,GAAI,cAAc,KAAA,IAAY,EAAE,kBAAkB,UAAU,IAAI,CAAC;GACjE,GAAI,kBAAkB,KAAA,IAAY,EAAE,sBAAsB,cAAc,IAAI,CAAC;GAC7E,GAAI,aAAa,KAAA,IAAY,EAAE,SAAS,IAAI,CAAC;EAC/C,CAAC;CACH;CACA,OAAO;AACT;;;;;;;;AASA,SAAS,WAAW,aAAiC,OAAkC;CACrF,MAAM,QAAkB,CAAC;CACzB,MAAM,uBAAO,IAAI,IAAY;CAC7B,KAAK,MAAM,OAAO,CAAC,GAAI,aAAa,MAAM,KAAK,KAAK,CAAC,GAAI,GAAG,KAAK,GAAG;EAClE,MAAM,SAAS,IAAI,KAAK;EACxB,IAAI,OAAO,WAAW,KAAK,KAAK,IAAI,MAAM,GAAG;EAC7C,KAAK,IAAI,MAAM;EACf,MAAM,KAAK,MAAM;CACnB;CACA,OAAO,MAAM,KAAK,KAAK;AACzB;AAOA,SAAS,WACP,YACA,UACA,YACa;CACb,MAAM,SAAS,SAAS,WAAW,KAAK;CACxC,MAAM,YACJ,SAAS,QAAQ,UAAU,KAAK,WAAW,iBAAiB,WAAW;CACzE,IAAI,cAAc,KAAA,KAAa,UAAU,WAAW,GAClD,MAAM,IAAI,MACR,+HACF;CAEF,MAAM,UAAU,cAAc,QAAQ,SAAS;CAC/C,MAAM,aAAa,cAAc,QAAQ,YAAY;CACrD,MAAM,UAAU,YAAY;CAC5B,MAAM,cAAc,SAAS,YAAY,GAAG;CAC5C,MAAM,OAAO,QAAQ;CACrB,MAAM,gBAAgB,QAAQ;CAI9B,MAAM,cAAc,eAAe,WAAW,KAAK;CACnD,MAAM,QAAmC,MAAM,QAAQ,aAAa,IAC/D,gBACD,YAAY,SAAS,IACnB,cACA,KAAA;CACN,MAAM,WAAW,gBAAgB,WAAW,KAAK;CACjD,MAAM,OAAoB;EACxB,QAAQ;EACR,YAAY;EAGZ,mBAAmB,YAAY,SAAS,QAAQ,iBAAiB,KAAK;EACtE,QAAQ,SAAS,QAAQ,MAAM,KAAK,WAAW,cAAc;EAG7D,eAAe,SAAS,QAAQ,aAAa,KAAK;EAClD,cAAc,SAAS,QAAQ,YAAY,KAAK;EAChD,YAAY,qBAAqB,QAAQ,UAAU,KAAK;EACxD,iBAAiB,qBAAqB,QAAQ,eAAe,KAAK;EAClE,MAAM,cAAc,SAAS,IAAmB,IAAK,OAAuB;EAC5E,MAAM,SAAS,QAAQ,SAAS;EAChC,QAAQ,WAAW,QAAQ,WAAW,KAAK;EAC3C;EACA,WAAW,WAAW,MAAM,oBAAoB,CAAC;EACjD,GAAI,UAAU,KAAA,IAAY,EAAE,OAAO,MAAM,IAAI,CAAC;EAC9C,SAAS;GAGP,QAAQ;GACR,eAAe;GACf,SAAS;GACT,SAASA,WAAS,SAAS,OAAO,IAAI,QAAQ,UAAU,CAAC;GACzD,cAAc,OAAO,SAAS,iBAAiB,YAAY,QAAQ,eAAe;GAClF,cACE,OAAO,SAAS,iBAAiB,YAC7B,QAAQ,eACR,WAAW,6BAA6B,KAAA;GAC9C,OAAO,SAAS,SAAS,KAAK,KAAK;GAMnC,gBAAgB,SAAS,mBAAmB;EAC9C;EACA,MAAM,SAAS,WAAW,aAAa;EACvC,WAAW,cAAc,MAAM;EAC/B,YAAY;GACV,aAAa,SAAS,YAAY,WAAW,KAAK;GAClD,SAAS,iBAAiB,SAAS,OAAyB,IACvD,UACD;GACJ,KAAK,WAAW,aAAa,CAAC,iBAAiB,CAAC;GAQhD,GAAIA,WAAS,YAAY,cAAc,IACnC,EAAE,gBAAgB,WAAW,eAAgC,IAC7D,CAAC;EACP;CACF;CACA,kBAAkB,MAAM,8CAA8C,WAAW;CACjF,OAAO;AACT;;;;;;;;;;AAWA,SAAgB,qBACd,YACA,UAA6B,CAAC,GACf;CACf,MAAM,cAAc,QAAQ,MAAM,qBAAK,IAAI,KAAK,EAAA,CAAG,YAAY;CAC/D,MAAM,QAAuB,CAAC;CAC9B,MAAM,QAAQ,MAAwB,UAAyB,WAA8B;EAC3F,MAAM,OAAO,WAAW,MAAM,UAAU,UAAU;EAClD,IAAI,OAAO,IAAI,KAAK,UAAU,GAC5B,MAAM,IAAI,MAAM,qDAAqD,KAAK,YAAY;EAExF,OAAO,IAAI,KAAK,UAAU;EAC1B,MAAM,KAAK,IAAI;EACf,KAAK,MAAM,SAAS,KAAK,yBAAyB,CAAC,GACjD,KAAK,OAAO,KAAK,YAAY,MAAM;EAErC,OAAO,OAAO,KAAK,UAAU;CAC/B;CACA,KAAK,YAAY,sBAAM,IAAI,IAAI,CAAC;CAChC,OAAO;AACT;;;;;;;;;;;;;;;AAgBA,SAAgB,qBACd,OACA,OACe;CACf,IAAI,CAAC,eAAe,SAAS,KAAK,GAChC,MAAM,IAAI,MAAM,uCAAuC,OAAO,KAAK,GAAG;CAExE,OAAO,MAAM,KAAK,UAAU;EAAE,GAAG;EAAM,MAAM;GAAE,GAAG,KAAK;GAAM;EAAM;CAAE,EAAE;AACzE;;;;;;;;;;;;;;;;;;;ACv7BA,MAAa,sBAAsB,KAAK,QAAQ,GAAG,UAAU,SAAS,YAAY,aAAa;;AA4B/F,eAAsB,eACpB,OAAe,qBACe;CAC9B,IAAI;EAIF,MAAM,EAAE,iBADY,cAAc,YAAY,GACX,CAAC,CAAC,aAAa;EAClD,MAAM,KAAK,IAAI,aAAa,MAAM,EAAE,UAAU,KAAK,CAAC;EAEpD,GAAG,QAAQ,gCAAgC,CAAC,CAAC,IAAI;EACjD,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;AAEA,MAAM,YAAY,MAChB,OAAO,MAAM,YAAY,MAAM,QAAQ,CAAC,MAAM,QAAQ,CAAC;AAEzD,SAAS,eAAe,OAA+B;CACrD,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,KAAK,SAAS,IAAI,QAAQ;AACrF;AAEA,SAAS,gBAAgB,KAAkD;CACzE,IAAI,QAAqC;CACzC,IAAI,OAAO,IAAI,UAAU,YAAY,IAAI,MAAM,SAAS,GACtD,IAAI;EACF,MAAM,SAAkB,KAAK,MAAM,IAAI,KAAK;EAC5C,IAAI,SAAS,MAAM,GAAG,QAAQ;CAChC,QAAQ;EACN,QAAQ;CACV;CAEF,OAAO;EACL,IAAI,OAAO,IAAI,EAAE;EACjB,UAAU,IAAI,cAAc,QAAQ,IAAI,cAAc,KAAA,IAAY,OAAO,OAAO,IAAI,SAAS;EAC7F,WAAW,OAAO,IAAI,SAAS;EAC/B,OAAO,IAAI,UAAU,QAAQ,IAAI,UAAU,KAAA,IAAY,OAAO,OAAO,IAAI,KAAK;EAC9E;EACA,SAAS,eAAe,IAAI,IAAI;EAChC,aAAa,eAAe,IAAI,YAAY;EAC5C,cAAc,eAAe,IAAI,aAAa;EAC9C,iBAAiB,eAAe,IAAI,gBAAgB;EACpD,iBAAiB,eAAe,IAAI,iBAAiB;EACrD,kBAAkB,eAAe,IAAI,kBAAkB;EACvD,aAAa,OAAO,IAAI,gBAAgB,CAAC;EACzC,aAAa,OAAO,IAAI,gBAAgB,CAAC;CAC3C;AACF;AAEA,MAAM,kBACJ;;AAGF,SAAgB,gCACd,IACA,WACsB;CAItB,OAHa,GACV,QAAQ,UAAU,gBAAgB,wDAAwD,CAAC,CAC3F,IAAI,SACG,CAAC,CAAC,IAAI,eAAe;AACjC;AAEA,MAAM,uCAAuB,IAAI,IAAI;CACnC;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAUD,SAAS,kBAAkB,QAAyB;CAClD,IAAI,OAAO,WAAW,UAAU,OAAO;CACvC,IAAI,WAAW,QAAQ,WAAW,KAAA,GAAW,OAAO;CACpD,OAAO,KAAK,UAAU,MAAM;AAC9B;;;;;;;AAQA,SAAgB,4BAA4B,IAAkB,WAAkC;CAC9F,MAAM,cAAc,GACjB,QAAQ,6EAA6E,CAAC,CACtF,IAAI,SAAS;CAChB,MAAM,YAAY,GAAG,QAAQ,4DAA4D;CAEzF,MAAM,WAA0B,CAAC;CACjC,KAAK,MAAM,cAAc,aAAa;EACpC,IAAI;EACJ,IAAI;GACF,MAAM,SAAkB,KAAK,MAAM,WAAW,IAAI;GAClD,IAAI,CAAC,SAAS,MAAM,GAClB,MAAM,IAAI,MAAM,kCAAkC;GAEpD,OAAO;EACT,SAAS,OAAO;GACd,MAAM,IAAI,MAAM,oBAAoB,UAAU,WAAW,WAAW,GAAG,sBAAsB,EAC3F,OAAO,MACT,CAAC;EACH;EACA,MAAM,QAAwB,CAAC;EAC/B,KAAK,MAAM,OAAO,UAAU,IAAI,WAAW,EAAE,GAA0C;GACrF,IAAI;GACJ,IAAI;IACF,SAAS,KAAK,MAAM,IAAI,IAAI;GAC9B,SAAS,OAAO;IACd,MAAM,IAAI,MAAM,oBAAoB,UAAU,QAAQ,IAAI,GAAG,mBAAmB,EAC9E,OAAO,MACT,CAAC;GACH;GACA,IAAI,CAAC,SAAS,MAAM,GAClB,MAAM,IAAI,MAAM,oBAAoB,UAAU,QAAQ,IAAI,GAAG,2BAA2B;GAE1F,IAAI,OAAO,OAAO,SAAS,YAAY,CAAC,qBAAqB,IAAI,OAAO,IAAI,GAC1E,MAAM,IAAI,MACR,oBAAoB,UAAU,QAAQ,IAAI,GAAG,0BAA0B,OAAO,OAAO,IAAI,GAC3F;GAEF,KACI,OAAO,SAAS,UAAU,OAAO,SAAS,gBAC1C,OAAO,OAAO,SAAS,YACxB,OAAO,SAAS,WACd,OAAO,OAAO,WAAW,YACxB,CAAC,OAAO,UACR,CAAC,SAAS,OAAO,KAAK,KACtB,OAAO,MAAM,WAAW,eACxB,OAAO,MAAM,WAAW,KAAA,IAE5B,MAAM,IAAI,MACR,oBAAoB,UAAU,QAAQ,IAAI,GAAG,eAAe,OAAO,KAAK,MAC1E;GAEF,MAAM,KAAK,MAAsB;EACnC;EAEA,IAAI,KAAK,SAAS,QAAQ;GACxB,MAAM,OAAO,MACV,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,SAAS,QAAQ,CAAC,CAC9D,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,SAAS,KAAK;IAAE,MAAM;IAAQ,SAAS;GAAK,CAAC;GAC7C;EACF;EACA,IAAI,KAAK,SAAS,aAChB,MAAM,IAAI,MACR,oBAAoB,UAAU,WAAW,WAAW,GAAG,qBAAqB,OAAO,KAAK,IAAI,GAC9F;EAKF,MAAM,QAA0B,CAAC;EACjC,IAAI,UAA0B,CAAC;EAC/B,KAAK,MAAM,QAAQ,OAAO;GACxB,IAAI,KAAK,SAAS,cAAc;IAC9B,IAAI,QAAQ,SAAS,GAAG,MAAM,KAAK,OAAO;IAC1C,UAAU,CAAC;IACX;GACF;GACA,IAAI,KAAK,SAAS,iBAAiB,KAAK,SAAS,cAAc,KAAK,SAAS,SAAS;GACtF,QAAQ,KAAK,IAAI;EACnB;EACA,IAAI,QAAQ,SAAS,GAAG,MAAM,KAAK,OAAO;EAE1C,KAAK,MAAM,QAAQ,OAAO;GACxB,MAAM,YAAY,KACf,QAAQ,MAAM,EAAE,SAAS,eAAe,OAAO,EAAE,SAAS,YAAY,EAAE,KAAK,SAAS,CAAC,CAAC,CACxF,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,MAAM,OAAO,KACV,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,SAAS,QAAQ,CAAC,CAC9D,KAAK,MAAM,EAAE,IAAc,CAAC,CAC5B,KAAK,IAAI;GACZ,MAAM,YAAY,KAAK,QAAQ,MAAM,EAAE,SAAS,UAAU,OAAO,EAAE,WAAW,QAAQ;GACtF,MAAM,YAA4B,UAAU,KAAK,OAAO;IACtD,IAAI,EAAE;IACN,MAAM;IACN,UAAU;KACR,MAAM,EAAE,QAAQ;KAChB,WAAW,KAAK,UAAU,EAAE,OAAO,SAAS,CAAC,CAAC;IAChD;GACF,EAAE;GACF,IAAI,UAAU,WAAW,KAAK,KAAK,WAAW,KAAK,UAAU,WAAW,GAAG;GAC3E,SAAS,KAAK;IACZ,MAAM;IACN,SAAS,KAAK,SAAS,IAAI,OAAO;IAClC,GAAI,UAAU,SAAS,IAAI,EAAE,mBAAmB,UAAU,IAAI,CAAC;IAC/D,GAAI,UAAU,SAAS,IAAI,EAAE,YAAY,UAAU,IAAI,CAAC;GAC1D,CAAC;GACD,KAAK,MAAM,KAAK,WACd,SAAS,KAAK;IACZ,MAAM;IACN,cAAc,EAAE;IAChB,MAAM,EAAE,QAAQ;IAChB,SAAS,kBAAkB,EAAE,OAAO,MAAM;GAC5C,CAAC;EAEL;CACF;CACA,OAAO;AACT"}

package/dist/run-record-BRlM99rX.js.map DELETED Viewed

@@ -1 +0,0 @@

- {"version":3,"file":"run-record-BRlM99rX.js","names":[],"sources":["../src/run-record.ts"],"sourcesContent":["/**\n * Paper-grade RunRecord schema + runtime validator.\n *\n * Every run that participates in a promotion gate, paper table, or\n * researcher loop SHOULD be recorded as a `RunRecord`. The mandatory\n * fields are exactly those the paper \"Two Loops, Three Roles\" requires\n * for reproducibility: who/what/when/cost/seed/hash, plus the search vs\n * holdout split tag. A task score is optional because execution-only records\n * must preserve missing labels instead of converting errors into zero quality.\n *\n * This is intentionally NOT a replacement for the rich `Run` /\n * `ProposeReviewReport` / `ScenarioResult` types already in the\n * package. Those are runtime structures with full provenance. A\n * `RunRecord` is the analysis-time projection — the JSON-friendly\n * row you'd put in a parquet file or paste into a notebook.\n *\n * Validate at the boundary:\n *\n * const rec = validateRunRecord(rawJson) // throws on missing\n * const ok = isRunRecord(rawJson) // boolean check\n * const rec = parseRunRecordSafe(rawJson) // { ok, value | error }\n *\n * The validator runs in pure TS — zod is intentionally NOT a\n * dependency.\n */\n\nimport type { AgentProfileCell } from './agent-profile-cell'\nimport { validateAgentProfileCell } from './agent-profile-cell'\nimport type { CostProvenance } from './cost-ledger'\nimport { ValidationError } from './errors'\n// Value import of a leaf module that itself imports only this file's TYPES —\n// no runtime cycle. It keeps the raw split-score derivation spelled in exactly\n// one place (see `rollout/score-derivation-guard`).\nimport { observedScore } from './rollout/reward'\nimport { FAILURE_CLASSES, type FailureClass } from './trace/schema'\n\n/** Search/dev/holdout split tag. 'search' is the paper-grade alias for the\n * combined train+test pool that the optimizer is allowed to read. */\nexport type RunSplitTag = 'search' | 'dev' | 'holdout'\n\n/**\n * Explicit execution-lifecycle result for a run.\n *\n * This is separate from task quality (`outcome`) and failure classification.\n * Producers set it only from root-run or process evidence.\n */\nexport type RunTerminalOutcome = 'succeeded' | 'failed' | 'cancelled' | 'incomplete' | 'unknown'\n\n/** Explicit model value for a row that never produced a served model snapshot. */\nexport const UNKNOWN_MODEL = 'unknown'\n\nexport interface RunTokenUsage {\n input: number\n /** All generated tokens charged as output, including reasoning tokens. */\n output: number\n /** Present only when one or more paid calls did not report token usage.\n * In that case, every numeric field is a known subtotal, not a measured total. */\n tokensKnown?: false\n /** Reasoning-token subset of `output`, when the provider reports it. */\n reasoning?: number\n /** Prompt tokens served from a provider cache. */\n cached?: number\n /** Prompt tokens written into a provider cache. */\n cacheWrite?: number\n}\n\n/**\n * How a run's USD amount was obtained.\n *\n * `costUsd` is the run's total, or null when the total is unknown. A run whose\n * receipts prove only part of its spend is a `lower-bound`: its total is\n * unknown, so `costUsd` and `usd` stay null, and the proven part is\n * `knownLowerBoundUsd`. A reader of the total therefore never takes a floor\n * for a total. A floor of $0 proves nothing, so it is written `uncaptured`.\n *\n * Only a RunRecord needs this kind. Every other cost shape keeps a known\n * subtotal beside `CostProvenance`, where `uncaptured` already means \"at\n * least the subtotal\"; a RunRecord's nullable `costUsd` cannot hold one.\n */\nexport type RunCostProvenance =\n | CostProvenance\n | { kind: 'lower-bound'; usd: null; knownLowerBoundUsd: number }\n\n/**\n * The run's place in a search: one attempt at one cell. The search ledger\n * holds the lineage (the node's parents, depth and order); a RunRecord only\n * points at it.\n */\nexport interface RunSearchCoordinates {\n searchId: string\n nodeId: string\n /** The cell id hashes the search, node, task, split and repeat, so it is\n * unique across candidates. */\n cellId: string\n /** Attempt number within the cell, counted from 1. */\n attempt: number\n}\n\n/** Where the run's execution record lives. */\nexport interface RunTraceRef {\n /** Trace id shared by every span of the run. */\n traceId: string\n /** Root run of the execution tree the run spawned, when one exists. */\n execRunId?: string\n}\n\n/**\n * The run id of one attempt at a search cell. It is unique across candidates\n * and attempts, unlike a `${scenario}:${rep}` id, which every candidate that\n * ran the same scenario shares.\n */\nexport function searchCellRunId(search: Pick<RunSearchCoordinates, 'cellId' | 'attempt'>): string {\n return `${search.cellId}:${search.attempt}`\n}\n\n/** The run's proven spend: its total when known, its floor when only a floor is proven, else 0. */\nexport function runCostFloorUsd(record: Pick<RunRecord, 'costUsd' | 'costProvenance'>): number {\n if (record.costProvenance.kind === 'lower-bound') return record.costProvenance.knownLowerBoundUsd\n return record.costUsd ?? 0\n}\n\nexport interface RunJudgeMetadata {\n model: string\n promptVersion: string\n /** [0,1] confidence the judge declared. Constant judge confidence\n * across many runs is a fallback signal (see `canary.ts`). */\n confidence: number\n /** True if the judge degraded to a fallback path (rules-only,\n * prior-call cache, etc.). The canary uses this to alert. */\n fallback: boolean\n}\n\n/**\n * Per-judge / per-dimension breakdown for runs scored by an ensemble of\n * judges over a multi-dimensional rubric.\n *\n * The collapsed `outcome.searchScore` / `holdoutScore` carries the\n * composite the gate uses. The full breakdown belongs here so consumers\n * can answer \"which judge disagreed?\", \"which dimension dragged the\n * composite down?\", and \"did half the panel fail?\" without re-running.\n *\n * `perJudge[judgeId][dim]` is the canonical source; `perDimMean` and\n * `composite` are convenience projections — derivable but precomputed so\n * downstream IRR primitives (`interRaterReliability`,\n * `corpusInterRaterAgreement`) and reporters don't pay the same\n * aggregation twice.\n *\n * Fail-loud discipline: judges that errored out land in `failedJudges`\n * by id. A missing key in `perJudge` is ambiguous (silent zero vs not\n * run); the explicit list makes a partial-failure recorded as such.\n */\nexport interface JudgeScoresRecord {\n /** Per-judge per-dimension scores. `{ \"kimi-k2.6\": { helpfulness: 0.8, clarity: 0.7 }, ... }`. */\n perJudge: Record<string, Record<string, number>>\n /** Per-dim mean across judges. Convenience — derivable from `perJudge`. */\n perDimMean: Record<string, number>\n /** Composite mean across successful judges. Mirrors the task score only\n * when `failedJudges` is empty. */\n composite: number\n /** Judges that errored or returned an unparseable verdict. Recorded\n * by id (e.g. `['glm-5.1']`) so a partial-failure case is explicit,\n * not inferred from missing keys in `perJudge`. */\n failedJudges?: string[]\n /** Free-form notes the judges emitted (joined across judges or\n * first-judge only — consumer's choice). */\n notes?: string\n}\n\nexport interface RunOutcome {\n /** Score on the search/optimization split. Optional for holdout-only and\n * execution-only records. */\n searchScore?: number\n /** Score on the held-out split. Optional for search-only and execution-only\n * records. When both scores are absent, the run is explicitly unlabeled. */\n holdoutScore?: number\n /** Bag of any other metric the run produced — judge dimensions,\n * pass/fail counters, latency stats, etc. Numeric only — keeps\n * reporters honest. */\n raw: Record<string, number>\n /** Per-judge / per-dim breakdown. Consumers writing ensemble\n * judgements populate this; substrate primitives like\n * `interRaterReliability` and `corpusInterRaterAgreement` accept\n * these records as input. Optional — single-judge or scalar-only\n * runs leave it unset. */\n judgeScores?: JudgeScoresRecord\n /** Authenticity / realness verdict — did the run build the REAL thing on the\n * intended infra, or fake it (see `./authenticity`)? Optional: only domains\n * with an authenticity config populate it. Carried in the corpus so the\n * flywheel / off-policy learning can optimize for real completion, not gamed\n * pass-rate. `score` is 0-1; `gated` is the anti-Goodhart flag — a gated run\n * must not count as a real success regardless of `score`. */\n realness?: { score: number; gated: boolean; reason?: string }\n}\n\n/**\n * Mandatory paper-grade fields for a single evaluation run. Optional\n * fields are extension points; mandatory fields throw if missing.\n *\n * Hash discipline:\n * - `promptHash` is the sha256 of the EFFECTIVE prompt sent to the\n * model (after any steering bundle merge).\n * - `configHash` is the sha256 of the effective run config (model,\n * temperature, tools, judges, splits). The pair (promptHash,\n * configHash) uniquely identifies an experiment cell.\n *\n * Model snapshot discipline:\n * - successful rows MUST encode a snapshot version. Bare aliases like\n * `claude-sonnet-4` or `gpt-4o` are banned — they remap silently.\n * Use `claude-sonnet-4-6@2025-04-15` or `gpt-4o-2024-11-20`.\n * - a failed, cancelled, incomplete, or otherwise unknown row may use\n * `UNKNOWN_MODEL` when no served model was observed. This is an explicit\n * absence marker, not a fabricated snapshot.\n */\nexport interface RunRecord {\n /** UUID for the run. */\n runId: string\n /** Logical experiment grouping (a treatment vs a baseline within\n * the same sweep should share `experimentId`). */\n experimentId: string\n /** Stable identifier for the candidate (variant) being run. The\n * promotion gate compares two `candidateId`s on matched items. */\n candidateId: string\n /** RNG seed for the run. Always recorded — silent re-seeding is\n * the most common cause of non-reproducible numbers. */\n seed: number\n /** Model identifier WITH snapshot version. */\n model: string\n /** sha256 of the effective prompt (post-steering). */\n promptHash: string\n /** sha256 of the effective config. */\n configHash: string\n /** Git SHA the harness was run from. */\n commitSha: string\n /** End-to-end wall-clock duration in milliseconds. */\n wallMs: number\n /** Time spent queued before execution started, if known. */\n queueMs?: number\n /** Total USD cost, or null when the total is unknown. */\n costUsd: number | null\n /** Whether `costUsd` came from billing data or a price calculation, is only\n * proven down to a floor, or is unavailable. */\n costProvenance: RunCostProvenance\n /** Token usage breakdown. */\n tokenUsage: RunTokenUsage\n /** Root-run or process terminal result. Never inferred from a child span. */\n terminalOutcome: RunTerminalOutcome\n /** Root-run or process failure reason. Valid only for a failed, cancelled,\n * or incomplete terminal result; never populated from a child span. */\n terminalFailureReason?: string\n /** Judge-side metadata, if a judge was used. */\n judgeMetadata?: RunJudgeMetadata\n /** Per-split scores + raw bag. */\n outcome: RunOutcome\n /** Canonical task-failure class drawn from the shared\n * `FAILURE_CLASSES` taxonomy. Producers set it only from task-result\n * evidence. Execution errors belong in\n * `outcome.raw.execution_error_count`. */\n failureClass?: FailureClass\n /** Free-form task-failure detail scoped under a non-success\n * `failureClass`. It is invalid without that class. */\n failureMode?: string\n /** Which split this run was drawn from. */\n splitTag: RunSplitTag\n /**\n * Stable scenario identifier the run observed or was scored against.\n * Comparison primitives match this identity rather than input order.\n */\n scenarioId: string\n /**\n * Canonical identity for the agent profile cell that produced this row:\n * profile artifact hash plus optional harness/model/prompt/reporting\n * dimensions. Use `agentProfile.cellId` to group persona sweeps and\n * longitudinal reports by the complete source profile, not by a loose\n * candidate label or opaque config hash.\n */\n agentProfile?: AgentProfileCell\n /** The search cell this run settled. When present, `runId` is\n * `searchCellRunId(search)`. */\n search?: RunSearchCoordinates\n /** The run's trace and execution tree. Absent means the producer did not\n * record where they are. */\n traceRef?: RunTraceRef\n}\n\n/**\n * Canonical task-result classification.\n *\n * A producer may omit classification, record explicit success, or attach\n * domain-specific detail to a non-success class. Detail can never stand alone.\n * Execution errors belong in `outcome.raw.execution_error_count`.\n */\nexport type RunTaskFailure =\n | { failureClass?: undefined; failureMode?: undefined }\n | { failureClass: 'success'; failureMode?: undefined }\n | {\n failureClass: Exclude<FailureClass, 'success'>\n failureMode?: string\n }\n\n/**\n * Return task quality, preferring held-out evidence when both scores exist.\n *\n * RAW: no realness protection is applied. Built on `observedScore` rather\n * than repeating the split derivation, so only `rollout/reward.ts` reads the\n * raw fields. Anything that becomes training data must use `trainingScore` or\n * `trainingReward` instead.\n */\nexport function runTaskScore(record: RunRecord): number | undefined {\n const score = observedScore(record)\n return typeof score === 'number' && Number.isFinite(score) ? score : undefined\n}\n\n// ── Validation ───────────────────────────────────────────────────────\n\nconst MANDATORY_TOP_LEVEL = [\n 'runId',\n 'experimentId',\n 'candidateId',\n 'seed',\n 'model',\n 'promptHash',\n 'configHash',\n 'commitSha',\n 'wallMs',\n 'costUsd',\n 'costProvenance',\n 'tokenUsage',\n 'terminalOutcome',\n 'outcome',\n 'splitTag',\n 'scenarioId',\n] as const\n\nconst SPLIT_TAGS: ReadonlyArray<RunSplitTag> = ['search', 'dev', 'holdout']\nconst TERMINAL_OUTCOMES: ReadonlyArray<RunTerminalOutcome> = [\n 'succeeded',\n 'failed',\n 'cancelled',\n 'incomplete',\n 'unknown',\n]\n\nexport class RunRecordValidationError extends ValidationError {\n readonly path: string\n constructor(message: string, path = '') {\n super(path ? `${message} (at ${path})` : message)\n this.path = path\n }\n}\n\n/**\n * Strict validator. Throws `RunRecordValidationError` on the first\n * missing or wrongly-typed field. Returns the input cast to\n * `RunRecord` on success — the validator does not coerce.\n */\nexport function validateRunRecord(input: unknown): RunRecord {\n if (input === null || typeof input !== 'object') {\n throw new RunRecordValidationError('expected object')\n }\n const obj = input as Record<string, unknown>\n\n for (const key of MANDATORY_TOP_LEVEL) {\n if (!(key in obj)) {\n throw new RunRecordValidationError(`missing mandatory field \"${key}\"`)\n }\n }\n\n expectString(obj.runId, 'runId')\n expectString(obj.experimentId, 'experimentId')\n expectString(obj.candidateId, 'candidateId')\n expectFiniteNumber(obj.seed, 'seed')\n expectString(obj.model, 'model')\n expectString(obj.promptHash, 'promptHash')\n expectString(obj.configHash, 'configHash')\n expectString(obj.commitSha, 'commitSha')\n expectNonNegativeNumber(obj.wallMs, 'wallMs')\n if (obj.queueMs !== undefined) expectNonNegativeNumber(obj.queueMs, 'queueMs')\n validateCost(obj.costUsd, obj.costProvenance)\n\n // Snapshot discipline: successful rows require a served model snapshot.\n // Non-success rows may carry the explicit absence marker when execution\n // stopped before a model identity was observed.\n if (\n !modelHasSnapshot(obj.model as string) &&\n !(obj.model === UNKNOWN_MODEL && obj.terminalOutcome !== 'succeeded')\n ) {\n throw new RunRecordValidationError(\n `model \"${obj.model}\" lacks a snapshot version (use 'name@YYYY-MM-DD' or 'name-YYYYMMDD', or '${UNKNOWN_MODEL}' for a non-success row without a served model)`,\n 'model',\n )\n }\n\n // Token usage.\n const tu = obj.tokenUsage\n if (tu === null || typeof tu !== 'object') {\n throw new RunRecordValidationError('tokenUsage must be an object', 'tokenUsage')\n }\n const tuRec = tu as Record<string, unknown>\n expectNonNegativeNumber(tuRec.input, 'tokenUsage.input')\n expectNonNegativeNumber(tuRec.output, 'tokenUsage.output')\n if (tuRec.tokensKnown !== undefined && tuRec.tokensKnown !== false) {\n throw new RunRecordValidationError(\n 'tokensKnown must be false when present; omit it when token usage is complete',\n 'tokenUsage.tokensKnown',\n )\n }\n if (tuRec.reasoning !== undefined) {\n expectNonNegativeNumber(tuRec.reasoning, 'tokenUsage.reasoning')\n if ((tuRec.reasoning as number) > (tuRec.output as number)) {\n throw new RunRecordValidationError(\n 'reasoning tokens must be a subset of output tokens',\n 'tokenUsage.reasoning',\n )\n }\n }\n if (tuRec.cached !== undefined) expectNonNegativeNumber(tuRec.cached, 'tokenUsage.cached')\n if (tuRec.cacheWrite !== undefined) {\n expectNonNegativeNumber(tuRec.cacheWrite, 'tokenUsage.cacheWrite')\n }\n\n // Judge metadata, optional.\n if (obj.judgeMetadata !== undefined) {\n const jm = obj.judgeMetadata\n if (jm === null || typeof jm !== 'object') {\n throw new RunRecordValidationError('judgeMetadata must be an object', 'judgeMetadata')\n }\n const jmRec = jm as Record<string, unknown>\n expectString(jmRec.model, 'judgeMetadata.model')\n expectString(jmRec.promptVersion, 'judgeMetadata.promptVersion')\n expectFiniteNumber(jmRec.confidence, 'judgeMetadata.confidence')\n if (typeof jmRec.fallback !== 'boolean') {\n throw new RunRecordValidationError(\n 'judgeMetadata.fallback must be boolean',\n 'judgeMetadata.fallback',\n )\n }\n }\n\n // Outcome.\n const out = obj.outcome\n if (out === null || typeof out !== 'object') {\n throw new RunRecordValidationError('outcome must be an object', 'outcome')\n }\n const outRec = out as Record<string, unknown>\n if (outRec.searchScore !== undefined)\n expectFiniteNumber(outRec.searchScore, 'outcome.searchScore')\n if (outRec.holdoutScore !== undefined)\n expectFiniteNumber(outRec.holdoutScore, 'outcome.holdoutScore')\n const raw = outRec.raw\n if (raw === null || typeof raw !== 'object') {\n throw new RunRecordValidationError('outcome.raw must be an object', 'outcome.raw')\n }\n for (const [k, v] of Object.entries(raw as Record<string, unknown>)) {\n expectFiniteNumber(v, `outcome.raw.${k}`)\n }\n // Realness verdict, optional.\n if (outRec.realness !== undefined) {\n const r = outRec.realness\n if (r === null || typeof r !== 'object') {\n throw new RunRecordValidationError('outcome.realness must be an object', 'outcome.realness')\n }\n const rr = r as Record<string, unknown>\n expectFiniteNumber(rr.score, 'outcome.realness.score')\n if (typeof rr.gated !== 'boolean') {\n throw new RunRecordValidationError(\n 'outcome.realness.gated must be a boolean',\n 'outcome.realness.gated',\n )\n }\n }\n\n // Per-judge / per-dim breakdown, optional.\n if (outRec.judgeScores !== undefined) {\n validateJudgeScores(outRec.judgeScores, 'outcome.judgeScores')\n }\n\n // Failure mode optional.\n if (\n obj.failureClass !== undefined &&\n (typeof obj.failureClass !== 'string' ||\n !FAILURE_CLASSES.includes(obj.failureClass as FailureClass))\n ) {\n throw new RunRecordValidationError(\n `failureClass must be one of ${FAILURE_CLASSES.join(', ')}`,\n 'failureClass',\n )\n }\n if (obj.failureMode !== undefined) {\n expectString(obj.failureMode, 'failureMode')\n if (obj.failureClass === undefined || obj.failureClass === 'success') {\n throw new RunRecordValidationError(\n 'failureMode requires a non-success failureClass',\n 'failureMode',\n )\n }\n }\n\n if (\n typeof obj.terminalOutcome !== 'string' ||\n !TERMINAL_OUTCOMES.includes(obj.terminalOutcome as RunTerminalOutcome)\n ) {\n throw new RunRecordValidationError(\n `terminalOutcome must be one of ${TERMINAL_OUTCOMES.join(', ')}`,\n 'terminalOutcome',\n )\n }\n if (obj.terminalFailureReason !== undefined) {\n expectString(obj.terminalFailureReason, 'terminalFailureReason')\n if (\n obj.terminalOutcome !== 'failed' &&\n obj.terminalOutcome !== 'cancelled' &&\n obj.terminalOutcome !== 'incomplete'\n ) {\n throw new RunRecordValidationError(\n 'terminalFailureReason requires terminalOutcome failed, cancelled, or incomplete',\n 'terminalFailureReason',\n )\n }\n }\n\n if (obj.agentProfile !== undefined) {\n try {\n const profile = validateAgentProfileCell(obj.agentProfile)\n if (profile.model !== undefined && profile.model !== obj.model) {\n throw new RunRecordValidationError(\n `agentProfile.model \"${profile.model}\" does not match model \"${obj.model}\"`,\n 'agentProfile.model',\n )\n }\n if (profile.promptHash !== undefined && profile.promptHash !== obj.promptHash) {\n throw new RunRecordValidationError(\n `agentProfile.promptHash \"${profile.promptHash}\" does not match promptHash \"${obj.promptHash}\"`,\n 'agentProfile.promptHash',\n )\n }\n } catch (error) {\n if (error instanceof RunRecordValidationError) throw error\n if (error instanceof Error) {\n throw new RunRecordValidationError(error.message, 'agentProfile')\n }\n throw error\n }\n }\n\n expectString(obj.scenarioId, 'scenarioId')\n\n if (obj.search !== undefined) validateSearchCoordinates(obj.search, obj.runId as string)\n if (obj.traceRef !== undefined) validateTraceRef(obj.traceRef)\n\n // Split tag.\n if (typeof obj.splitTag !== 'string' || !SPLIT_TAGS.includes(obj.splitTag as RunSplitTag)) {\n throw new RunRecordValidationError(\n `splitTag must be one of ${SPLIT_TAGS.join(', ')}, got ${String(obj.splitTag)}`,\n 'splitTag',\n )\n }\n\n return input as RunRecord\n}\n\nfunction validateCost(costUsd: unknown, provenance: unknown): void {\n if (provenance === null || typeof provenance !== 'object') {\n throw new RunRecordValidationError('costProvenance must be an object', 'costProvenance')\n }\n const value = provenance as Record<string, unknown>\n if (\n value.kind !== 'observed' &&\n value.kind !== 'estimated' &&\n value.kind !== 'lower-bound' &&\n value.kind !== 'uncaptured'\n ) {\n throw new RunRecordValidationError(\n 'costProvenance.kind must be observed, estimated, lower-bound, or uncaptured',\n 'costProvenance.kind',\n )\n }\n if (value.kind === 'lower-bound') {\n if (value.usd !== null) {\n throw new RunRecordValidationError(\n 'lower-bound costProvenance.usd must be null; the floor is knownLowerBoundUsd',\n 'costProvenance.usd',\n )\n }\n if (costUsd !== null) {\n throw new RunRecordValidationError(\n 'lower-bound cost requires costUsd to be null; a floor is not a total',\n 'costUsd',\n )\n }\n expectFiniteNumber(value.knownLowerBoundUsd, 'costProvenance.knownLowerBoundUsd')\n if ((value.knownLowerBoundUsd as number) <= 0) {\n throw new RunRecordValidationError(\n \"a lower bound must be positive; write a $0 floor as { kind: 'uncaptured', usd: null }\",\n 'costProvenance.knownLowerBoundUsd',\n )\n }\n return\n }\n // A record must never read as a total it cannot support, so an uncaptured\n // cost carries no number at all. A matrix `CellResult` keeps its known\n // subtotal instead, because a cost ceiling must charge the part it can see;\n // converting one to the other drops that subtotal.\n if (value.kind === 'uncaptured') {\n if (value.usd !== null) {\n throw new RunRecordValidationError(\n 'uncaptured costProvenance.usd must be null',\n 'costProvenance.usd',\n )\n }\n if (costUsd !== null) {\n throw new RunRecordValidationError('uncaptured cost requires costUsd to be null', 'costUsd')\n }\n return\n }\n expectNonNegativeNumber(costUsd, 'costUsd')\n expectNonNegativeNumber(value.usd, 'costProvenance.usd')\n if (value.usd !== costUsd) {\n throw new RunRecordValidationError(\n 'costProvenance.usd must equal costUsd',\n 'costProvenance.usd',\n )\n }\n}\n\nfunction validateSearchCoordinates(value: unknown, runId: string): void {\n if (value === null || typeof value !== 'object') {\n throw new RunRecordValidationError('search must be an object', 'search')\n }\n const search = value as Record<string, unknown>\n expectString(search.searchId, 'search.searchId')\n expectString(search.nodeId, 'search.nodeId')\n expectString(search.cellId, 'search.cellId')\n if (!Number.isSafeInteger(search.attempt) || (search.attempt as number) < 1) {\n throw new RunRecordValidationError('expected an integer from 1', 'search.attempt')\n }\n const expected = searchCellRunId(search as unknown as RunSearchCoordinates)\n if (runId !== expected) {\n throw new RunRecordValidationError(\n `runId \"${runId}\" must be \"${expected}\", the search cell and attempt it settled`,\n 'runId',\n )\n }\n}\n\nfunction validateTraceRef(value: unknown): void {\n if (value === null || typeof value !== 'object') {\n throw new RunRecordValidationError('traceRef must be an object', 'traceRef')\n }\n const ref = value as Record<string, unknown>\n expectString(ref.traceId, 'traceRef.traceId')\n if (ref.execRunId !== undefined) expectString(ref.execRunId, 'traceRef.execRunId')\n}\n\n/** Boolean validator — convenience for filtering arrays. */\nexport function isRunRecord(input: unknown): input is RunRecord {\n try {\n validateRunRecord(input)\n return true\n } catch {\n return false\n }\n}\n\n/** Non-throwing validator — returns a discriminated union. */\nexport function parseRunRecordSafe(\n input: unknown,\n): { ok: true; value: RunRecord } | { ok: false; error: RunRecordValidationError } {\n try {\n return { ok: true, value: validateRunRecord(input) }\n } catch (e) {\n if (e instanceof RunRecordValidationError) return { ok: false, error: e }\n throw e\n }\n}\n\n/** Round-trip helper — `JSON.parse(JSON.stringify(record))` then validate. */\nexport function roundTripRunRecord(record: RunRecord): RunRecord {\n const json = JSON.stringify(record)\n return validateRunRecord(JSON.parse(json))\n}\n\n// ── Internals ────────────────────────────────────────────────────────\n\nfunction expectString(value: unknown, path: string): void {\n if (typeof value !== 'string' || value.length === 0) {\n throw new RunRecordValidationError(`expected non-empty string`, path)\n }\n}\n\nfunction expectFiniteNumber(value: unknown, path: string): void {\n if (typeof value !== 'number' || !Number.isFinite(value)) {\n throw new RunRecordValidationError(`expected finite number`, path)\n }\n}\n\nfunction expectNonNegativeNumber(value: unknown, path: string): void {\n expectFiniteNumber(value, path)\n if ((value as number) < 0) {\n throw new RunRecordValidationError('expected non-negative number', path)\n }\n}\n\nfunction validateJudgeScores(value: unknown, path: string): void {\n if (value === null || typeof value !== 'object') {\n throw new RunRecordValidationError('judgeScores must be an object', path)\n }\n const rec = value as Record<string, unknown>\n\n const perJudge = rec.perJudge\n if (perJudge === null || typeof perJudge !== 'object') {\n throw new RunRecordValidationError('perJudge must be an object', `${path}.perJudge`)\n }\n for (const [judgeId, dims] of Object.entries(perJudge as Record<string, unknown>)) {\n if (dims === null || typeof dims !== 'object') {\n throw new RunRecordValidationError(\n 'per-judge entry must be an object of dimension scores',\n `${path}.perJudge.${judgeId}`,\n )\n }\n for (const [dim, score] of Object.entries(dims as Record<string, unknown>)) {\n expectFiniteNumber(score, `${path}.perJudge.${judgeId}.${dim}`)\n }\n }\n\n const perDimMean = rec.perDimMean\n if (perDimMean === null || typeof perDimMean !== 'object') {\n throw new RunRecordValidationError('perDimMean must be an object', `${path}.perDimMean`)\n }\n for (const [dim, mean] of Object.entries(perDimMean as Record<string, unknown>)) {\n expectFiniteNumber(mean, `${path}.perDimMean.${dim}`)\n }\n\n expectFiniteNumber(rec.composite, `${path}.composite`)\n\n if (rec.failedJudges !== undefined) {\n if (!Array.isArray(rec.failedJudges)) {\n throw new RunRecordValidationError(\n 'failedJudges must be an array of strings',\n `${path}.failedJudges`,\n )\n }\n for (let i = 0; i < rec.failedJudges.length; i++) {\n const id = rec.failedJudges[i]\n if (typeof id !== 'string' || id.length === 0) {\n throw new RunRecordValidationError(\n 'failedJudges entry must be a non-empty string',\n `${path}.failedJudges[${i}]`,\n )\n }\n }\n }\n\n if (rec.notes !== undefined && typeof rec.notes !== 'string') {\n throw new RunRecordValidationError('notes must be a string', `${path}.notes`)\n }\n}\n\n/**\n * Snapshot check for provider model identifiers. Accepts ISO and compact\n * dates, Router's `-MMDD` snapshots, one opaque `@token`, and Vertex-style\n * `:date-token` suffixes. Routing selectors such as `@preset/name` are not\n * immutable model identities.\n */\nexport function modelHasSnapshot(model: string): boolean {\n if (model.length === 0 || model.trim() !== model) return false\n\n const opaqueAt = model.lastIndexOf('@')\n if (opaqueAt > 0) {\n const base = model.slice(0, opaqueAt)\n const token = model.slice(opaqueAt + 1)\n if (!base.includes('@') && /^[A-Za-z0-9](?:[A-Za-z0-9._-]*[A-Za-z0-9])?$/u.test(token)) {\n return true\n }\n }\n\n const isoDate = model.match(/-(\\d{4})-(\\d{2})-(\\d{2})$/u)\n if (isoDate && validSnapshotDate(isoDate[1]!, isoDate[2]!, isoDate[3]!)) return true\n\n const compactDate = model.match(/-(\\d{4})(\\d{2})(\\d{2})$/u)\n if (compactDate && validSnapshotDate(compactDate[1]!, compactDate[2]!, compactDate[3]!)) {\n return true\n }\n\n const routerDate = model.match(/-(\\d{2})(\\d{2})$/u)\n if (routerDate && validSnapshotDate(undefined, routerDate[1]!, routerDate[2]!)) return true\n\n return /:date-[A-Za-z0-9](?:[A-Za-z0-9._-]*[A-Za-z0-9])?$/u.test(model)\n}\n\nfunction validSnapshotDate(year: string | undefined, month: string, day: string): boolean {\n const monthNumber = Number(month)\n const dayNumber = Number(day)\n if (!Number.isInteger(monthNumber) || monthNumber < 1 || monthNumber > 12) return false\n\n const yearNumber = year === undefined ? undefined : Number(year)\n const leapYear =\n yearNumber === undefined ||\n (yearNumber % 4 === 0 && (yearNumber % 100 !== 0 || yearNumber % 400 === 0))\n const daysInMonth = [31, leapYear ? 29 : 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31]\n return Number.isInteger(dayNumber) && dayNumber >= 1 && dayNumber <= daysInMonth[monthNumber - 1]!\n}\n"],"mappings":";;;;;;AAiDA,MAAa,gBAAgB;;;;;;AA8D7B,SAAgB,gBAAgB,QAAkE;CAChG,OAAO,GAAG,OAAO,OAAO,GAAG,OAAO;AACpC;;AAGA,SAAgB,gBAAgB,QAA+D;CAC7F,IAAI,OAAO,eAAe,SAAS,eAAe,OAAO,OAAO,eAAe;CAC/E,OAAO,OAAO,WAAW;AAC3B;;;;;;;;;AA4LA,SAAgB,aAAa,QAAuC;CAClE,MAAM,QAAQ,cAAc,MAAM;CAClC,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,IAAI,QAAQ,KAAA;AACvE;AAIA,MAAM,sBAAsB;CAC1B;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;AAEA,MAAM,aAAyC;CAAC;CAAU;CAAO;AAAS;AAC1E,MAAM,oBAAuD;CAC3D;CACA;CACA;CACA;CACA;AACF;AAEA,IAAa,2BAAb,cAA8C,gBAAgB;CAC5D;CACA,YAAY,SAAiB,OAAO,IAAI;EACtC,MAAM,OAAO,GAAG,QAAQ,OAAO,KAAK,KAAK,OAAO;EAChD,KAAK,OAAO;CACd;AACF;;;;;;AAOA,SAAgB,kBAAkB,OAA2B;CAC3D,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,iBAAiB;CAEtD,MAAM,MAAM;CAEZ,KAAK,MAAM,OAAO,qBAChB,IAAI,EAAE,OAAO,MACX,MAAM,IAAI,yBAAyB,4BAA4B,IAAI,EAAE;CAIzE,aAAa,IAAI,OAAO,OAAO;CAC/B,aAAa,IAAI,cAAc,cAAc;CAC7C,aAAa,IAAI,aAAa,aAAa;CAC3C,mBAAmB,IAAI,MAAM,MAAM;CACnC,aAAa,IAAI,OAAO,OAAO;CAC/B,aAAa,IAAI,YAAY,YAAY;CACzC,aAAa,IAAI,YAAY,YAAY;CACzC,aAAa,IAAI,WAAW,WAAW;CACvC,wBAAwB,IAAI,QAAQ,QAAQ;CAC5C,IAAI,IAAI,YAAY,KAAA,GAAW,wBAAwB,IAAI,SAAS,SAAS;CAC7E,aAAa,IAAI,SAAS,IAAI,cAAc;CAK5C,IACE,CAAC,iBAAiB,IAAI,KAAe,KACrC,EAAE,IAAI,UAAA,aAA2B,IAAI,oBAAoB,cAEzD,MAAM,IAAI,yBACR,UAAU,IAAI,MAAM,4EAA4E,cAAc,kDAC9G,OACF;CAIF,MAAM,KAAK,IAAI;CACf,IAAI,OAAO,QAAQ,OAAO,OAAO,UAC/B,MAAM,IAAI,yBAAyB,gCAAgC,YAAY;CAEjF,MAAM,QAAQ;CACd,wBAAwB,MAAM,OAAO,kBAAkB;CACvD,wBAAwB,MAAM,QAAQ,mBAAmB;CACzD,IAAI,MAAM,gBAAgB,KAAA,KAAa,MAAM,gBAAgB,OAC3D,MAAM,IAAI,yBACR,gFACA,wBACF;CAEF,IAAI,MAAM,cAAc,KAAA,GAAW;EACjC,wBAAwB,MAAM,WAAW,sBAAsB;EAC/D,IAAK,MAAM,YAAwB,MAAM,QACvC,MAAM,IAAI,yBACR,sDACA,sBACF;CAEJ;CACA,IAAI,MAAM,WAAW,KAAA,GAAW,wBAAwB,MAAM,QAAQ,mBAAmB;CACzF,IAAI,MAAM,eAAe,KAAA,GACvB,wBAAwB,MAAM,YAAY,uBAAuB;CAInE,IAAI,IAAI,kBAAkB,KAAA,GAAW;EACnC,MAAM,KAAK,IAAI;EACf,IAAI,OAAO,QAAQ,OAAO,OAAO,UAC/B,MAAM,IAAI,yBAAyB,mCAAmC,eAAe;EAEvF,MAAM,QAAQ;EACd,aAAa,MAAM,OAAO,qBAAqB;EAC/C,aAAa,MAAM,eAAe,6BAA6B;EAC/D,mBAAmB,MAAM,YAAY,0BAA0B;EAC/D,IAAI,OAAO,MAAM,aAAa,WAC5B,MAAM,IAAI,yBACR,0CACA,wBACF;CAEJ;CAGA,MAAM,MAAM,IAAI;CAChB,IAAI,QAAQ,QAAQ,OAAO,QAAQ,UACjC,MAAM,IAAI,yBAAyB,6BAA6B,SAAS;CAE3E,MAAM,SAAS;CACf,IAAI,OAAO,gBAAgB,KAAA,GACzB,mBAAmB,OAAO,aAAa,qBAAqB;CAC9D,IAAI,OAAO,iBAAiB,KAAA,GAC1B,mBAAmB,OAAO,cAAc,sBAAsB;CAChE,MAAM,MAAM,OAAO;CACnB,IAAI,QAAQ,QAAQ,OAAO,QAAQ,UACjC,MAAM,IAAI,yBAAyB,iCAAiC,aAAa;CAEnF,KAAK,MAAM,CAAC,GAAG,MAAM,OAAO,QAAQ,GAA8B,GAChE,mBAAmB,GAAG,eAAe,GAAG;CAG1C,IAAI,OAAO,aAAa,KAAA,GAAW;EACjC,MAAM,IAAI,OAAO;EACjB,IAAI,MAAM,QAAQ,OAAO,MAAM,UAC7B,MAAM,IAAI,yBAAyB,sCAAsC,kBAAkB;EAE7F,MAAM,KAAK;EACX,mBAAmB,GAAG,OAAO,wBAAwB;EACrD,IAAI,OAAO,GAAG,UAAU,WACtB,MAAM,IAAI,yBACR,4CACA,wBACF;CAEJ;CAGA,IAAI,OAAO,gBAAgB,KAAA,GACzB,oBAAoB,OAAO,aAAa,qBAAqB;CAI/D,IACE,IAAI,iBAAiB,KAAA,MACpB,OAAO,IAAI,iBAAiB,YAC3B,CAAC,gBAAgB,SAAS,IAAI,YAA4B,IAE5D,MAAM,IAAI,yBACR,+BAA+B,gBAAgB,KAAK,IAAI,KACxD,cACF;CAEF,IAAI,IAAI,gBAAgB,KAAA,GAAW;EACjC,aAAa,IAAI,aAAa,aAAa;EAC3C,IAAI,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB,WACzD,MAAM,IAAI,yBACR,mDACA,aACF;CAEJ;CAEA,IACE,OAAO,IAAI,oBAAoB,YAC/B,CAAC,kBAAkB,SAAS,IAAI,eAAqC,GAErE,MAAM,IAAI,yBACR,kCAAkC,kBAAkB,KAAK,IAAI,KAC7D,iBACF;CAEF,IAAI,IAAI,0BAA0B,KAAA,GAAW;EAC3C,aAAa,IAAI,uBAAuB,uBAAuB;EAC/D,IACE,IAAI,oBAAoB,YACxB,IAAI,oBAAoB,eACxB,IAAI,oBAAoB,cAExB,MAAM,IAAI,yBACR,mFACA,uBACF;CAEJ;CAEA,IAAI,IAAI,iBAAiB,KAAA,GACvB,IAAI;EACF,MAAM,UAAU,yBAAyB,IAAI,YAAY;EACzD,IAAI,QAAQ,UAAU,KAAA,KAAa,QAAQ,UAAU,IAAI,OACvD,MAAM,IAAI,yBACR,uBAAuB,QAAQ,MAAM,0BAA0B,IAAI,MAAM,IACzE,oBACF;EAEF,IAAI,QAAQ,eAAe,KAAA,KAAa,QAAQ,eAAe,IAAI,YACjE,MAAM,IAAI,yBACR,4BAA4B,QAAQ,WAAW,+BAA+B,IAAI,WAAW,IAC7F,yBACF;CAEJ,SAAS,OAAO;EACd,IAAI,iBAAiB,0BAA0B,MAAM;EACrD,IAAI,iBAAiB,OACnB,MAAM,IAAI,yBAAyB,MAAM,SAAS,cAAc;EAElE,MAAM;CACR;CAGF,aAAa,IAAI,YAAY,YAAY;CAEzC,IAAI,IAAI,WAAW,KAAA,GAAW,0BAA0B,IAAI,QAAQ,IAAI,KAAe;CACvF,IAAI,IAAI,aAAa,KAAA,GAAW,iBAAiB,IAAI,QAAQ;CAG7D,IAAI,OAAO,IAAI,aAAa,YAAY,CAAC,WAAW,SAAS,IAAI,QAAuB,GACtF,MAAM,IAAI,yBACR,2BAA2B,WAAW,KAAK,IAAI,EAAE,QAAQ,OAAO,IAAI,QAAQ,KAC5E,UACF;CAGF,OAAO;AACT;AAEA,SAAS,aAAa,SAAkB,YAA2B;CACjE,IAAI,eAAe,QAAQ,OAAO,eAAe,UAC/C,MAAM,IAAI,yBAAyB,oCAAoC,gBAAgB;CAEzF,MAAM,QAAQ;CACd,IACE,MAAM,SAAS,cACf,MAAM,SAAS,eACf,MAAM,SAAS,iBACf,MAAM,SAAS,cAEf,MAAM,IAAI,yBACR,+EACA,qBACF;CAEF,IAAI,MAAM,SAAS,eAAe;EAChC,IAAI,MAAM,QAAQ,MAChB,MAAM,IAAI,yBACR,gFACA,oBACF;EAEF,IAAI,YAAY,MACd,MAAM,IAAI,yBACR,wEACA,SACF;EAEF,mBAAmB,MAAM,oBAAoB,mCAAmC;EAChF,IAAK,MAAM,sBAAiC,GAC1C,MAAM,IAAI,yBACR,yFACA,mCACF;EAEF;CACF;CAKA,IAAI,MAAM,SAAS,cAAc;EAC/B,IAAI,MAAM,QAAQ,MAChB,MAAM,IAAI,yBACR,8CACA,oBACF;EAEF,IAAI,YAAY,MACd,MAAM,IAAI,yBAAyB,+CAA+C,SAAS;EAE7F;CACF;CACA,wBAAwB,SAAS,SAAS;CAC1C,wBAAwB,MAAM,KAAK,oBAAoB;CACvD,IAAI,MAAM,QAAQ,SAChB,MAAM,IAAI,yBACR,yCACA,oBACF;AAEJ;AAEA,SAAS,0BAA0B,OAAgB,OAAqB;CACtE,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,4BAA4B,QAAQ;CAEzE,MAAM,SAAS;CACf,aAAa,OAAO,UAAU,iBAAiB;CAC/C,aAAa,OAAO,QAAQ,eAAe;CAC3C,aAAa,OAAO,QAAQ,eAAe;CAC3C,IAAI,CAAC,OAAO,cAAc,OAAO,OAAO,KAAM,OAAO,UAAqB,GACxE,MAAM,IAAI,yBAAyB,8BAA8B,gBAAgB;CAEnF,MAAM,WAAW,gBAAgB,MAAyC;CAC1E,IAAI,UAAU,UACZ,MAAM,IAAI,yBACR,UAAU,MAAM,aAAa,SAAS,4CACtC,OACF;AAEJ;AAEA,SAAS,iBAAiB,OAAsB;CAC9C,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,8BAA8B,UAAU;CAE7E,MAAM,MAAM;CACZ,aAAa,IAAI,SAAS,kBAAkB;CAC5C,IAAI,IAAI,cAAc,KAAA,GAAW,aAAa,IAAI,WAAW,oBAAoB;AACnF;;AAGA,SAAgB,YAAY,OAAoC;CAC9D,IAAI;EACF,kBAAkB,KAAK;EACvB,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;;AAGA,SAAgB,mBACd,OACiF;CACjF,IAAI;EACF,OAAO;GAAE,IAAI;GAAM,OAAO,kBAAkB,KAAK;EAAE;CACrD,SAAS,GAAG;EACV,IAAI,aAAa,0BAA0B,OAAO;GAAE,IAAI;GAAO,OAAO;EAAE;EACxE,MAAM;CACR;AACF;;AAGA,SAAgB,mBAAmB,QAA8B;CAC/D,MAAM,OAAO,KAAK,UAAU,MAAM;CAClC,OAAO,kBAAkB,KAAK,MAAM,IAAI,CAAC;AAC3C;AAIA,SAAS,aAAa,OAAgB,MAAoB;CACxD,IAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAChD,MAAM,IAAI,yBAAyB,6BAA6B,IAAI;AAExE;AAEA,SAAS,mBAAmB,OAAgB,MAAoB;CAC9D,IAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GACrD,MAAM,IAAI,yBAAyB,0BAA0B,IAAI;AAErE;AAEA,SAAS,wBAAwB,OAAgB,MAAoB;CACnE,mBAAmB,OAAO,IAAI;CAC9B,IAAK,QAAmB,GACtB,MAAM,IAAI,yBAAyB,gCAAgC,IAAI;AAE3E;AAEA,SAAS,oBAAoB,OAAgB,MAAoB;CAC/D,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,iCAAiC,IAAI;CAE1E,MAAM,MAAM;CAEZ,MAAM,WAAW,IAAI;CACrB,IAAI,aAAa,QAAQ,OAAO,aAAa,UAC3C,MAAM,IAAI,yBAAyB,8BAA8B,GAAG,KAAK,UAAU;CAErF,KAAK,MAAM,CAAC,SAAS,SAAS,OAAO,QAAQ,QAAmC,GAAG;EACjF,IAAI,SAAS,QAAQ,OAAO,SAAS,UACnC,MAAM,IAAI,yBACR,yDACA,GAAG,KAAK,YAAY,SACtB;EAEF,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,IAA+B,GACvE,mBAAmB,OAAO,GAAG,KAAK,YAAY,QAAQ,GAAG,KAAK;CAElE;CAEA,MAAM,aAAa,IAAI;CACvB,IAAI,eAAe,QAAQ,OAAO,eAAe,UAC/C,MAAM,IAAI,yBAAyB,gCAAgC,GAAG,KAAK,YAAY;CAEzF,KAAK,MAAM,CAAC,KAAK,SAAS,OAAO,QAAQ,UAAqC,GAC5E,mBAAmB,MAAM,GAAG,KAAK,cAAc,KAAK;CAGtD,mBAAmB,IAAI,WAAW,GAAG,KAAK,WAAW;CAErD,IAAI,IAAI,iBAAiB,KAAA,GAAW;EAClC,IAAI,CAAC,MAAM,QAAQ,IAAI,YAAY,GACjC,MAAM,IAAI,yBACR,4CACA,GAAG,KAAK,cACV;EAEF,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,aAAa,QAAQ,KAAK;GAChD,MAAM,KAAK,IAAI,aAAa;GAC5B,IAAI,OAAO,OAAO,YAAY,GAAG,WAAW,GAC1C,MAAM,IAAI,yBACR,iDACA,GAAG,KAAK,gBAAgB,EAAE,EAC5B;EAEJ;CACF;CAEA,IAAI,IAAI,UAAU,KAAA,KAAa,OAAO,IAAI,UAAU,UAClD,MAAM,IAAI,yBAAyB,0BAA0B,GAAG,KAAK,OAAO;AAEhF;;;;;;;AAQA,SAAgB,iBAAiB,OAAwB;CACvD,IAAI,MAAM,WAAW,KAAK,MAAM,KAAK,MAAM,OAAO,OAAO;CAEzD,MAAM,WAAW,MAAM,YAAY,GAAG;CACtC,IAAI,WAAW,GAAG;EAChB,MAAM,OAAO,MAAM,MAAM,GAAG,QAAQ;EACpC,MAAM,QAAQ,MAAM,MAAM,WAAW,CAAC;EACtC,IAAI,CAAC,KAAK,SAAS,GAAG,KAAK,gDAAgD,KAAK,KAAK,GACnF,OAAO;CAEX;CAEA,MAAM,UAAU,MAAM,MAAM,4BAA4B;CACxD,IAAI,WAAW,kBAAkB,QAAQ,IAAK,QAAQ,IAAK,QAAQ,EAAG,GAAG,OAAO;CAEhF,MAAM,cAAc,MAAM,MAAM,0BAA0B;CAC1D,IAAI,eAAe,kBAAkB,YAAY,IAAK,YAAY,IAAK,YAAY,EAAG,GACpF,OAAO;CAGT,MAAM,aAAa,MAAM,MAAM,mBAAmB;CAClD,IAAI,cAAc,kBAAkB,KAAA,GAAW,WAAW,IAAK,WAAW,EAAG,GAAG,OAAO;CAEvF,OAAO,qDAAqD,KAAK,KAAK;AACxE;AAEA,SAAS,kBAAkB,MAA0B,OAAe,KAAsB;CACxF,MAAM,cAAc,OAAO,KAAK;CAChC,MAAM,YAAY,OAAO,GAAG;CAC5B,IAAI,CAAC,OAAO,UAAU,WAAW,KAAK,cAAc,KAAK,cAAc,IAAI,OAAO;CAElF,MAAM,aAAa,SAAS,KAAA,IAAY,KAAA,IAAY,OAAO,IAAI;CAI/D,MAAM,cAAc;EAAC;EAFnB,eAAe,KAAA,KACd,aAAa,MAAM,MAAM,aAAa,QAAQ,KAAK,aAAa,QAAQ,KACvC,KAAK;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;EAAI;CAAE;CACnF,OAAO,OAAO,UAAU,SAAS,KAAK,aAAa,KAAK,aAAa,YAAY,cAAc;AACjG"}

package/dist/run-record-DF-fuu31.d.ts.map DELETED Viewed

@@ -1 +0,0 @@

- {"version":3,"file":"run-record-DF-fuu31.d.ts","names":[],"sources":["../src/run-record.ts"],"mappings":";;;;;;;KAsCY;;;;;;;KAQA;;cAGC;UAEI;EACf;;EAEA;;;EAGA;;EAEA;;EAEA;;EAEA;;;;;;;;;;;;;;;KAgBU,oBACR;EACE;EAAqB;EAAW;;;;;;;UAOrB;EACf;EACA;;;EAGA;;EAEA;;;UAIe;;EAEf;;EAEA;;;;;;;iBAQc,gBAAgB,QAAQ,KAAK;;iBAK7B,gBAAgB,QAAQ,KAAK;UAK5B;EACf;EACA;;;EAGA;;;EAGA;;;;;;;;;;;;;;;;;;;;;UAsBe;;EAEf,UAAU,eAAe;;EAEzB,YAAY;;;EAGZ;;;;EAIA;;;EAGA;;UAGe;;;EAGf;;;EAGA;;;;EAIA,KAAK;;;;;;EAML,cAAc;;;;;;;EAOd;IAAa;IAAe;IAAgB;;;;;;;;;;;;;;;;;;;;;;UAsB7B;;EAEf;;;EAGA;;;EAGA;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,gBAAgB;;EAEhB,YAAY;;EAEZ,iBAAiB;;;EAGjB;;EAEA,gBAAgB;;EAEhB,SAAS;;;;;EAKT,eAAe;;;EAGf;;EAEA,UAAU;;;;;EAKV;;;;;;;;EAQA,eAAe;;;EAGf,SAAS;;;EAGT,WAAW;;;;;;;;;KAUD;EACN;EAA0B;;EAC1B;EAAyB;;EAEzB,cAAc,QAAQ;EACtB;;;;;;;;;;iBAWU,aAAa,QAAQ;cAmCxB,iCAAiC;WACnC;EACT,YAAY,iBAAiB;;;;;;;iBAWf,kBAAkB,iBAAiB;;iBA2SnC,YAAY,iBAAiB,SAAS;;iBAUtC,mBACd;EACG;EAAU,OAAO;;EAAgB;EAAW,OAAO;;;iBAUxC,mBAAmB,QAAQ,YAAY;;;;;;;iBAuFvC,iBAAiB"}