@tangle-network/agent-eval 0.95.0 → 0.96.0

package/dist/meta-eval/sentinel.d.ts

@@ -1,169 +0,0 @@
-/**
- * Judge sentinel — eval trustworthiness as a continuously measured,
- * alarmed trend.
- *
- * Judges are models; models change underneath us; calibration decays.
- * This module composes three existing instruments into one loop:
- *
- *   snapshot — adapters turn real instrument outputs (`calibrateJudge` /
- *              `calibrateJudgeContinuous`, `corpusInterRaterAgreement` /
- *              `continuousAgreement`, a rerun sentinel golden set) into
- *              `SentinelSnapshot`s
- *   series   — `analyzeSeries` (src/series-convergence) classifies each
- *              judge×metric history as stabilized / drifting / noisy
- *   alarm    — `judgeSentinelReport` turns trends + thresholds into named
- *              alarms; `evalHealthStamp` is the tiny object a campaign
- *              attaches to its verdicts
- *
- * Alarm conditions:
- *   - convergence state `drifting-down` on any tracked metric
- *   - irr below `minIrr`
- *   - calibrationKappa / sentinelPassRate dropped vs the series baseline
- *     beyond `maxKappaDrop` / `maxSentinelDrop`
- *   - judge model changed with no post-change golden-grounded snapshot
- *     (the silent-upgrade trap: judges agreeing with each other after a
- *     model swap proves nothing — only re-measuring against gold does)
- *   - newest snapshot older than `staleAfterDays` relative to `asOf`
- *
- * Wire-in: run the snapshot adapters from a post-campaign hook or a
- * nightly job, append to a `SentinelStore`, then compute
- * `judgeSentinelReport(await store.history(), { asOf })` with `asOf` =
- * the campaign timestamp. Attach `evalHealthStamp(report)` to campaign
- * verdicts: an alarmed sentinel means every downstream verdict carries
- * an integrity warning until the judge is recalibrated.
- *
- * The module is clock-free — every timestamp (`at`, `asOf`) is
- * caller-supplied ISO-8601, so reports are reproducible.
- */
-import type { CalibrationResult, CandidateScore, ContinuousAgreement, ContinuousCalibrationResult, GoldenItem } from '../judge-calibration';
-import { type SeriesConvergenceOptions, type SeriesConvergenceResult } from '../series-convergence';
-import type { CorpusAgreementReport } from '../statistics';
-export declare const SENTINEL_METRIC_NAMES: readonly ["irr", "calibrationKappa", "sentinelPassRate"];
-export type SentinelMetricName = (typeof SENTINEL_METRIC_NAMES)[number];
-export interface SentinelMetrics {
-    /** Inter-rater reliability — ICC(2,1) from agreement instruments. */
-    irr?: number;
-    /** Weighted κ vs the human golden set (`calibrateJudge*`). */
-    calibrationKappa?: number;
-    /** Fraction of a fixed sentinel golden set the judge still scores within tolerance. */
-    sentinelPassRate?: number;
-}
-export interface SentinelSnapshot {
-    /** Caller-supplied ISO-8601 timestamp — the module never reads a clock. */
-    at: string;
-    judgeId: string;
-    /** Exact model identity behind the judge (pin the full version string). */
-    judgeModel: string;
-    /**
-     * Metrics measured at `at`. May be empty: an empty-metrics snapshot is a
-     * model-change marker — it records the new `judgeModel` immediately and
-     * the silent-upgrade alarm stays raised until a golden-grounded snapshot
-     * (calibrationKappa or sentinelPassRate) follows.
-     */
-    metrics: SentinelMetrics;
-}
-/** Identity + timestamp the caller supplies alongside an instrument output. */
-export interface SnapshotMeta {
-    at: string;
-    judgeId: string;
-    judgeModel: string;
-}
-/** Shape guard for snapshots crossing a boundary (store append, JSONL load, report input). */
-export declare function validateSentinelSnapshot(snapshot: SentinelSnapshot, source?: string): void;
-/**
- * From `calibrateJudge` / `calibrateJudgeContinuous` output. Prefers the
- * un-rounded continuous κ_w when the report carries one (integer κ
- * discards information for [0,1] judges).
- */
-export declare function snapshotFromCalibration(report: CalibrationResult | ContinuousCalibrationResult, meta: SnapshotMeta): SentinelSnapshot;
-/**
- * From `corpusInterRaterAgreement` (statistics.ts) or `continuousAgreement`
- * (judge-calibration.ts) output. IRR = ICC(2,1) — the reliability
- * coefficient both instruments compute. For ensemble-level agreement,
- * `meta.judgeId` names the ensemble and `meta.judgeModel` pins its
- * composition (e.g. a joined list of member model strings).
- */
-export declare function snapshotFromAgreement(report: CorpusAgreementReport | ContinuousAgreement, meta: SnapshotMeta): SentinelSnapshot;
-export interface SentinelSetOptions {
-    /** Max |judge − human| for an item to count as passed. Default 0.1 (tuned for [0,1] scores). */
-    tolerance?: number;
-}
-/**
- * From a rerun of a fixed sentinel golden set: the same `GoldenItem`s the
- * judge was originally calibrated on, re-scored periodically. Pass rate =
- * fraction of joined items where the judge stays within `tolerance` of the
- * human score. Items the judge didn't score are excluded from the
- * denominator; zero joined items is an error, not a 0% pass rate.
- */
-export declare function snapshotFromSentinelSet(scores: CandidateScore[], golden: GoldenItem[], meta: SnapshotMeta, options?: SentinelSetOptions): SentinelSnapshot;
-export interface SentinelStore {
-    append(snapshot: SentinelSnapshot): Promise<void>;
-    /** All snapshots (optionally for one judge), in append order. */
-    history(judgeId?: string): Promise<SentinelSnapshot[]>;
-}
-export declare function inMemorySentinelStore(initial?: SentinelSnapshot[]): SentinelStore;
-/**
- * JSONL store — one snapshot per line, appended atomically per call.
- * A corrupt or shape-invalid line is a loud error naming the file and
- * line number: a sentinel history that silently drops records would
- * defeat the drift detection it exists to provide.
- */
-export declare function fileSentinelStore(path: string): SentinelStore;
-export interface SentinelThresholds {
-    /** Floor for irr — below it the judge ensemble is unreliable regardless of trend. Default 0.6. */
-    minIrr?: number;
-    /** Max tolerated calibrationKappa drop vs the series baseline. Default 0.15. */
-    maxKappaDrop?: number;
-    /** Max tolerated sentinelPassRate drop vs the series baseline. Default 0.1. */
-    maxSentinelDrop?: number;
-    /** Newest snapshot older than this (days, vs asOf) = the sentinel itself went blind. Default 30. */
-    staleAfterDays?: number;
-}
-export interface JudgeSentinelOptions {
-    /** Caller-supplied ISO timestamp staleness is measured against. */
-    asOf: string;
-    thresholds?: SentinelThresholds;
-    /** Forwarded to `analyzeSeries` (window, stableCv, driftRun). */
-    convergence?: SeriesConvergenceOptions;
-}
-export interface SentinelTrend {
-    judgeId: string;
-    metric: SentinelMetricName;
-    /** Verbatim `analyzeSeries` state for this judge×metric history. */
-    state: SeriesConvergenceResult['state'];
-    /** Newest value in the series. */
-    current: number;
-    /** Oldest value in the series — the reference the drop thresholds compare against. */
-    baseline: number;
-    /** current − baseline (signed; negative = decayed). */
-    drift: number;
-    alarmed: boolean;
-    reason?: string;
-}
-export interface SentinelReport {
-    perJudge: SentinelTrend[];
-    alarms: string[];
-    healthy: boolean;
-    /**
-     * `judgeId:metric` pairs with too few snapshots for the convergence
-     * machine. Named, not counted — a blind spot you can't see is a blind
-     * spot you won't fix. Floor/drop checks still apply to thin series, so
-     * insufficient history alone never masks an alarm.
-     */
-    insufficientHistory: string[];
-}
-export declare function judgeSentinelReport(history: SentinelSnapshot[], opts: JudgeSentinelOptions): SentinelReport;
-export interface EvalHealthStamp {
-    healthy: boolean;
-    alarms: string[];
-}
-/**
- * The object a campaign attaches to its verdicts. Compute it from the
- * sentinel report in a post-campaign hook (or a nightly job feeding the
- * next day's campaigns) and store it alongside the verdict payload.
- * `healthy: false` means the judges that produced those verdicts have an
- * unresolved drift / decay / silent-upgrade / staleness alarm — treat the
- * verdicts as carrying an integrity warning until recalibration clears it.
- */
-export declare function evalHealthStamp(report: SentinelReport): EvalHealthStamp;
-//# sourceMappingURL=sentinel.d.ts.map

package/dist/meta-eval/sentinel.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"sentinel.d.ts","sourceRoot":"","sources":["../../src/meta-eval/sentinel.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoCG;AAGH,OAAO,KAAK,EACV,iBAAiB,EACjB,cAAc,EACd,mBAAmB,EACnB,2BAA2B,EAC3B,UAAU,EACX,MAAM,sBAAsB,CAAA;AAC7B,OAAO,EAEL,KAAK,wBAAwB,EAC7B,KAAK,uBAAuB,EAC7B,MAAM,uBAAuB,CAAA;AAC9B,OAAO,KAAK,EAAE,qBAAqB,EAAE,MAAM,eAAe,CAAA;AAE1D,eAAO,MAAM,qBAAqB,0DAA2D,CAAA;AAE7F,MAAM,MAAM,kBAAkB,GAAG,CAAC,OAAO,qBAAqB,CAAC,CAAC,MAAM,CAAC,CAAA;AAEvE,MAAM,WAAW,eAAe;IAC9B,qEAAqE;IACrE,GAAG,CAAC,EAAE,MAAM,CAAA;IACZ,8DAA8D;IAC9D,gBAAgB,CAAC,EAAE,MAAM,CAAA;IACzB,uFAAuF;IACvF,gBAAgB,CAAC,EAAE,MAAM,CAAA;CAC1B;AAED,MAAM,WAAW,gBAAgB;IAC/B,2EAA2E;IAC3E,EAAE,EAAE,MAAM,CAAA;IACV,OAAO,EAAE,MAAM,CAAA;IACf,2EAA2E;IAC3E,UAAU,EAAE,MAAM,CAAA;IAClB;;;;;OAKG;IACH,OAAO,EAAE,eAAe,CAAA;CACzB;AAED,+EAA+E;AAC/E,MAAM,WAAW,YAAY;IAC3B,EAAE,EAAE,MAAM,CAAA;IACV,OAAO,EAAE,MAAM,CAAA;IACf,UAAU,EAAE,MAAM,CAAA;CACnB;AAYD,8FAA8F;AAC9F,wBAAgB,wBAAwB,CAAC,QAAQ,EAAE,gBAAgB,EAAE,MAAM,SAAa,GAAG,IAAI,CAwB9F;AAID;;;;GAIG;AACH,wBAAgB,uBAAuB,CACrC,MAAM,EAAE,iBAAiB,GAAG,2BAA2B,EACvD,IAAI,EAAE,YAAY,GACjB,gBAAgB,CAUlB;AAED;;;;;;GAMG;AACH,wBAAgB,qBAAqB,CACnC,MAAM,EAAE,qBAAqB,GAAG,mBAAmB,EACnD,IAAI,EAAE,YAAY,GACjB,gBAAgB,CAUlB;AAED,MAAM,WAAW,kBAAkB;IACjC,gGAAgG;IAChG,SAAS,CAAC,EAAE,MAAM,CAAA;CACnB;AAED;;;;;;GAMG;AACH,wBAAgB,uBAAuB,CACrC,MAAM,EAAE,cAAc,EAAE,EACxB,MAAM,EAAE,UAAU,EAAE,EACpB,IAAI,EAAE,YAAY,EAClB,OAAO,GAAE,kBAAuB,GAC/B,gBAAgB,CAwClB;AAID,MAAM,WAAW,aAAa;IAC5B,MAAM,CAAC,QAAQ,EAAE,gBAAgB,GAAG,OAAO,CAAC,IAAI,CAAC,CAAA;IACjD,iEAAiE;IACjE,OAAO,CAAC,OAAO,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,gBAAgB,EAAE,CAAC,CAAA;CACvD;AAED,wBAAgB,qBAAqB,CAAC,OAAO,GAAE,gBAAgB,EAAO,GAAG,aAAa,CAarF;AAED;;;;;GAKG;AACH,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,MAAM,GAAG,aAAa,CAsC7D;AAID,MAAM,WAAW,kBAAkB;IACjC,kGAAkG;IAClG,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,gFAAgF;IAChF,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,+EAA+E;IAC/E,eAAe,CAAC,EAAE,MAAM,CAAA;IACxB,oGAAoG;IACpG,cAAc,CAAC,EAAE,MAAM,CAAA;CACxB;AAED,MAAM,WAAW,oBAAoB;IACnC,mEAAmE;IACnE,IAAI,EAAE,MAAM,CAAA;IACZ,UAAU,CAAC,EAAE,kBAAkB,CAAA;IAC/B,iEAAiE;IACjE,WAAW,CAAC,EAAE,wBAAwB,CAAA;CACvC;AAED,MAAM,WAAW,aAAa;IAC5B,OAAO,EAAE,MAAM,CAAA;IACf,MAAM,EAAE,kBAAkB,CAAA;IAC1B,oEAAoE;IACpE,KAAK,EAAE,uBAAuB,CAAC,OAAO,CAAC,CAAA;IACvC,kCAAkC;IAClC,OAAO,EAAE,MAAM,CAAA;IACf,sFAAsF;IACtF,QAAQ,EAAE,MAAM,CAAA;IAChB,uDAAuD;IACvD,KAAK,EAAE,MAAM,CAAA;IACb,OAAO,EAAE,OAAO,CAAA;IAChB,MAAM,CAAC,EAAE,MAAM,CAAA;CAChB;AAED,MAAM,WAAW,cAAc;IAC7B,QAAQ,EAAE,aAAa,EAAE,CAAA;IACzB,MAAM,EAAE,MAAM,EAAE,CAAA;IAChB,OAAO,EAAE,OAAO,CAAA;IAChB;;;;;OAKG;IACH,mBAAmB,EAAE,MAAM,EAAE,CAAA;CAC9B;AAID,wBAAgB,mBAAmB,CACjC,OAAO,EAAE,gBAAgB,EAAE,EAC3B,IAAI,EAAE,oBAAoB,GACzB,cAAc,CAyGhB;AAID,MAAM,WAAW,eAAe;IAC9B,OAAO,EAAE,OAAO,CAAA;IAChB,MAAM,EAAE,MAAM,EAAE,CAAA;CACjB;AAED;;;;;;;GAOG;AACH,wBAAgB,eAAe,CAAC,MAAM,EAAE,cAAc,GAAG,eAAe,CAEvE"}

package/dist/metrics.d.ts

@@ -1,63 +0,0 @@
-import type { ProductClient } from './client';
-import type { DriverState, TurnMetrics } from './types';
-interface TokenPrice {
-    input: number;
-    output: number;
-}
-/** Per-1K token pricing for exact model ids. */
-export declare const MODEL_PRICING: Record<string, TokenPrice>;
-/** Resolve pricing for a model id: exact table, then family fallback.
- *  Returns null when the id matches nothing (caller decides — never a
- *  silent-zero masquerading as a real $0 cost). */
-export declare function resolveModelPricing(model: string): TokenPrice | null;
-/** True when `model` has known pricing (exact or family). Lets cost-aware
- *  callers distinguish a real $0 from an unpriced model. */
-export declare function isModelPriced(model: string): boolean;
-/** Estimate token count from string length (chars / 4 approximation) */
-export declare function estimateTokens(text: string): number;
-/** Calculate cost in USD from token counts and model. Unknown models warn
- *  once (not a silent zero) and return 0 so callers that ignore pricing keep
- *  working; cost-sensitive callers should gate on {@link isModelPriced}. */
-export declare function estimateCost(inputTokens: number, outputTokens: number, model: string): number;
-/**
- * TokenCounter — accumulates token usage and cost across turns.
- */
-export declare class TokenCounter {
-    private totalInput;
-    private totalOutput;
-    private totalCost;
-    private model;
-    constructor(model?: string);
-    /** Record tokens for a turn, returns per-turn cost */
-    record(inputTokens: number, outputTokens: number): number;
-    /** Estimate and record from raw text */
-    recordFromText(inputText: string, outputText: string): {
-        inputTokens: number;
-        outputTokens: number;
-        cost: number;
-    };
-    getTotalInput(): number;
-    getTotalOutput(): number;
-    getTotalCost(): number;
-}
-/**
- * MetricsCollector — collects per-turn metrics from the product.
- *
- * After each turn, queries the product's APIs to measure state changes.
- */
-export declare class MetricsCollector {
-    private client;
-    private workspaceId;
-    private metrics;
-    constructor(client: ProductClient, workspaceId: string);
-    /** Collect metrics after a turn completes */
-    collect(turn: number, responseLatencyMs: number, responseChars: number, codeBlocksProduced: number, blocksExtracted: number, completionCriteriaMet: number, completionCriteriaTotal: number, qualityScore?: number, inputTokens?: number, outputTokens?: number, estimatedCostUsd?: number): Promise<TurnMetrics>;
-    /** Get current product state */
-    getState(): Promise<DriverState>;
-    /** Get all collected metrics */
-    getMetrics(): TurnMetrics[];
-    /** Get convergence curve (completion% over turns) */
-    getConvergenceCurve(): number[];
-}
-export {};
-//# sourceMappingURL=metrics.d.ts.map

package/dist/metrics.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"metrics.d.ts","sourceRoot":"","sources":["../src/metrics.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,UAAU,CAAA;AAC7C,OAAO,KAAK,EAAE,WAAW,EAAE,WAAW,EAAE,MAAM,SAAS,CAAA;AAEvD,UAAU,UAAU;IAClB,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,MAAM,CAAA;CACf;AAED,gDAAgD;AAChD,eAAO,MAAM,aAAa,EAAE,MAAM,CAAC,MAAM,EAAE,UAAU,CAOpD,CAAA;AAgCD;;mDAEmD;AACnD,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,MAAM,GAAG,UAAU,GAAG,IAAI,CAQpE;AAED;4DAC4D;AAC5D,wBAAgB,aAAa,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAEpD;AAID,wEAAwE;AACxE,wBAAgB,cAAc,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAEnD;AAED;;4EAE4E;AAC5E,wBAAgB,YAAY,CAAC,WAAW,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,MAAM,CAa7F;AAED;;GAEG;AACH,qBAAa,YAAY;IACvB,OAAO,CAAC,UAAU,CAAI;IACtB,OAAO,CAAC,WAAW,CAAI;IACvB,OAAO,CAAC,SAAS,CAAI;IACrB,OAAO,CAAC,KAAK,CAAQ;gBAET,KAAK,SAAW;IAI5B,sDAAsD;IACtD,MAAM,CAAC,WAAW,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,GAAG,MAAM;IAQzD,wCAAwC;IACxC,cAAc,CACZ,SAAS,EAAE,MAAM,EACjB,UAAU,EAAE,MAAM,GACjB;QAAE,WAAW,EAAE,MAAM,CAAC;QAAC,YAAY,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE;IAO9D,aAAa,IAAI,MAAM;IAGvB,cAAc,IAAI,MAAM;IAGxB,YAAY,IAAI,MAAM;CAGvB;AAED;;;;GAIG;AACH,qBAAa,gBAAgB;IAC3B,OAAO,CAAC,MAAM,CAAe;IAC7B,OAAO,CAAC,WAAW,CAAQ;IAC3B,OAAO,CAAC,OAAO,CAAoB;gBACvB,MAAM,EAAE,aAAa,EAAE,WAAW,EAAE,MAAM;IAKtD,6CAA6C;IACvC,OAAO,CACX,IAAI,EAAE,MAAM,EACZ,iBAAiB,EAAE,MAAM,EACzB,aAAa,EAAE,MAAM,EACrB,kBAAkB,EAAE,MAAM,EAC1B,eAAe,EAAE,MAAM,EACvB,qBAAqB,EAAE,MAAM,EAC7B,uBAAuB,EAAE,MAAM,EAC/B,YAAY,CAAC,EAAE,MAAM,EACrB,WAAW,SAAI,EACf,YAAY,SAAI,EAChB,gBAAgB,SAAI,GACnB,OAAO,CAAC,WAAW,CAAC;IA2BvB,gCAAgC;IAC1B,QAAQ,IAAI,OAAO,CAAC,WAAW,CAAC;IAsBtC,gCAAgC;IAChC,UAAU,IAAI,WAAW,EAAE;IAI3B,qDAAqD;IACrD,mBAAmB,IAAI,MAAM,EAAE;CAGhC"}

package/dist/model-seats.d.ts

@@ -1,71 +0,0 @@
-/**
- * ModelSeats — the program's model seating chart.
- *
- * One object names which model fills each role in an eval program: the worker
- * under evaluation, the judge panel, the analyst, the reflection/driver model,
- * and the verifier. Re-tiering an entire program (economy ↔ frontier) is one
- * swapped object instead of a hunt through call sites.
- *
- * Wiring points — consumers thread seats; this module implements none of them
- * (those files belong to other surfaces):
- *  - `judges`     → `ensembleJudge({ models: seats.judges, … })` (src/judge-panel.ts)
- *                   and the `JudgeConfig`s handed to `makeEvalTools({ judges })`
- *                   (src/eval-tools.ts).
- *  - `reflection` → `selfImprove({ llm: { model: seats.reflection } })` — the
- *                   `gepaProposer` reflection model (src/contract/self-improve.ts);
- *                   same seat for any custom `SurfaceProposer`'s LLM.
- *  - `worker`     → the dispatch model the agent itself calls — the model an
- *                   `AgentProfile` declares.
- *  - `analyst`    → the LLM behind `analyzeRuns` / analyst-registry kinds.
- *  - `verifier`   → completion-verifier / objective-checker model.
- *  - campaign cells thread `judges` + driver models the same way; that wiring
- *    lands with the campaign surface, not here.
- *
- * `resolveSeat` is the only read path: an unset seat with no explicit fallback
- * throws — a model id is a budget decision, never a silent default.
- */
-import { ConfigError } from './errors';
-export interface ModelSeats {
-    /** The model under evaluation — what the agent itself dispatches with. */
-    worker?: string;
-    /** Judge-panel model ids — thread into `ensembleJudge({ models })`. */
-    judges?: string[];
-    /** Analyst model — `analyzeRuns` / analyst-registry LLM calls. */
-    analyst?: string;
-    /** Reflection/proposer model — `gepaProposer` mutation proposals. */
-    reflection?: string;
-    /** Verifier model — completion/objective checking. */
-    verifier?: string;
-}
-export type SeatName = keyof ModelSeats;
-export type SeatPresetName = keyof typeof seatPresets;
-/**
- * Tier presets — plain data, swap or spread freely.
- *
- * `economy` uses the fleet-policy ids: every id resolves through the
- * substrate's family pricing (no costUnknown axis) and the judge trio spans
- * three provider families (moonshot / deepseek / openai), so it passes
- * `assertCrossFamily` as-is.
- *
- * `frontier` is deliberately EMPTY: entitled frontier ids vary per router
- * account, and a hardcoded claude/gpt-5 id 401s on keys that lack it. Supply
- * your own: `{ ...seatPresets.frontier, worker: '<your-frontier-id>', … }` —
- * `resolveSeat` throws on every seat you haven't filled.
- */
-export declare const seatPresets: Record<'economy' | 'frontier', ModelSeats>;
-/** Thrown by `resolveSeat` when a seat is unset and no fallback was given. */
-export declare class SeatUnsetError extends ConfigError {
-    readonly seat: SeatName;
-    constructor(seat: SeatName);
-}
-/**
- * Read one seat. Blank strings and empty arrays count as unset (env-var
- * plumbing produces them); malformed values (non-string seat, non-array or
- * blank-entry `judges`) throw `ValidationError`. When the seat is unset, an
- * explicit `fallback` is returned (`[fallback]` for `judges` — a one-model
- * panel); without one, `SeatUnsetError`.
- */
-export declare function resolveSeat(seats: ModelSeats, seat: 'judges', fallback?: string): string[];
-export declare function resolveSeat(seats: ModelSeats, seat: Exclude<SeatName, 'judges'>, fallback?: string): string;
-export declare function resolveSeat(seats: ModelSeats, seat: SeatName, fallback?: string): string | string[];
-//# sourceMappingURL=model-seats.d.ts.map

package/dist/model-seats.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"model-seats.d.ts","sourceRoot":"","sources":["../src/model-seats.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAEH,OAAO,EAAE,WAAW,EAAmB,MAAM,UAAU,CAAA;AAEvD,MAAM,WAAW,UAAU;IACzB,0EAA0E;IAC1E,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,uEAAuE;IACvE,MAAM,CAAC,EAAE,MAAM,EAAE,CAAA;IACjB,kEAAkE;IAClE,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,qEAAqE;IACrE,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,sDAAsD;IACtD,QAAQ,CAAC,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,MAAM,QAAQ,GAAG,MAAM,UAAU,CAAA;AAEvC,MAAM,MAAM,cAAc,GAAG,MAAM,OAAO,WAAW,CAAA;AAErD;;;;;;;;;;;;GAYG;AACH,eAAO,MAAM,WAAW,EAAE,MAAM,CAAC,SAAS,GAAG,UAAU,EAAE,UAAU,CASlE,CAAA;AAED,8EAA8E;AAC9E,qBAAa,cAAe,SAAQ,WAAW;aACjB,IAAI,EAAE,QAAQ;gBAAd,IAAI,EAAE,QAAQ;CAM3C;AAED;;;;;;GAMG;AACH,wBAAgB,WAAW,CAAC,KAAK,EAAE,UAAU,EAAE,IAAI,EAAE,QAAQ,EAAE,QAAQ,CAAC,EAAE,MAAM,GAAG,MAAM,EAAE,CAAA;AAC3F,wBAAgB,WAAW,CACzB,KAAK,EAAE,UAAU,EACjB,IAAI,EAAE,OAAO,CAAC,QAAQ,EAAE,QAAQ,CAAC,EACjC,QAAQ,CAAC,EAAE,MAAM,GAChB,MAAM,CAAA;AACT,wBAAgB,WAAW,CAAC,KAAK,EAAE,UAAU,EAAE,IAAI,EAAE,QAAQ,EAAE,QAAQ,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,MAAM,EAAE,CAAA"}

package/dist/model-seats.test.d.ts

@@ -1,2 +0,0 @@
-export {};
-//# sourceMappingURL=model-seats.test.d.ts.map

package/dist/model-seats.test.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"model-seats.test.d.ts","sourceRoot":"","sources":["../src/model-seats.test.ts"],"names":[],"mappings":""}

package/dist/muffled-gate-scanner.d.ts

@@ -1,102 +0,0 @@
-/**
- * muffled-gate-scanner — test helper that greps consumer source for
- * gate + measurement anti-patterns and fails with file:line locations.
- *
- * Named pattern lives at starter-foundry's `.evolve/patterns/muffled-gate.md`;
- * same shape applies to every consumer (a gate that should fail loud
- * returns silent success; a metric that should emit a real number
- * reports noise/empty).
- *
- * Usage (in a consumer project's test file):
- *
- *   import { scanForMuffledGates, DEFAULT_FINDERS } from '@tangle-network/agent-eval'
- *
- *   test('no muffled gates in eval surface', () => {
- *     const findings = scanForMuffledGates({
- *       repoRoot: process.cwd(),
- *       scanFiles: ['src/eval/scaffold.ts', 'scripts/promote.mjs'],
- *       finders: DEFAULT_FINDERS,
- *     })
- *     if (findings.length) assert.fail(formatFindings(findings))
- *   })
- *
- * Customize by passing your own `finders` — each finder is
- * `(file, text) => Finding[]` and runs per-file.
- *
- * Escape hatch: any line containing `muffle-ok:` is excluded from all
- * finders, letting consumers opt a legitimate fallback out explicitly.
- */
-export interface MuffledFinding {
-    file: string;
-    line: number;
-    lineText: string;
-    pattern: string;
-}
-export type MuffledFinder = (file: string, text: string) => MuffledFinding[];
-export interface ScanOptions {
-    /** Absolute path to the repo root. */
-    repoRoot: string;
-    /** Explicit file list (paths relative to repoRoot) for context-specific finders. */
-    scanFiles: string[];
-    /**
-     * Auto-derived scan: walk these dirs for files matching importGlob + the
-     * string `importsContain` and run the universal finders on them. Pattern
-     * from starter-foundry H4 (research/decisions/001) — catches new files
-     * with agent-eval import that would otherwise escape context-specific
-     * scan lists.
-     */
-    autoDerive?: {
-        roots: string[];
-        extensions: RegExp;
-        importsContain: string;
-        universalFinders: MuffledFinder[];
-    };
-    /** Per-file finders (context-specific patterns). */
-    finders: MuffledFinder[];
-}
-/**
- * Default finder: `command || true` in a testCommand/setupCommand/cmd/command
- * string. Swallows exit codes.
- */
-export declare const findFallbackToPass: MuffledFinder;
-/**
- * `testCommand: 'true'` literal silent-pass — an unknown-language dispatch
- * arm that returns a no-op instead of throwing.
- */
-export declare const findLiteralTruePass: MuffledFinder;
-/**
- * `new SubprocessSandboxDriver({ cwd: ... })` — constructor arg silently
- * dropped in agent-eval <0.7.1. 0.7.1+ honors as fallback, but the form
- * still invites confusion; prefer `new SubprocessSandboxDriver()` with
- * cwd in the per-call HarnessConfig.
- */
-export declare const findConstructorCwdDropped: MuffledFinder;
-/**
- * `if (!expected) return true` — matcher auto-passes when ground truth is
- * absent. Inflates accuracy metrics for scenarios without expectations.
- */
-export declare const findAutoMatchNoExpectation: MuffledFinder;
-/**
- * `if (p.skipped) return true` — skip-counts-as-pass in quality scorers.
- * Use three-valued `true | false | 'skipped'` return + explicit partial
- * credit instead.
- */
-export declare const findSkipCountsAsPass: MuffledFinder;
-/**
- * The canonical default bundle. Callers can import these individually,
- * replace them, or append custom finders for project-specific patterns.
- */
-export declare const DEFAULT_FINDERS: MuffledFinder[];
-/** Finders that should run on EVERY file with the target import, not just SCAN_FILES. */
-export declare const UNIVERSAL_FINDERS: MuffledFinder[];
-/**
- * Run all finders against the configured files. Returns a flat list of
- * findings. Callers format + assert as they prefer.
- */
-export declare function scanForMuffledGates(opts: ScanOptions): MuffledFinding[];
-/**
- * Format findings into a single assert.fail-ready message. Each finding
- * carries file:line + pattern name + the offending line.
- */
-export declare function formatFindings(findings: MuffledFinding[]): string;
-//# sourceMappingURL=muffled-gate-scanner.d.ts.map

package/dist/muffled-gate-scanner.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"muffled-gate-scanner.d.ts","sourceRoot":"","sources":["../src/muffled-gate-scanner.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AAKH,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAA;IACZ,IAAI,EAAE,MAAM,CAAA;IACZ,QAAQ,EAAE,MAAM,CAAA;IAChB,OAAO,EAAE,MAAM,CAAA;CAChB;AAED,MAAM,MAAM,aAAa,GAAG,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,KAAK,cAAc,EAAE,CAAA;AAE5E,MAAM,WAAW,WAAW;IAC1B,sCAAsC;IACtC,QAAQ,EAAE,MAAM,CAAA;IAChB,oFAAoF;IACpF,SAAS,EAAE,MAAM,EAAE,CAAA;IACnB;;;;;;OAMG;IACH,UAAU,CAAC,EAAE;QACX,KAAK,EAAE,MAAM,EAAE,CAAA;QACf,UAAU,EAAE,MAAM,CAAA;QAClB,cAAc,EAAE,MAAM,CAAA;QACtB,gBAAgB,EAAE,aAAa,EAAE,CAAA;KAClC,CAAA;IACD,oDAAoD;IACpD,OAAO,EAAE,aAAa,EAAE,CAAA;CACzB;AAeD;;;GAGG;AACH,eAAO,MAAM,kBAAkB,EAAE,aAkBhC,CAAA;AAED;;;GAGG;AACH,eAAO,MAAM,mBAAmB,EAAE,aAkBjC,CAAA;AAED;;;;;GAKG;AACH,eAAO,MAAM,yBAAyB,EAAE,aAmBvC,CAAA;AAED;;;GAGG;AACH,eAAO,MAAM,0BAA0B,EAAE,aAkBxC,CAAA;AAED;;;;GAIG;AACH,eAAO,MAAM,oBAAoB,EAAE,aAkBlC,CAAA;AAED;;;GAGG;AACH,eAAO,MAAM,eAAe,EAAE,aAAa,EAK1C,CAAA;AAED,yFAAyF;AACzF,eAAO,MAAM,iBAAiB,EAAE,aAAa,EAAgC,CAAA;AAmD7E;;;GAGG;AACH,wBAAgB,mBAAmB,CAAC,IAAI,EAAE,WAAW,GAAG,cAAc,EAAE,CA+BvE;AAED;;;GAGG;AACH,wBAAgB,cAAc,CAAC,QAAQ,EAAE,cAAc,EAAE,GAAG,MAAM,CAQjE"}

package/dist/multi-layer-verifier.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"multi-layer-verifier.d.ts","sourceRoot":"","sources":["../src/multi-layer-verifier.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAEH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,WAAW,CAAA;AAI/C,MAAM,MAAM,WAAW,GAAG,MAAM,GAAG,MAAM,GAAG,SAAS,GAAG,OAAO,GAAG,SAAS,CAAA;AAE3E,MAAM,MAAM,QAAQ,GAAG,UAAU,GAAG,OAAO,GAAG,OAAO,GAAG,MAAM,CAAA;AAE9D,MAAM,WAAW,OAAO;IACtB,QAAQ,EAAE,QAAQ,CAAA;IAClB,OAAO,EAAE,MAAM,CAAA;IACf,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,mFAAmF;IACnF,KAAK,CAAC,EAAE,MAAM,CAAA;IACd;;;;OAIG;IACH,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACjC;AAED,MAAM,WAAW,WAAW;IAC1B,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,WAAW,CAAA;IACnB,6EAA6E;IAC7E,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,UAAU,EAAE,MAAM,CAAA;IAClB,QAAQ,EAAE,OAAO,EAAE,CAAA;IACnB,+CAA+C;IAC/C,MAAM,CAAC,EAAE,MAAM,CAAA;IACf;;;;;;;OAOG;IACH,WAAW,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC,CAAA;IAC3C,mFAAmF;IACnF,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACjC;AAED,MAAM,WAAW,aAAa,CAAC,GAAG,GAAG,OAAO;IAC1C,qFAAqF;IACrF,GAAG,EAAE,GAAG,CAAA;IACR,gEAAgE;IAChE,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,WAAW,CAAC,CAAA;IAClC,oEAAoE;IACpE,MAAM,EAAE,WAAW,CAAA;CACpB;AAED,MAAM,WAAW,KAAK,CAAC,GAAG,GAAG,OAAO;IAClC,IAAI,EAAE,MAAM,CAAA;IACZ,qEAAqE;IACrE,SAAS,CAAC,EAAE,MAAM,EAAE,CAAA;IACpB;;;OAGG;IACH,MAAM,CAAC,EAAE,MAAM,CAAA;IACf;;;;OAIG;IACH,sBAAsB,CAAC,EAAE,OAAO,CAAA;IAChC,gFAAgF;IAChF,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,GAAG,EAAE,CAAC,GAAG,EAAE,aAAa,CAAC,GAAG,CAAC,KAAK,OAAO,CAAC,WAAW,CAAC,GAAG,WAAW,CAAA;CACrE;AAED,MAAM,WAAW,aAAa,CAAC,GAAG,GAAG,OAAO;IAC1C,GAAG,EAAE,GAAG,CAAA;IACR;;;OAGG;IACH,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,qDAAqD;IACrD,OAAO,CAAC,EAAE,CAAC,MAAM,EAAE,WAAW,KAAK,IAAI,CAAA;CACxC;AAED;;sEAEsE;AACtE,MAAM,WAAW,kBAAmB,SAAQ,cAAc;IACxD,MAAM,EAAE,WAAW,EAAE,CAAA;IACrB,SAAS,EAAE,MAAM,CAAA;IACjB,SAAS,EAAE,MAAM,CAAA;IACjB,YAAY,EAAE,MAAM,CAAA;IACpB,UAAU,EAAE,MAAM,CAAA;IAClB,4EAA4E;IAC5E,OAAO,EAAE,OAAO,CAAA;IAChB;;;OAGG;IACH,YAAY,EAAE,MAAM,CAAA;IACpB,UAAU,EAAE,MAAM,CAAA;IAClB,SAAS,EAAE,MAAM,CAAA;IACjB,UAAU,EAAE,MAAM,CAAA;CACnB;AAID;;;;;;;;;;GAUG;AACH,wBAAgB,mBAAmB,CAAC,KAAK,EAAE;IACzC,KAAK,EAAE,MAAM,CAAA;IACb,QAAQ,EAAE,KAAK,CAAC;QAAE,QAAQ,EAAE,QAAQ,CAAC;QAAC,OAAO,CAAC,EAAE,OAAO,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;IAC1E,SAAS,EAAE,OAAO,CAAA;IAClB,SAAS,CAAC,EAAE,MAAM,CAAA;CACnB,GAAG,WAAW,CAOd;AAID,qBAAa,kBAAkB,CAAC,GAAG,GAAG,OAAO;IAC/B,OAAO,CAAC,QAAQ,CAAC,MAAM;gBAAN,MAAM,EAAE,KAAK,CAAC,GAAG,CAAC,EAAE;IAiB3C,GAAG,CAAC,IAAI,EAAE,aAAa,CAAC,GAAG,CAAC,GAAG,OAAO,CAAC,kBAAkB,CAAC;CAgFjE"}

package/dist/multi-layer-verifier.test.d.ts

@@ -1,2 +0,0 @@
-export {};
-//# sourceMappingURL=multi-layer-verifier.test.d.ts.map

package/dist/multi-layer-verifier.test.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"multi-layer-verifier.test.d.ts","sourceRoot":"","sources":["../src/multi-layer-verifier.test.ts"],"names":[],"mappings":""}

package/dist/multi-toolchain-layer.d.ts

@@ -1,80 +0,0 @@
-/**
- * Multi-toolchain layer factory + merge helper.
- *
- * Some verification stages (install, typecheck, build, lint) run the
- * SAME logical layer across multiple parallel adapters — pnpm AND npm
- * AND cargo AND forge for a polyglot scaffold. The verifier presents
- * one row per stage; the toolchain breakdown lives in `findings.detail`.
- *
- * This module provides the merge: take N independent `LayerResult`s
- * (one per adapter) and reduce them to a single `LayerResult` whose
- * status is the worst of the parts and whose findings cite the adapter
- * that produced each one. Plus a {@link multiToolchainLayer} factory
- * that runs the adapter calls in parallel + applies the reducer.
- *
- * Pure utility — composes with {@link MultiLayerVerifier}.{run}.
- */
-import type { Layer, LayerResult, VerifyContext } from './multi-layer-verifier';
-export interface AdapterRun {
-    /** Identifier for the adapter (e.g. 'pnpm', 'npm', 'cargo', 'forge'). */
-    adapter: string;
-    result: LayerResult;
-}
-export interface MergeOptions {
-    /**
-     * How to combine per-adapter `durationMs`. Default `'max'` (parallel
-     * wall-clock). Set `'sum'` when reporting total work done across
-     * adapters rather than wall time.
-     */
-    mergeDuration?: 'max' | 'sum';
-    /**
-     * Prefix finding messages with a per-adapter tag (e.g. `[pnpm] typecheck failed`).
-     * Default: no prefix (renderers read `detail.adapter` instead).
-     */
-    messagePrefixer?: (adapter: string) => string;
-    /**
-     * How to reduce per-adapter `LayerResult.diagnostics` into the merged
-     * result's diagnostics. `'max'` (default) — for each key, merged =
-     * max across adapters where value is non-null (matches "if ANY adapter
-     * saw N errors, merged saw N"). `'sum'` — sum non-null values.
-     */
-    mergeDiagnostics?: 'max' | 'sum';
-}
-/**
- * Reduce N adapter runs to a single `LayerResult` for a logical layer.
- *
- *   - status: worst of the parts (pass < skipped < fail < timeout < error)
- *   - score: weighted mean of numeric scores (skip = no contribution)
- *   - findings: union, each tagged with `detail.adapter`
- *   - durationMs: `mergeDuration` option (default 'max' for parallel wall-clock)
- *   - diagnostics: `mergeDiagnostics` option (default 'max' per key)
- *   - reason: " · "-joined `name: status` per adapter
- */
-export declare function mergeLayerResults(name: string, perAdapter: AdapterRun[], options?: MergeOptions): LayerResult;
-export interface MultiToolchainLayerConfig<Env, Adapter> {
-    name: string;
-    adapters: ReadonlyArray<Adapter>;
-    /** Adapter identifier — used in findings + reason. */
-    adapterName: (a: Adapter) => string;
-    /** Run a single adapter against the verify context. */
-    run: (a: Adapter, ctx: VerifyContext<Env>) => Promise<LayerResult> | LayerResult;
-    dependsOn?: string[];
-    weight?: number;
-    failContributesToScore?: boolean;
-    capMs?: number;
-    /**
-     * Per-adapter parallel cap. Defaults to 8 — defense in depth against a
-     * caller passing 50 adapters and fanning out 50 simultaneous subprocesses.
-     * Adapters that need higher concurrency raise this explicitly.
-     */
-    maxParallel?: number;
-}
-/**
- * Build a {@link Layer} that fans the same logical stage across N adapters
- * in parallel and merges via {@link mergeLayerResults}.
- *
- * Per-adapter throws are caught + converted to `status: 'error'` results
- * so one bad adapter doesn't poison the whole layer.
- */
-export declare function multiToolchainLayer<Env, Adapter>(config: MultiToolchainLayerConfig<Env, Adapter>): Layer<Env>;
-//# sourceMappingURL=multi-toolchain-layer.d.ts.map

package/dist/multi-toolchain-layer.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"multi-toolchain-layer.d.ts","sourceRoot":"","sources":["../src/multi-toolchain-layer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,OAAO,KAAK,EACV,KAAK,EACL,WAAW,EAGX,aAAa,EACd,MAAM,wBAAwB,CAAA;AAiC/B,MAAM,WAAW,UAAU;IACzB,yEAAyE;IACzE,OAAO,EAAE,MAAM,CAAA;IACf,MAAM,EAAE,WAAW,CAAA;CACpB;AAED,MAAM,WAAW,YAAY;IAC3B;;;;OAIG;IACH,aAAa,CAAC,EAAE,KAAK,GAAG,KAAK,CAAA;IAC7B;;;OAGG;IACH,eAAe,CAAC,EAAE,CAAC,OAAO,EAAE,MAAM,KAAK,MAAM,CAAA;IAC7C;;;;;OAKG;IACH,gBAAgB,CAAC,EAAE,KAAK,GAAG,KAAK,CAAA;CACjC;AAED;;;;;;;;;GASG;AACH,wBAAgB,iBAAiB,CAC/B,IAAI,EAAE,MAAM,EACZ,UAAU,EAAE,UAAU,EAAE,EACxB,OAAO,GAAE,YAAiB,GACzB,WAAW,CAiFb;AAID,MAAM,WAAW,yBAAyB,CAAC,GAAG,EAAE,OAAO;IACrD,IAAI,EAAE,MAAM,CAAA;IACZ,QAAQ,EAAE,aAAa,CAAC,OAAO,CAAC,CAAA;IAChC,sDAAsD;IACtD,WAAW,EAAE,CAAC,CAAC,EAAE,OAAO,KAAK,MAAM,CAAA;IACnC,uDAAuD;IACvD,GAAG,EAAE,CAAC,CAAC,EAAE,OAAO,EAAE,GAAG,EAAE,aAAa,CAAC,GAAG,CAAC,KAAK,OAAO,CAAC,WAAW,CAAC,GAAG,WAAW,CAAA;IAChF,SAAS,CAAC,EAAE,MAAM,EAAE,CAAA;IACpB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,sBAAsB,CAAC,EAAE,OAAO,CAAA;IAChC,KAAK,CAAC,EAAE,MAAM,CAAA;IACd;;;;OAIG;IACH,WAAW,CAAC,EAAE,MAAM,CAAA;CACrB;AAED;;;;;;GAMG;AACH,wBAAgB,mBAAmB,CAAC,GAAG,EAAE,OAAO,EAC9C,MAAM,EAAE,yBAAyB,CAAC,GAAG,EAAE,OAAO,CAAC,GAC9C,KAAK,CAAC,GAAG,CAAC,CAuDZ"}

package/dist/multi-toolchain-layer.test.d.ts

@@ -1,2 +0,0 @@
-export {};
-//# sourceMappingURL=multi-toolchain-layer.test.d.ts.map

package/dist/multi-toolchain-layer.test.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"multi-toolchain-layer.test.d.ts","sourceRoot":"","sources":["../src/multi-toolchain-layer.test.ts"],"names":[],"mappings":""}

package/dist/multishot/default-tools.d.ts

@@ -1,34 +0,0 @@
-import type { MultishotToolDefinition, MultishotToolExecutor } from './types';
-export declare const DEFAULT_RESEARCHER_MODEL = "openai/gpt-4o-mini";
-export declare const DEFAULT_CODER_MODEL = "openai/gpt-4o-mini";
-export interface DefaultResearcherConfig {
-    /** Replace the system prompt to bias the researcher toward a domain's
-     *  citation style. Defaults to a generic "cite sources by name" prompt. */
-    systemPrompt?: string;
-    model?: string;
-}
-export interface DefaultCoderConfig {
-    /** Replace the system prompt to bias the coder toward a language /
-     *  framework / artifact style. */
-    systemPrompt?: string;
-    model?: string;
-}
-export declare const DEFAULT_DELEGATE_RESEARCH_TOOL: MultishotToolDefinition;
-export declare const DEFAULT_DELEGATE_CODE_TOOL: MultishotToolDefinition;
-export declare function createResearchExecutor(config?: DefaultResearcherConfig): MultishotToolExecutor;
-export declare function createCodeExecutor(config?: DefaultCoderConfig): MultishotToolExecutor;
-export interface DefaultToolsConfig {
-    research?: DefaultResearcherConfig;
-    code?: DefaultCoderConfig;
-    /** When true (default), each tool result is recorded as a typed artifact:
-     *  research → type='research', code → type='code'. */
-    recordArtifacts?: boolean;
-}
-export interface DefaultToolsBundle {
-    tools: MultishotToolDefinition[];
-    executors: Record<string, MultishotToolExecutor>;
-    artifactTypeFor: (toolName: string) => string | undefined;
-}
-export declare function defaultDelegationTools(config?: DefaultToolsConfig): DefaultToolsBundle;
-export { defaultRouterBaseUrl } from './router';
-//# sourceMappingURL=default-tools.d.ts.map

package/dist/multishot/default-tools.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"default-tools.d.ts","sourceRoot":"","sources":["../../src/multishot/default-tools.ts"],"names":[],"mappings":"AAQA,OAAO,KAAK,EAAE,uBAAuB,EAAE,qBAAqB,EAAE,MAAM,SAAS,CAAA;AAE7E,eAAO,MAAM,wBAAwB,uBAAuB,CAAA;AAC5D,eAAO,MAAM,mBAAmB,uBAAuB,CAAA;AAEvD,MAAM,WAAW,uBAAuB;IACtC;+EAC2E;IAC3E,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,KAAK,CAAC,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,kBAAkB;IACjC;sCACkC;IAClC,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,KAAK,CAAC,EAAE,MAAM,CAAA;CACf;AAQD,eAAO,MAAM,8BAA8B,EAAE,uBAkB5C,CAAA;AAED,eAAO,MAAM,0BAA0B,EAAE,uBAkBxC,CAAA;AAED,wBAAgB,sBAAsB,CACpC,MAAM,GAAE,uBAA4B,GACnC,qBAAqB,CAoBvB;AAED,wBAAgB,kBAAkB,CAAC,MAAM,GAAE,kBAAuB,GAAG,qBAAqB,CAoBzF;AAED,MAAM,WAAW,kBAAkB;IACjC,QAAQ,CAAC,EAAE,uBAAuB,CAAA;IAClC,IAAI,CAAC,EAAE,kBAAkB,CAAA;IACzB;0DACsD;IACtD,eAAe,CAAC,EAAE,OAAO,CAAA;CAC1B;AAED,MAAM,WAAW,kBAAkB;IACjC,KAAK,EAAE,uBAAuB,EAAE,CAAA;IAChC,SAAS,EAAE,MAAM,CAAC,MAAM,EAAE,qBAAqB,CAAC,CAAA;IAChD,eAAe,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,MAAM,GAAG,SAAS,CAAA;CAC1D;AAED,wBAAgB,sBAAsB,CAAC,MAAM,GAAE,kBAAuB,GAAG,kBAAkB,CAU1F;AAED,OAAO,EAAE,oBAAoB,EAAE,MAAM,UAAU,CAAA"}

package/dist/multishot/index.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/multishot/index.ts"],"names":[],"mappings":"AAEA,OAAO,EACL,kBAAkB,EAClB,sBAAsB,EACtB,mBAAmB,EACnB,0BAA0B,EAC1B,8BAA8B,EAC9B,wBAAwB,EACxB,KAAK,kBAAkB,EACvB,KAAK,uBAAuB,EAC5B,KAAK,kBAAkB,EACvB,KAAK,kBAAkB,EACvB,sBAAsB,GACvB,MAAM,iBAAiB,CAAA;AAExB,OAAO,EACL,mBAAmB,EACnB,KAAK,WAAW,EAChB,KAAK,cAAc,EACnB,KAAK,UAAU,EACf,gBAAgB,EAChB,gBAAgB,EAChB,QAAQ,GACT,MAAM,UAAU,CAAA;AACjB,OAAO,EACL,KAAK,kBAAkB,EACvB,KAAK,kBAAkB,EACvB,KAAK,kBAAkB,EACvB,KAAK,sBAAsB,EAC3B,oBAAoB,EACpB,KAAK,eAAe,EACpB,KAAK,yBAAyB,EAC9B,KAAK,wBAAwB,EAC7B,kBAAkB,GACnB,MAAM,UAAU,CAAA;AACjB,OAAO,EAAE,KAAK,mBAAmB,EAAE,YAAY,EAAE,MAAM,aAAa,CAAA;AAEpE,OAAO,EACL,oBAAoB,EACpB,kBAAkB,EAClB,KAAK,uBAAuB,EAC5B,KAAK,wBAAwB,EAC7B,KAAK,cAAc,EACnB,mBAAmB,EACnB,gBAAgB,GACjB,MAAM,UAAU,CAAA;AAEjB,OAAO,EACL,KAAK,iBAAiB,EACtB,yBAAyB,EACzB,KAAK,gBAAgB,EACrB,KAAK,gBAAgB,EACrB,KAAK,eAAe,EACpB,KAAK,cAAc,EACnB,KAAK,uBAAuB,EAC5B,KAAK,qBAAqB,GAC3B,MAAM,SAAS,CAAA"}

package/dist/multishot/judges.d.ts

@@ -1,32 +0,0 @@
-import type { JudgeScore } from '../campaign/types';
-export type { JudgeScore } from '../campaign/types';
-export declare const DEFAULT_JUDGE_MODEL = "openai/gpt-4o-mini";
-export interface JudgeDimension {
-    /** JSON field name + score key. */
-    key: string;
-    /** Description shown in the judge's user prompt. */
-    description: string;
-}
-export interface JudgeConfig<TInput> {
-    /** Display name (for trace + log). */
-    name: string;
-    /** Model used for this judge. */
-    model?: string;
-    /** 0-10 scored dimensions. */
-    dimensions: JudgeDimension[];
-    /** Judge system prompt — sets persona + JSON-only constraint. */
-    systemPrompt: string;
-    /** Build the user prompt from the typed input. Must include "Respond with
-     *  ONLY this JSON: { ... }" listing each dimension key. */
-    buildPrompt: (input: TInput) => string;
-    /** Optional model + api overrides. */
-    apiKey?: string;
-    baseUrl?: string;
-}
-export declare function runJudge<TInput>(judge: JudgeConfig<TInput>, input: TInput): Promise<JudgeScore>;
-/** Convenience: stringified dimension list for inclusion in a judge prompt.
- *  Returns lines like `- audience_fit: Does this match what the audience cares about? (0-10)`. */
-export declare function renderDimensions(dims: readonly JudgeDimension[]): string;
-/** Convenience: build the "Respond with ONLY this JSON" footer for a judge prompt. */
-export declare function renderJsonFooter(dims: readonly JudgeDimension[]): string;
-//# sourceMappingURL=judges.d.ts.map

package/dist/multishot/judges.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"judges.d.ts","sourceRoot":"","sources":["../../src/multishot/judges.ts"],"names":[],"mappings":"AASA,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,mBAAmB,CAAA;AAKnD,YAAY,EAAE,UAAU,EAAE,MAAM,mBAAmB,CAAA;AAEnD,eAAO,MAAM,mBAAmB,uBAAuB,CAAA;AAEvD,MAAM,WAAW,cAAc;IAC7B,mCAAmC;IACnC,GAAG,EAAE,MAAM,CAAA;IACX,oDAAoD;IACpD,WAAW,EAAE,MAAM,CAAA;CACpB;AAED,MAAM,WAAW,WAAW,CAAC,MAAM;IACjC,sCAAsC;IACtC,IAAI,EAAE,MAAM,CAAA;IACZ,iCAAiC;IACjC,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,8BAA8B;IAC9B,UAAU,EAAE,cAAc,EAAE,CAAA;IAC5B,iEAAiE;IACjE,YAAY,EAAE,MAAM,CAAA;IACpB;+DAC2D;IAC3D,WAAW,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,MAAM,CAAA;IACtC,sCAAsC;IACtC,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,OAAO,CAAC,EAAE,MAAM,CAAA;CACjB;AAED,wBAAsB,QAAQ,CAAC,MAAM,EACnC,KAAK,EAAE,WAAW,CAAC,MAAM,CAAC,EAC1B,KAAK,EAAE,MAAM,GACZ,OAAO,CAAC,UAAU,CAAC,CA2DrB;AAED;kGACkG;AAClG,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,SAAS,cAAc,EAAE,GAAG,MAAM,CAExE;AAED,sFAAsF;AACtF,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,SAAS,cAAc,EAAE,GAAG,MAAM,CAGxE"}