@swarmdo/cli 1.4.8

package/dist/src/swarmvector/model-router.js

@@ -0,0 +1,1162 @@
+/**
+ * Intelligent Model Router — lexical complexity heuristic + Thompson bandit
+ *
+ * Dynamically routes requests to the optimal Claude model (haiku/sonnet/opus)
+ * based on task complexity, uncertainty, and online-learned routing outcomes.
+ *
+ * Mechanism (shipped):
+ * - Complexity score = blend of lexical, semantic-depth, task-scope, and
+ *   uncertainty heuristics (see `computeLexicalComplexity` and friends).
+ *   Pure JS arithmetic — no model load, no tensor math.
+ * - Model selection = Thompson-sampling Beta-Bernoulli bandit with
+ *   complexity-bucketed Beta(α,β) priors, persisted to
+ *   `.swarm/model-router-state.json` and updated by `recordOutcome` after
+ *   each routing decision.
+ * - Uncertainty quantification + a circuit breaker drive escalation when
+ *   the bandit's confidence is low or downstream failures are observed.
+ *
+ * Routing Strategy:
+ * - Haiku: high confidence, low complexity (fast, cheap)
+ * - Sonnet: medium confidence, moderate complexity (balanced)
+ * - Opus: low confidence, high complexity (most capable)
+ *
+ * Note (#2329): An earlier design (ADR-026 + this file's previous header)
+ * described a Tiny-Dancer / FastGRNN neural router with embedding-based
+ * complexity scoring. That path was never wired in directly.
+ *
+ * Note (ADR-148, #2334): The cost-optimal neural router is now wired as an
+ * optional, gated addition via `./neural-router.ts` (which uses
+ * `@metaharness/router`, optionally accelerated by `@swarmvector/tiny-dancer`).
+ * It is double-gated on `SWARMDO_ROUTER_NEURAL=1` + an embedding being
+ * supplied + a corpus/artifact being loadable. When any gate is closed the
+ * shipped heuristic + bandit path runs unchanged and the result carries
+ * `routedBy: 'heuristic'` (default) or `'bandit-fallback'` (neural enabled
+ * but declined). When all gates are open and a backend resolves, the result
+ * carries `routedBy: 'metaharness-knn' | 'metaharness-krr' | 'fastgrnn'`.
+ *
+ * @module model-router
+ */
+import { existsSync, mkdirSync, readFileSync, writeFileSync } from 'fs';
+import { dirname, join } from 'path';
+// ----------------------------------------------------------------------------
+// Lazy module loaders — initialised once per process. The dynamic-import calls
+// keep these modules off the critical-path-cold-start, but we want subsequent
+// route() calls to pay only a Map lookup, not a new Promise per call.
+// ----------------------------------------------------------------------------
+let _neuralRouterMod = null;
+function loadNeuralRouter() {
+    if (_neuralRouterMod === null)
+        _neuralRouterMod = import('./neural-router.js');
+    return _neuralRouterMod;
+}
+let _trajectoryMod = null;
+function loadTrajectoryRecorder() {
+    if (_trajectoryMod === null)
+        _trajectoryMod = import('./router-trajectory.js');
+    return _trajectoryMod;
+}
+// ADR-150 iter 11–12 — parallel-decision recorder for the SelfEvolvingRouter
+// promotion gate. Dynamic-imported lazily so the routing hot path never
+// pays the load cost when SWARMDO_ROUTER_PARALLEL_LOG is unset.
+let _parallelRecorderMod = null;
+function loadParallelRecorder() {
+    if (_parallelRecorderMod === null)
+        _parallelRecorderMod = import('./router-parallel-recorder.js');
+    return _parallelRecorderMod;
+}
+let _altsCache = null;
+let _altsProbeDone = false;
+function loadOpenRouterAlts() {
+    if (_altsProbeDone)
+        return _altsCache;
+    _altsProbeDone = true;
+    try {
+        // Probe candidate paths: explicit env override, then asset locations
+        // relative to this file (src dev) and the dist build.
+        const explicit = process.env.SWARMDO_ROUTER_OPENROUTER_ALTS;
+        const candidates = [];
+        if (explicit)
+            candidates.push(explicit);
+        // Probe asset dirs without using import.meta.url so this stays compatible
+        // with both CJS and ESM consumers of the compiled .js.
+        candidates.push(join(process.cwd(), 'v3', '@swarmdo', 'cli', 'assets', 'model-router', 'openrouter-alts.json'));
+        candidates.push(join(process.cwd(), 'assets', 'model-router', 'openrouter-alts.json'));
+        for (const p of candidates) {
+            if (existsSync(p)) {
+                _altsCache = JSON.parse(readFileSync(p, 'utf8'));
+                return _altsCache;
+            }
+        }
+    }
+    catch {
+        // Silent — alts are optional.
+    }
+    return null;
+}
+/** Return the resolved provider + OpenRouter model for the picked tier. */
+function resolveExecutionProvider(model) {
+    const explicit = process.env.SWARMDO_ROUTER_PROVIDER?.toLowerCase();
+    // Default: anthropic unless explicitly set to openrouter, or OPENROUTER_API_KEY
+    // is the only credential present (matches agent-execute-core's selection).
+    const hasOpenRouter = !!process.env.OPENROUTER_API_KEY;
+    const hasAnthropic = !!process.env.ANTHROPIC_API_KEY;
+    const wantOR = explicit === 'openrouter' ||
+        (!hasAnthropic && hasOpenRouter && explicit !== 'anthropic');
+    if (!wantOR)
+        return { provider: 'anthropic' };
+    const alts = loadOpenRouterAlts();
+    if (!alts)
+        return { provider: 'openrouter' }; // OR provider but no alt slug
+    const entry = alts.tiers[model];
+    if (!entry?.openrouter_alt)
+        return { provider: 'openrouter' };
+    return { provider: 'openrouter', openrouterModel: entry.openrouter_alt };
+}
+/**
+ * Model capabilities and characteristics
+ */
+export const MODEL_CAPABILITIES = {
+    haiku: {
+        maxComplexity: 0.4,
+        costMultiplier: 0.04, // ~25x cheaper than Opus
+        speedMultiplier: 3.0, // ~3x faster than Sonnet
+        description: 'Fast, cost-effective for simple tasks',
+    },
+    sonnet: {
+        maxComplexity: 0.7,
+        costMultiplier: 0.2, // ~5x cheaper than Opus
+        speedMultiplier: 1.5, // ~1.5x faster than Opus
+        description: 'Balanced capability and cost',
+    },
+    opus: {
+        maxComplexity: 1.0,
+        costMultiplier: 1.0, // Baseline
+        speedMultiplier: 1.0, // Baseline
+        description: 'Most capable for complex reasoning',
+    },
+    inherit: {
+        maxComplexity: 1.0,
+        costMultiplier: 1.0,
+        speedMultiplier: 1.0,
+        description: 'Use parent model selection',
+    },
+};
+/**
+ * Complexity indicators for task classification
+ */
+export const COMPLEXITY_INDICATORS = {
+    high: [
+        'architect', 'design', 'refactor', 'optimize', 'security', 'audit',
+        'complex', 'analyze', 'investigate', 'debug', 'performance', 'scale',
+        'distributed', 'concurrent', 'algorithm', 'system', 'integration',
+    ],
+    medium: [
+        'implement', 'feature', 'add', 'update', 'modify', 'fix', 'test',
+        'review', 'validate', 'check', 'improve', 'enhance', 'extend',
+    ],
+    low: [
+        'simple', 'typo', 'comment', 'format', 'rename', 'move', 'copy',
+        'delete', 'documentation', 'readme', 'config', 'version', 'bump',
+    ],
+};
+/**
+ * Cost-adjusted Bernoulli rewards for Thompson sampling updates. Higher
+ * reward when the right tier is chosen — Haiku-success > Sonnet-success >
+ * Opus-success because Opus-success on a simple task is wasteful even when
+ * the answer is correct. Escalations get partial credit at best (Sonnet) or
+ * zero (Haiku/Opus) since they signal the initial choice was wrong.
+ */
+const BANDIT_REWARDS = {
+    haiku: { success: 1.0, failure: 0.0, escalated: 0.0 },
+    sonnet: { success: 0.7, failure: 0.0, escalated: 0.1 },
+    opus: { success: 0.4, failure: 0.0, escalated: 0.0 },
+    inherit: { success: 0.5, failure: 0.0, escalated: 0.0 },
+};
+function complexityBucket(score) {
+    if (score < 0.4)
+        return 'low'; // haiku territory
+    if (score < 0.7)
+        return 'med'; // sonnet territory
+    return 'high'; // opus territory
+}
+// ============================================================================
+// Beta Sampling for Thompson Sampling Bandit
+// ============================================================================
+/**
+ * Standard normal sample via Box-Muller. Used by Marsaglia-Tsang Gamma.
+ * Module-local so the bandit doesn't pull in a heavy stats dep.
+ */
+function sampleStandardNormal() {
+    const u1 = Math.random() || 1e-12; // avoid log(0)
+    const u2 = Math.random();
+    return Math.sqrt(-2 * Math.log(u1)) * Math.cos(2 * Math.PI * u2);
+}
+/**
+ * Sample from Gamma(shape α, scale=1). Marsaglia & Tsang (2000), with the
+ * standard "boost α<1 by α+1 then scale by U^(1/α)" trick for shape parameters
+ * smaller than 1. O(1) expected, no rejection-loop pathology in practice.
+ */
+function sampleGamma(alpha) {
+    if (alpha < 1) {
+        const u = Math.random() || 1e-12;
+        return sampleGamma(alpha + 1) * Math.pow(u, 1 / alpha);
+    }
+    const d = alpha - 1 / 3;
+    const c = 1 / Math.sqrt(9 * d);
+    while (true) {
+        let x;
+        let v;
+        do {
+            x = sampleStandardNormal();
+            v = 1 + c * x;
+        } while (v <= 0);
+        v = v * v * v;
+        const u = Math.random();
+        const xx = x * x;
+        if (u < 1 - 0.0331 * xx * xx)
+            return d * v;
+        if (Math.log(u) < 0.5 * xx + d * (1 - v + Math.log(v)))
+            return d * v;
+    }
+}
+/**
+ * Sample θ ~ Beta(α, β) via the identity Beta(α,β) = X / (X+Y) where
+ * X ~ Gamma(α), Y ~ Gamma(β). Returns the mean for degenerate α+β=0
+ * (shouldn't happen in practice but defensive).
+ */
+export function sampleBeta(alpha, beta) {
+    if (alpha <= 0 || beta <= 0)
+        return 0.5;
+    const x = sampleGamma(alpha);
+    const y = sampleGamma(beta);
+    const denom = x + y;
+    return denom > 0 ? x / denom : 0.5;
+}
+/**
+ * Default uniform priors (no prior knowledge). Beta(1,1) is the standard
+ * Bayesian-Bernoulli starting point — uniform over [0,1].
+ */
+function defaultBanditPriors() {
+    return {
+        haiku: { alpha: 1, beta: 1 },
+        sonnet: { alpha: 1, beta: 1 },
+        opus: { alpha: 1, beta: 1 },
+        inherit: { alpha: 1, beta: 1 },
+    };
+}
+/** Uniform priors for every complexity bucket (cold start). */
+function defaultBucketedPriors() {
+    return { low: defaultBanditPriors(), med: defaultBanditPriors(), high: defaultBanditPriors() };
+}
+/** ADR-149 — empty per-modelId shadow priors. Each bucket starts as `{}`; entries
+ *  populate on first `recordOutcomeByModelId(task, modelId, outcome)` call. */
+function defaultBucketedPriorsById() {
+    return { low: {}, med: {}, high: {} };
+}
+function clonePriors(p) {
+    return { haiku: { ...p.haiku }, sonnet: { ...p.sonnet }, opus: { ...p.opus }, inherit: { ...p.inherit } };
+}
+/**
+ * Forward-migrate a persisted `priors` field of any layout to the bucketed
+ * shape, never throwing (ADR-142):
+ *  - missing/garbage → fresh uniform buckets
+ *  - already bucketed (has `low.haiku`) → kept, backfilling any missing bucket
+ *  - flat per-model (v1 bandit) → seed ALL buckets from it (lossless: prior
+ *    learning becomes a shared starting point that then diverges per bucket)
+ */
+function migratePriors(p) {
+    if (!p || typeof p !== 'object')
+        return defaultBucketedPriors();
+    const obj = p;
+    if (obj.low && typeof obj.low === 'object' && obj.low.haiku) {
+        return {
+            low: obj.low,
+            med: obj.med ?? clonePriors(obj.low),
+            high: obj.high ?? clonePriors(obj.low),
+        };
+    }
+    if (obj.haiku && typeof obj.haiku.alpha === 'number') {
+        const flat = obj;
+        return { low: clonePriors(flat), med: clonePriors(flat), high: clonePriors(flat) };
+    }
+    return defaultBucketedPriors();
+}
+// ============================================================================
+// Default Configuration
+// ============================================================================
+// #2250 — env override for maxUncertainty so callers can suppress the
+// escalation without recompiling. Parsed once at module load; invalid /
+// out-of-range values fall through to the default below.
+function envMaxUncertainty() {
+    const raw = process.env.SWARMDO_MAX_UNCERTAINTY;
+    if (!raw)
+        return undefined;
+    const n = Number(raw);
+    if (!Number.isFinite(n) || n < 0 || n > 1)
+        return undefined;
+    return n;
+}
+const DEFAULT_CONFIG = {
+    confidenceThreshold: 0.85,
+    maxUncertainty: envMaxUncertainty() ?? 0.15,
+    enableCircuitBreaker: true,
+    circuitBreakerThreshold: 5,
+    statePath: '.swarm/model-router-state.json',
+    autoSaveInterval: 1, // Save after every decision for CLI persistence
+    enableCostOptimization: true,
+    preferSpeed: true,
+};
+// Posterior mean of a Beta(α,β) prior — used by the #2250 escalation guard
+// to detect when the bandit has *learned* the escalation target is worse.
+function priorMean(p) {
+    return p.alpha / (p.alpha + p.beta);
+}
+// ============================================================================
+// Model Router Implementation
+// ============================================================================
+/**
+ * Intelligent Model Router using complexity-based routing
+ */
+export class ModelRouter {
+    config;
+    state;
+    decisionCount = 0;
+    consecutiveFailures = {
+        haiku: 0,
+        sonnet: 0,
+        opus: 0,
+        inherit: 0,
+    };
+    /**
+     * ADR-148 — in-memory counters surfaced via `getStats()` and read by the
+     * `hooks_intelligence_stats` MCP tool. Process-local, not persisted (these
+     * are operational metrics, not authoritative state — see ADR-074/086).
+     */
+    routedByCounts = {
+        heuristic: 0,
+        'bandit-fallback': 0,
+        hybrid: 0,
+    };
+    neuralBackendCounts = {
+        'metaharness-knn': 0,
+        'metaharness-krr': 0,
+        fastgrnn: 0,
+    };
+    abDisagreements = 0;
+    abComparisons = 0;
+    constructor(config = {}) {
+        this.config = { ...DEFAULT_CONFIG, ...config };
+        this.state = this.loadState();
+    }
+    /**
+     * Route a task to the optimal model.
+     *
+     * When `embedding` is supplied and `SWARMDO_ROUTER_NEURAL=1` is set,
+     * the cost-optimal neural backend (ADR-148) is consulted first; its
+     * decision is used when its `metBar` clears the configured quality bar
+     * and `routedBy` reflects which backend produced the decision. Otherwise
+     * the shipped heuristic + Thompson bandit path runs (byte-identical to
+     * the pre-ADR-148 behavior) and the result carries `routedBy:
+     * 'bandit-fallback'` (neural was enabled but declined) or
+     * `'heuristic'` (neural was disabled).
+     */
+    async route(task, embedding) {
+        const startTime = performance.now();
+        // Analyze task complexity
+        const complexity = this.analyzeComplexity(task, embedding);
+        // ADR-148 — optional neural cost-optimal path (gated, opt-in).
+        //
+        // Hybrid math: we use the neural's per-candidate predicted quality as a
+        // weighted prior on the bandit's Beta(α,β) posterior, rather than
+        // overriding the bandit's pick. With weight=w (env-tunable, default 5)
+        // each candidate's Beta becomes Beta(α + q*w, β + (1-q)*w). Cold start
+        // → neural dominates (α+β ≈ 2 + w); many real outcomes → bandit
+        // dominates (α+β >> w). The persistent bandit state is unchanged.
+        //
+        // `bandit-fallback` is reserved for the case where neural was enabled
+        // but the backend returned no decision at all (artifact load failed,
+        // dim mismatch, etc.). In that case we route via pure bandit.
+        let neuralPrior = null;
+        let neuralBackend = undefined;
+        let neuralModelId = undefined;
+        let neuralDeclined = false;
+        // iter 46 — capture iter 45's ensemble-disagreement diagnostic so we can
+        // persist it to the trajectory below; downstream tuners (future) can
+        // analyze the distribution to recommend an iter 44 threshold.
+        let neuralEnsembleDisagreement = undefined;
+        if (embedding && embedding.length > 0 && process.env.SWARMDO_ROUTER_NEURAL === '1') {
+            try {
+                const { tryCostOptimalRoute } = await loadNeuralRouter();
+                // ADR-149 iter 15 — pass the task's complexity bucket through so
+                // the neural-router's per-modelId Thompson (when gated on) can
+                // use the bucket-specific prior instead of marginalising.
+                const nr = await tryCostOptimalRoute(embedding, { complexityBucket: complexityBucket(complexity.score) });
+                if (nr) {
+                    // ADR-149: capture the concrete picked model id (the cost-optimal
+                    // pick across all candidates, not just within a tier).
+                    neuralModelId = nr.modelId;
+                    neuralEnsembleDisagreement = nr.ensembleDisagreement;
+                    // Build per-tier quality map for the bandit prior. The neural
+                    // backend returns per-model alternatives (e.g. 7 distinct slugs);
+                    // for the Beta-prior bump we collapse to tier by taking the MAX
+                    // predicted quality within each tier — the tier whose best
+                    // candidate is best overall is the tier the bandit should favor.
+                    // Then we apply rank-based scaling [1.0, 0.5, 0.2] across the
+                    // three tiers so the Beta bumps are well-separated (raw KRR
+                    // outputs are typically in a narrow band — see ADR-148 phase 1).
+                    const sorted = [...nr.alternatives].sort((a, b) => b.predictedQuality - a.predictedQuality);
+                    if (!sorted.find(s => s.model === nr.model)) {
+                        sorted.unshift({ model: nr.model, modelId: nr.modelId, predictedQuality: nr.predictedQuality, costPerMTok: 0 });
+                    }
+                    // Collapse per-model alternatives → per-tier MAX quality, preserving
+                    // first-occurrence order (which is best-quality first thanks to the sort).
+                    const tierMaxQ = {};
+                    const tierOrder = [];
+                    for (const a of sorted) {
+                        if (tierMaxQ[a.model] === undefined) {
+                            tierMaxQ[a.model] = a.predictedQuality;
+                            tierOrder.push(a.model);
+                        }
+                        else if (a.predictedQuality > (tierMaxQ[a.model] ?? 0)) {
+                            tierMaxQ[a.model] = a.predictedQuality;
+                        }
+                    }
+                    const rankQ = [1.0, 0.5, 0.2];
+                    const qualities = {};
+                    for (let i = 0; i < tierOrder.length && i < rankQ.length; i++) {
+                        qualities[tierOrder[i]] = rankQ[i];
+                    }
+                    const weight = parseFloat(process.env.SWARMDO_ROUTER_NEURAL_WEIGHT ?? '5') || 5;
+                    neuralPrior = { qualities, weight };
+                    neuralBackend = nr.routedBy;
+                }
+                else {
+                    neuralDeclined = true;
+                }
+            }
+            catch {
+                // Silent — neural path is best-effort.
+            }
+        }
+        const scores = this.computeModelScores(complexity);
+        const adjustedScores = this.applyCircuitBreaker(scores);
+        // A/B mode: compute the pure-bandit pick alongside the hybrid pick so
+        // we can log disagreement. Both samples are drawn from the same Beta
+        // posteriors but with/without the neural prior bump — useful for
+        // measuring real lift before flipping defaults.
+        //
+        // iter 37 — TWO knobs:
+        //   SWARMDO_ROUTER_AB=1                    → A/B on every call (orig)
+        //   SWARMDO_ROUTER_AB_SAMPLE_RATE=<0..1>   → A/B on a sampled subset,
+        //     keyed deterministically by task_hash so the same task always falls
+        //     in or out of the sample. Lets production accumulate ab_pair data
+        //     passively at low overhead (e.g. 0.05 = 5% of decisions).
+        //   Both set → SAMPLE_RATE wins (more specific).
+        const rateRaw = process.env.SWARMDO_ROUTER_AB_SAMPLE_RATE;
+        const rate = rateRaw ? Math.max(0, Math.min(1, parseFloat(rateRaw) || 0)) : 0;
+        const allOn = process.env.SWARMDO_ROUTER_AB === '1';
+        let inSample = false;
+        if (rate > 0) {
+            // Deterministic sample by FNV-1a-32 of the task text. Same task always
+            // gets the same A/B decision across re-runs (reproducible tests, stable
+            // population over time). Inlined to avoid the async router-trajectory
+            // import on the hot path.
+            let h = 0x811c9dc5 >>> 0;
+            for (let i = 0; i < task.length; i++) {
+                h ^= task.charCodeAt(i);
+                h = (h + ((h << 1) + (h << 4) + (h << 7) + (h << 8) + (h << 24))) >>> 0;
+            }
+            inSample = (h % 10000) / 10000 < rate;
+        }
+        const abEnabled = (rate > 0 ? inSample : allOn) && neuralPrior !== null;
+        let abPair;
+        if (abEnabled) {
+            // Pre-compute the bandit-only pick. Single extra Thompson sample per
+            // call — independent draws are noisy but cheap (~3 Beta samples ≈
+            // <10 μs). For a per-call disagreement signal one draw is fine; over
+            // N decisions the rate stabilises.
+            const banditOnly = this.selectModel(adjustedScores, complexity.score, undefined);
+            const picked = this.selectModel(adjustedScores, complexity.score, neuralPrior ?? undefined);
+            abPair = {
+                bandit_pick: banditOnly.model,
+                hybrid_pick: picked.model,
+                disagree: banditOnly.model !== picked.model,
+            };
+            this.abComparisons++;
+            if (abPair.disagree)
+                this.abDisagreements++;
+            // ADR-150 iter 12 — opt-in parallel-decision recorder. No-op when
+            // SWARMDO_ROUTER_PARALLEL_LOG is unset (the default), so this
+            // adds zero overhead to the default routing path. Fire-and-forget
+            // dynamic-import + recordPair; never blocks the route() return.
+            if (process.env.SWARMDO_ROUTER_PARALLEL_LOG === '1') {
+                loadParallelRecorder().then((mod) => {
+                    try {
+                        mod.recordPair({
+                            task,
+                            bandit: {
+                                pick: banditOnly.model,
+                                predictedQuality: banditOnly.confidence,
+                                predictedCostUsd: 0, // bandit doesn't price; analyzer uses outcome
+                                backend: 'thompson-bandit',
+                            },
+                            ser: {
+                                pick: picked.model,
+                                predictedQuality: picked.confidence,
+                                predictedCostUsd: 0,
+                                backend: neuralPrior ? 'metaharness-router-hybrid' : 'bandit-only',
+                            },
+                        });
+                    }
+                    catch {
+                        // ADR-150 rule #3 — never throw from the routing path.
+                    }
+                }).catch(() => { });
+            }
+            var pickedForResult = picked; // eslint-disable-line no-var
+        }
+        else {
+            var pickedForResult = this.selectModel(adjustedScores, complexity.score, neuralPrior ?? undefined); // eslint-disable-line no-var
+        }
+        const model = pickedForResult.model;
+        const confidence = pickedForResult.confidence;
+        const uncertainty = pickedForResult.uncertainty;
+        const routedBy = neuralPrior
+            ? 'hybrid'
+            : neuralDeclined ? 'bandit-fallback' : 'heuristic';
+        const inferenceTimeUs = (performance.now() - startTime) * 1000;
+        // ADR-148 phase 2 — resolve OpenRouter alt for the picked tier when the
+        // execution provider is OpenRouter. Free of side effects on the bandit;
+        // purely advisory metadata for downstream agent-execute-core.
+        //
+        // ADR-149 — when the neural backend returned a concrete `modelId` that
+        // is NOT an Anthropic slug, we override the provider to 'openrouter'
+        // and use the modelId as the openrouterModel. Otherwise consumers
+        // dispatching on `model` would call the Anthropic SDK for a model id
+        // the SDK can't reach (e.g. 'inclusionai/ling-2.6-flash'), losing the
+        // cost-optimal pick. When the neural picked an Anthropic id we keep
+        // the standard provider resolution since the Anthropic SDK can serve it.
+        let exec = resolveExecutionProvider(model);
+        if (neuralModelId && !neuralModelId.startsWith('anthropic/')) {
+            exec = { provider: 'openrouter', openrouterModel: neuralModelId };
+        }
+        // Build result
+        const result = {
+            model,
+            confidence,
+            uncertainty,
+            complexity: complexity.score,
+            reasoning: this.buildReasoning(model, complexity, confidence),
+            alternatives: Object.entries(adjustedScores)
+                .filter(([m]) => m !== model)
+                .map(([m, score]) => ({ model: m, score }))
+                .sort((a, b) => b.score - a.score),
+            inferenceTimeUs,
+            costMultiplier: MODEL_CAPABILITIES[model].costMultiplier,
+            routedBy,
+            ...(neuralBackend ? { neuralBackend } : {}),
+            provider: exec.provider,
+            ...(exec.openrouterModel ? { openrouterModel: exec.openrouterModel } : {}),
+            // ADR-149: surface the concrete neural pick when present. Prefer the
+            // explicit OpenRouter alt (resolveExecutionProvider) for execution,
+            // but expose the model id the neural backend chose so observers and
+            // consumers can see the cost-optimal decision.
+            ...(neuralModelId ? { modelId: neuralModelId } : {}),
+        };
+        // Track decision (in-memory bandit state)
+        this.trackDecision(task, result);
+        // ADR-148 — opt-in DRACO-shaped trajectory collection
+        if (process.env.SWARMDO_ROUTER_TRAJECTORY === '1') {
+            try {
+                const { recordDecision } = await loadTrajectoryRecorder();
+                recordDecision({
+                    task, embedding, complexity: complexity.score,
+                    model, confidence, uncertainty, routedBy,
+                    neuralBackend, abPair,
+                    provider: exec.provider,
+                    openrouterModel: exec.openrouterModel,
+                    ensembleDisagreement: neuralEnsembleDisagreement,
+                });
+            }
+            catch {
+                // Silent — trajectory recording must never break routing.
+            }
+        }
+        return result;
+    }
+    /**
+     * Analyze task complexity
+     */
+    analyzeComplexity(task, embedding) {
+        const taskLower = task.toLowerCase();
+        const words = taskLower.split(/\s+/);
+        // Find complexity indicators
+        const indicators = {
+            high: COMPLEXITY_INDICATORS.high.filter(ind => taskLower.includes(ind)),
+            medium: COMPLEXITY_INDICATORS.medium.filter(ind => taskLower.includes(ind)),
+            low: COMPLEXITY_INDICATORS.low.filter(ind => taskLower.includes(ind)),
+        };
+        // Compute feature scores
+        const lexicalComplexity = this.computeLexicalComplexity(task);
+        const semanticDepth = this.computeSemanticDepth(indicators, embedding);
+        const taskScope = this.computeTaskScope(task, words);
+        const uncertaintyLevel = this.computeUncertaintyLevel(task);
+        // Weighted combination
+        const score = Math.min(1, Math.max(0, lexicalComplexity * 0.2 +
+            semanticDepth * 0.35 +
+            taskScope * 0.25 +
+            uncertaintyLevel * 0.2));
+        return {
+            score,
+            indicators,
+            features: {
+                lexicalComplexity,
+                semanticDepth,
+                taskScope,
+                uncertaintyLevel,
+            },
+        };
+    }
+    /**
+     * Compute lexical complexity from text features
+     */
+    computeLexicalComplexity(task) {
+        const words = task.split(/\s+/);
+        const avgWordLength = words.reduce((sum, w) => sum + w.length, 0) / Math.max(1, words.length);
+        const sentenceLength = words.length;
+        // Normalize: longer sentences with longer words = more complex
+        const lengthScore = Math.min(1, sentenceLength / 50);
+        const wordScore = Math.min(1, (avgWordLength - 3) / 7); // 3-10 char words
+        return lengthScore * 0.4 + wordScore * 0.6;
+    }
+    /**
+     * Compute semantic depth from indicators and embedding
+     */
+    computeSemanticDepth(indicators, embedding) {
+        // Weight by indicator presence
+        const highWeight = indicators.high.length * 0.3;
+        const mediumWeight = indicators.medium.length * 0.15;
+        const lowWeight = indicators.low.length * -0.1;
+        let baseScore = Math.min(1, Math.max(0, 0.3 + highWeight + mediumWeight + lowWeight));
+        // Boost with embedding variance if available
+        if (embedding && embedding.length > 0) {
+            const mean = embedding.reduce((a, b) => a + b, 0) / embedding.length;
+            const variance = embedding.reduce((sum, v) => sum + Math.pow(v - mean, 2), 0) / embedding.length;
+            // Higher variance suggests more nuanced semantics
+            baseScore = baseScore * 0.7 + Math.min(1, variance * 10) * 0.3;
+        }
+        return baseScore;
+    }
+    /**
+     * Compute task scope from content analysis
+     */
+    computeTaskScope(task, words) {
+        // Multi-file indicators
+        const multiFilePatterns = [
+            /multiple files?/i, /across.*modules?/i, /refactor.*codebase/i,
+            /all.*files/i, /entire.*project/i, /system.*wide/i,
+        ];
+        const hasMultiFile = multiFilePatterns.some(p => p.test(task)) ? 0.4 : 0;
+        // Code generation indicators
+        const codeGenPatterns = [
+            /implement/i, /create.*feature/i, /build.*system/i,
+            /design.*api/i, /write.*tests/i, /add.*functionality/i,
+        ];
+        const hasCodeGen = codeGenPatterns.some(p => p.test(task)) ? 0.3 : 0;
+        // Word count contribution
+        const wordCountScore = Math.min(0.3, words.length / 100);
+        return hasMultiFile + hasCodeGen + wordCountScore;
+    }
+    /**
+     * Compute uncertainty level from task phrasing
+     */
+    computeUncertaintyLevel(task) {
+        const uncertainPatterns = [
+            /not sure/i, /might/i, /maybe/i, /possibly/i, /investigate/i,
+            /figure out/i, /unclear/i, /unknown/i, /debug/i, /strange/i,
+            /weird/i, /issue/i, /problem/i, /error/i, /bug/i,
+        ];
+        const matchCount = uncertainPatterns.filter(p => p.test(task)).length;
+        return Math.min(1, matchCount * 0.2);
+    }
+    /**
+     * Compute scores for each model
+     */
+    computeModelScores(complexity) {
+        const { score } = complexity;
+        // Base scoring: inverse relationship with complexity
+        // Low complexity → haiku scores high
+        // High complexity → opus scores high
+        return {
+            haiku: Math.max(0, 1 - score * 2), // Drops off quickly as complexity rises
+            sonnet: 1 - Math.abs(score - 0.5) * 2, // Peaks at medium complexity
+            opus: Math.min(1, score * 1.5), // Rises with complexity
+            inherit: 0.1, // Low baseline unless explicitly needed
+        };
+    }
+    /**
+     * Apply circuit breaker adjustments
+     */
+    applyCircuitBreaker(scores) {
+        if (!this.config.enableCircuitBreaker) {
+            return scores;
+        }
+        const adjusted = { ...scores };
+        for (const model of Object.keys(adjusted)) {
+            if (this.consecutiveFailures[model] >= this.config.circuitBreakerThreshold) {
+                // Circuit is open - heavily penalize this model
+                adjusted[model] *= 0.1;
+            }
+            else if (this.consecutiveFailures[model] > 0) {
+                // Partial penalty for recent failures
+                adjusted[model] *= 1 - (this.consecutiveFailures[model] / this.config.circuitBreakerThreshold) * 0.5;
+            }
+        }
+        return adjusted;
+    }
+    /**
+     * Select the best model from scores. Uses Thompson sampling (#1772):
+     * each model's deterministic complexity score is multiplied by a draw
+     * θ_m ~ Beta(α_m, β_m) from its bandit prior. Models with strong empirical
+     * track records get sampled higher; models with poor outcomes get sampled
+     * lower; the system auto-corrects against tier overuse without manual
+     * threshold tuning. Beta(1,1) = uniform on cold start so behavior matches
+     * the prior deterministic router until outcomes accumulate.
+     */
+    selectModel(scores, complexityScore, 
+    /**
+     * Optional neural prior (ADR-148 hybrid math). When supplied, each
+     * candidate's Beta(α, β) prior is perturbed by `weight` pseudo-counts of
+     * the neural's predicted quality before sampling. Cold start → neural
+     * dominates; many real outcomes → bandit dominates. The persistent
+     * bandit state is NOT modified — this is a per-call posterior bump only.
+     */
+    neuralPrior) {
+        // Thompson sampling: combine deterministic score with bandit posterior,
+        // keyed by complexity bucket (ADR-142) so learning is task-type-local.
+        const bucketed = this.state.priors ?? defaultBucketedPriors();
+        const priors = bucketed[complexityBucket(complexityScore)] ?? defaultBanditPriors();
+        // Apply the optional neural prior: Beta(α + q*w, β + (1-q)*w). Per-call,
+        // does not persist. Clamp `q` into [0, 1] so a bogus backend reading
+        // cannot push the prior into invalid territory.
+        const bump = (a, b, q, w) => {
+            if (q === undefined || w <= 0)
+                return { alpha: a, beta: b };
+            const clamped = Math.min(1, Math.max(0, q));
+            return { alpha: a + clamped * w, beta: b + (1 - clamped) * w };
+        };
+        const w = neuralPrior?.weight ?? 0;
+        const ph = bump(priors.haiku.alpha, priors.haiku.beta, neuralPrior?.qualities.haiku, w);
+        const ps = bump(priors.sonnet.alpha, priors.sonnet.beta, neuralPrior?.qualities.sonnet, w);
+        const po = bump(priors.opus.alpha, priors.opus.beta, neuralPrior?.qualities.opus, w);
+        const sampledScores = {
+            haiku: scores.haiku * sampleBeta(ph.alpha, ph.beta),
+            sonnet: scores.sonnet * sampleBeta(ps.alpha, ps.beta),
+            opus: scores.opus * sampleBeta(po.alpha, po.beta),
+            inherit: scores.inherit, // not bandit-controlled
+        };
+        // Get sorted models by sampled score (drops 'inherit' from selection)
+        const sorted = Object.entries(sampledScores)
+            .filter(([m]) => m !== 'inherit')
+            .sort((a, b) => b[1] - a[1]);
+        const [bestModel, bestScore] = sorted[0];
+        const [, secondScore] = sorted[1] || ['sonnet', 0];
+        // Confidence is how much better the best is vs second
+        const confidence = bestScore > 0 ? Math.min(1, bestScore / (bestScore + secondScore + 0.01)) : 0.5;
+        // Uncertainty based on score spread and complexity
+        const scoreSpread = bestScore - secondScore;
+        const uncertainty = Math.max(0, 1 - scoreSpread - confidence * 0.5);
+        // Escalate if uncertainty is too high.
+        //
+        // #2250 — `uncertainty` here is structurally ~0.6-0.7 for low-complexity
+        // tasks (formula: `1 - scoreSpread - confidence*0.5`, where `scoreSpread`
+        // is a raw 0-1 difference between bandit-sampled scores that rarely
+        // exceeds 0.1). With `maxUncertainty = 0.15` the gate fires on
+        // ~every trivial route, promoting `sonnet→opus` and `haiku→sonnet`
+        // even when the Thompson sampler has *already* suppressed the higher
+        // tier (e.g. opus `Beta(3.8, 17.2)`, mean ≈ 0.18). The learned
+        // suppression is computed and then discarded one line later.
+        //
+        // Guard: skip the escalation when EITHER (a) the bandit has confidently
+        // learned the escalation target performs WORSE than the selected model,
+        // OR (b) the bandit has a confident, decent posterior on the selected
+        // model — i.e. the Thompson sampler picked this tier on real evidence,
+        // not a coin flip. Cold-start priors (Beta(1,1), α+β=2, mean=0.5) fail
+        // both checks, so unlearned routers still escalate as before.
+        let model = bestModel;
+        if (uncertainty > this.config.maxUncertainty && bestModel !== 'opus') {
+            const escalateTo = bestModel === 'haiku' ? 'sonnet' : 'opus';
+            const selectedMean = priorMean(priors[bestModel]);
+            const targetMean = priorMean(priors[escalateTo]);
+            const targetWorse = targetMean < selectedMean - 0.10;
+            // Treat the selected model as trusted once the bandit has accumulated
+            // ~5 effective observations AND its mean is at least 0.45 (neutral-or-
+            // better). Both thresholds chosen to keep cold-start behavior identical
+            // while honoring any non-trivial learning.
+            const selectedSamples = priors[bestModel].alpha + priors[bestModel].beta;
+            const selectedTrusted = selectedSamples >= 5 && selectedMean >= 0.45;
+            // ADR-148 — additional trust path: when the neural prior agrees with
+            // the bandit's pick (i.e. neuralPrior.qualities[bestModel] is the
+            // highest), the neural backend's signal is treated as a vote of
+            // confidence that lets us skip escalation. Without this, cold-start
+            // installations stay stuck in the structurally-high-uncertainty
+            // regime (#2250) and the neural's clear preference never reaches
+            // the final pick. The neural prior's top is computed as the max
+            // quality over the supplied candidates.
+            let neuralEndorsesPick = false;
+            if (neuralPrior) {
+                let bestNeuralQ = -1;
+                let neuralTop = null;
+                for (const [m, q] of Object.entries(neuralPrior.qualities)) {
+                    if (q > bestNeuralQ) {
+                        bestNeuralQ = q;
+                        neuralTop = m;
+                    }
+                }
+                if (neuralTop === bestModel)
+                    neuralEndorsesPick = true;
+            }
+            if (!targetWorse && !selectedTrusted && !neuralEndorsesPick) {
+                model = escalateTo;
+            }
+        }
+        return { model, confidence, uncertainty };
+    }
+    /**
+     * Build human-readable reasoning
+     */
+    buildReasoning(model, complexity, confidence) {
+        const parts = [];
+        parts.push(`Complexity: ${(complexity.score * 100).toFixed(0)}%`);
+        if (complexity.indicators.high.length > 0) {
+            parts.push(`High-complexity indicators: ${complexity.indicators.high.join(', ')}`);
+        }
+        parts.push(`Confidence: ${(confidence * 100).toFixed(0)}%`);
+        parts.push(`Model: ${model} - ${MODEL_CAPABILITIES[model].description}`);
+        if (this.config.enableCostOptimization) {
+            parts.push(`Cost: ${MODEL_CAPABILITIES[model].costMultiplier}x baseline`);
+        }
+        return parts.join(' | ');
+    }
+    /**
+     * Track routing decision for learning
+     */
+    trackDecision(task, result) {
+        this.decisionCount++;
+        this.state.totalDecisions++;
+        this.state.modelDistribution[result.model] =
+            (this.state.modelDistribution[result.model] || 0) + 1;
+        // ADR-148 — operational counters for hooks_intelligence_stats.
+        this.routedByCounts[result.routedBy]++;
+        if (result.neuralBackend)
+            this.neuralBackendCounts[result.neuralBackend]++;
+        // Update running averages
+        const n = this.state.totalDecisions;
+        this.state.avgComplexity =
+            (this.state.avgComplexity * (n - 1) + result.complexity) / n;
+        this.state.avgConfidence =
+            (this.state.avgConfidence * (n - 1) + result.confidence) / n;
+        // Auto-save periodically
+        if (this.decisionCount % this.config.autoSaveInterval === 0) {
+            this.saveState();
+        }
+    }
+    /**
+     * Record outcome for learning
+     */
+    recordOutcome(task, model, outcome) {
+        // Update circuit breaker state
+        if (outcome === 'failure') {
+            this.consecutiveFailures[model]++;
+        }
+        else {
+            this.consecutiveFailures[model] = 0;
+        }
+        // Re-derive this task's complexity bucket from the task string (the MCP
+        // outcome payload carries no complexity), using the SAME analyzeComplexity
+        // path route() uses so record-time and select-time buckets match.
+        const taskScore = this.analyzeComplexity(task).score;
+        const bucket = complexityBucket(taskScore);
+        // Track in history (record THIS task's score, not the running average)
+        this.state.learningHistory.push({
+            task: task.slice(0, 100),
+            model,
+            complexity: taskScore,
+            outcome,
+            timestamp: new Date().toISOString(),
+        });
+        // Keep history bounded
+        if (this.state.learningHistory.length > 100) {
+            this.state.learningHistory = this.state.learningHistory.slice(-100);
+        }
+        if (outcome === 'failure') {
+            this.state.circuitBreakerTrips++;
+        }
+        // Thompson sampling update (#1772): cost-adjusted Bernoulli reward.
+        // Haiku-success > Sonnet-success > Opus-success (Opus on simple tasks
+        // is wasteful even when correct). Failure/escalation always β++.
+        if (!this.state.priors)
+            this.state.priors = defaultBucketedPriors();
+        const bp = this.state.priors[bucket] ?? (this.state.priors[bucket] = defaultBanditPriors());
+        const reward = BANDIT_REWARDS[model]?.[outcome] ?? 0.5;
+        bp[model].alpha += reward;
+        bp[model].beta += 1 - reward;
+        this.saveState();
+    }
+    /**
+     * ADR-149 — record an outcome keyed by the CONCRETE model id (e.g.
+     * 'inclusionai/ling-2.6-flash') rather than the tier label. Updates the
+     * shadow `priorsById` state without affecting `priors` (tier priors).
+     *
+     * This is the per-model learning signal the bandit needs to eventually
+     * distinguish GPT-4.1 from Sonnet within the 'sonnet' tier. Selection
+     * currently still uses tier priors; this state accumulates so a future
+     * refactor can switch the selector over once there's enough data.
+     *
+     * Cost-adjusted reward semantics: cheap models get the highest reward on
+     * success (their successes are most cost-efficient). We map modelId to
+     * its closest tier for the reward table — the routing math doesn't have
+     * a per-modelId reward configuration yet.
+     */
+    recordOutcomeByModelId(task, modelId, outcome) {
+        if (!modelId || typeof modelId !== 'string')
+            return;
+        const taskScore = this.analyzeComplexity(task).score;
+        const bucket = complexityBucket(taskScore);
+        if (!this.state.priorsById)
+            this.state.priorsById = defaultBucketedPriorsById();
+        let perBucket = this.state.priorsById[bucket];
+        if (!perBucket) {
+            perBucket = {};
+            this.state.priorsById[bucket] = perBucket;
+        }
+        if (!perBucket[modelId])
+            perBucket[modelId] = { alpha: 1, beta: 1 };
+        // Reward proxy: derive a tier-equivalent for cost weighting. Substring
+        // match keeps it cheap and accurate for the candidates in the registry.
+        const id = modelId.toLowerCase();
+        const tierProxy = id.includes('haiku') || id.includes('ling-') || id.includes('flash-lite')
+            || id.includes('nemotron-nano') || id.includes('ministral')
+            || id.includes('llama-3.2-3b') || id.includes('llama-3.1-8b')
+            ? 'haiku'
+            : id.includes('opus') ? 'opus' : 'sonnet';
+        const reward = BANDIT_REWARDS[tierProxy]?.[outcome] ?? 0.5;
+        perBucket[modelId].alpha += reward;
+        perBucket[modelId].beta += 1 - reward;
+        // Bump the schema version on first per-modelId write so downstream
+        // tooling can see v3 was reached (the version field stays at 3 once set).
+        if ((this.state.version ?? 0) < 3)
+            this.state.version = 3;
+        this.saveState();
+    }
+    /**
+     * Get router statistics
+     */
+    getStats() {
+        return {
+            totalDecisions: this.state.totalDecisions,
+            modelDistribution: { ...this.state.modelDistribution },
+            avgComplexity: this.state.avgComplexity,
+            avgConfidence: this.state.avgConfidence,
+            circuitBreakerTrips: this.state.circuitBreakerTrips,
+            consecutiveFailures: { ...this.consecutiveFailures },
+            routedByCounts: { ...this.routedByCounts },
+            neuralBackendCounts: { ...this.neuralBackendCounts },
+            ab: {
+                comparisons: this.abComparisons,
+                disagreements: this.abDisagreements,
+                disagreementRate: this.abComparisons > 0 ? this.abDisagreements / this.abComparisons : 0,
+            },
+            stateVersion: this.state.version ?? 2,
+            ...(this.state.priorsById ? { priorsById: this.state.priorsById } : {}),
+        };
+    }
+    /**
+     * Load state from disk
+     */
+    loadState() {
+        const defaultState = {
+            totalDecisions: 0,
+            modelDistribution: { haiku: 0, sonnet: 0, opus: 0, inherit: 0 },
+            avgComplexity: 0.5,
+            avgConfidence: 0.8,
+            circuitBreakerTrips: 0,
+            lastUpdated: new Date().toISOString(),
+            learningHistory: [],
+            version: 2,
+            priors: defaultBucketedPriors(),
+        };
+        try {
+            const fullPath = join(process.cwd(), this.config.statePath);
+            if (existsSync(fullPath)) {
+                const data = readFileSync(fullPath, 'utf-8');
+                const loaded = JSON.parse(data);
+                // ADR-142: forward-migrate priors of ANY layout (missing / flat v1 /
+                // already-bucketed) to the bucketed shape without data loss or throwing.
+                loaded.priors = migratePriors(loaded.priors);
+                loaded.version = 2;
+                return { ...defaultState, ...loaded };
+            }
+        }
+        catch {
+            // Ignore load errors
+        }
+        return defaultState;
+    }
+    /**
+     * Save state to disk
+     */
+    saveState() {
+        try {
+            const fullPath = join(process.cwd(), this.config.statePath);
+            const dir = dirname(fullPath);
+            if (!existsSync(dir)) {
+                mkdirSync(dir, { recursive: true });
+            }
+            this.state.lastUpdated = new Date().toISOString();
+            writeFileSync(fullPath, JSON.stringify(this.state, null, 2));
+        }
+        catch {
+            // Ignore save errors in non-critical scenarios
+        }
+    }
+    /**
+     * Reset router state
+     */
+    reset() {
+        this.state = {
+            totalDecisions: 0,
+            modelDistribution: { haiku: 0, sonnet: 0, opus: 0, inherit: 0 },
+            avgComplexity: 0.5,
+            avgConfidence: 0.8,
+            circuitBreakerTrips: 0,
+            lastUpdated: new Date().toISOString(),
+            learningHistory: [],
+            version: 2,
+            priors: defaultBucketedPriors(),
+        };
+        this.consecutiveFailures = { haiku: 0, sonnet: 0, opus: 0, inherit: 0 };
+        this.decisionCount = 0;
+        this.saveState();
+    }
+    /**
+     * Public read-only accessor for the bandit priors. Useful for tests,
+     * dashboards, and the pending hooks_intelligence_stats integration that
+     * surfaces convergence in the dashboard. Returns a copy.
+     */
+    getBanditPriors(bucket = 'med') {
+        const bucketed = this.state.priors ?? defaultBucketedPriors();
+        const p = bucketed[bucket] ?? defaultBanditPriors();
+        return {
+            haiku: { ...p.haiku },
+            sonnet: { ...p.sonnet },
+            opus: { ...p.opus },
+            inherit: { ...p.inherit },
+        };
+    }
+    /** All bucketed priors (copy) — for dashboards/tests. */
+    getBucketedPriors() {
+        const b = this.state.priors ?? defaultBucketedPriors();
+        return {
+            low: clonePriors(b.low ?? defaultBanditPriors()),
+            med: clonePriors(b.med ?? defaultBanditPriors()),
+            high: clonePriors(b.high ?? defaultBanditPriors()),
+        };
+    }
+}
+// ============================================================================
+// Singleton & Factory Functions
+// ============================================================================
+let modelRouterInstance = null;
+/**
+ * Get or create the singleton ModelRouter instance
+ */
+export function getModelRouter(config) {
+    if (!modelRouterInstance) {
+        modelRouterInstance = new ModelRouter(config);
+    }
+    return modelRouterInstance;
+}
+/**
+ * Reset the singleton instance
+ */
+export function resetModelRouter() {
+    modelRouterInstance = null;
+}
+/**
+ * Create a new ModelRouter instance (non-singleton)
+ */
+export function createModelRouter(config) {
+    return new ModelRouter(config);
+}
+// ============================================================================
+// Convenience Functions
+// ============================================================================
+/**
+ * Quick route function for common use case
+ */
+export async function routeToModel(task) {
+    const router = getModelRouter();
+    const result = await router.route(task);
+    return result.model;
+}
+/**
+ * Route with full result
+ */
+export async function routeToModelFull(task, embedding) {
+    const router = getModelRouter();
+    return router.route(task, embedding);
+}
+/**
+ * Analyze task complexity without routing
+ */
+export function analyzeTaskComplexity(task) {
+    const router = getModelRouter();
+    return router.analyzeComplexity(task, undefined);
+}
+/**
+ * Get model router statistics
+ */
+export function getModelRouterStats() {
+    const router = getModelRouter();
+    return router.getStats();
+}
+/**
+ * Record routing outcome for learning
+ */
+export function recordModelOutcome(task, model, outcome) {
+    const router = getModelRouter();
+    router.recordOutcome(task, model, outcome);
+}
+/**
+ * ADR-149 — record an outcome keyed by the concrete model id rather than
+ * the tier label. Updates the shadow `priorsById` state. Selection logic
+ * still uses tier priors; this data accumulates for a future per-modelId
+ * selector refactor.
+ *
+ * Safe to call alongside `recordModelOutcome` — they update independent
+ * state slices so double-counting is impossible.
+ */
+export function recordModelOutcomeByModelId(task, modelId, outcome) {
+    const router = getModelRouter();
+    router.recordOutcomeByModelId(task, modelId, outcome);
+}
+/**
+ * ADR-149 iter 14 — read-only access to the per-modelId Beta priors. The
+ * neural-router consumes this to apply per-model Thompson sampling on top
+ * of its predicted-quality vector when SWARMDO_ROUTER_BANDIT_PER_MODEL=1.
+ * Returns the legacy bucketed priors (`priorsById[bucket][modelId]`) when
+ * present, else null.
+ */
+export function getModelRouterPriorsById() {
+    const router = getModelRouter();
+    const stats = router.getStats();
+    return stats.priorsById ?? null;
+}
+/**
+ * Re-export the complexity-bucket helper so the neural-router (which gets
+ * the task text via the route() call) can map a complexity score to the
+ * matching bandit bucket.
+ */
+export { complexityBucket };
+//# sourceMappingURL=model-router.js.map