agentic-qe 3.12.1 → 3.12.2

package/dist/shared/llm/spend-ledger.d.ts

@@ -0,0 +1,64 @@
+/**
+ * Agentic QE v3 — LLM Spend Ledger (ADR-123)
+ *
+ * Issue #557: AQE declared budget caps (`maxCostPerHour/Day`,
+ * `COST_LIMIT_EXCEEDED`, `CostTracker.wouldExceedLimit`) but enforced none of
+ * them, and `CostTracker` is per-process in-memory — so a fleet of N agent
+ * processes each sees $0 cumulative spend and no cap can hold across them.
+ *
+ * This module persists every LLM charge to the unified `.agentic-qe/memory.db`
+ * (ADR: one DB, one schema — the `llm_spend` table is additive, never a new
+ * file) so a budget check reads the *shared* rolling-window spend across every
+ * process. When the unified DB is unavailable (e.g. a standalone eval script
+ * with no kernel), it degrades to an in-memory ledger that still enforces
+ * within the current process — never worse than today's behavior.
+ *
+ * better-sqlite3 is fully synchronous, so record/query are sync; multiple
+ * connections (one per process) to a WAL-mode DB is exactly the supported
+ * cross-process pattern.
+ */
+import type { Database as DatabaseType } from 'better-sqlite3';
+import type { CostSource } from './interfaces';
+/** A single persisted charge. */
+export interface SpendEntry {
+    provider: string;
+    model: string;
+    costUsd: number;
+    costSource: CostSource;
+    promptTokens: number;
+    completionTokens: number;
+    requestId: string;
+}
+/** Cross-process spend ledger. */
+export interface SpendLedger {
+    /** Persist a charge. Never throws for the caller — failures degrade silently. */
+    record(entry: SpendEntry): void;
+    /** Total USD charged within the last `windowMs` (rolling window ending now). */
+    spentSince(windowMs: number, nowMs?: number): number;
+}
+/**
+ * SQLite-backed ledger. Cross-process because every process opens its own
+ * connection to the same WAL-mode `memory.db`.
+ */
+export declare class SqliteSpendLedger implements SpendLedger {
+    private readonly db;
+    constructor(db: DatabaseType);
+    record(entry: SpendEntry): void;
+    spentSince(windowMs: number, nowMs?: number): number;
+}
+/**
+ * In-process fallback ledger. Used when no unified DB is available. Enforces
+ * within the current process only — no worse than the pre-ADR-123 behavior.
+ */
+export declare class InMemorySpendLedger implements SpendLedger {
+    private readonly entries;
+    record(entry: SpendEntry): void;
+    spentSince(windowMs: number, nowMs?: number): number;
+}
+/**
+ * Build the default ledger. Prefers the unified `memory.db` (cross-process);
+ * falls back to an in-memory ledger if the kernel DB can't be reached. Uses a
+ * dynamic import so the LLM layer stays statically decoupled from the kernel.
+ */
+export declare function createDefaultSpendLedger(): Promise<SpendLedger>;
+//# sourceMappingURL=spend-ledger.d.ts.map

package/dist/shared/llm/spend-ledger.js

@@ -0,0 +1,115 @@
+/**
+ * Agentic QE v3 — LLM Spend Ledger (ADR-123)
+ *
+ * Issue #557: AQE declared budget caps (`maxCostPerHour/Day`,
+ * `COST_LIMIT_EXCEEDED`, `CostTracker.wouldExceedLimit`) but enforced none of
+ * them, and `CostTracker` is per-process in-memory — so a fleet of N agent
+ * processes each sees $0 cumulative spend and no cap can hold across them.
+ *
+ * This module persists every LLM charge to the unified `.agentic-qe/memory.db`
+ * (ADR: one DB, one schema — the `llm_spend` table is additive, never a new
+ * file) so a budget check reads the *shared* rolling-window spend across every
+ * process. When the unified DB is unavailable (e.g. a standalone eval script
+ * with no kernel), it degrades to an in-memory ledger that still enforces
+ * within the current process — never worse than today's behavior.
+ *
+ * better-sqlite3 is fully synchronous, so record/query are sync; multiple
+ * connections (one per process) to a WAL-mode DB is exactly the supported
+ * cross-process pattern.
+ */
+const CREATE_TABLE_SQL = `
+CREATE TABLE IF NOT EXISTS llm_spend (
+  id INTEGER PRIMARY KEY AUTOINCREMENT,
+  ts INTEGER NOT NULL,
+  provider TEXT NOT NULL,
+  model TEXT NOT NULL,
+  cost_usd REAL NOT NULL,
+  cost_source TEXT NOT NULL,
+  prompt_tokens INTEGER NOT NULL DEFAULT 0,
+  completion_tokens INTEGER NOT NULL DEFAULT 0,
+  request_id TEXT
+);
+CREATE INDEX IF NOT EXISTS idx_llm_spend_ts ON llm_spend(ts);
+`;
+/**
+ * SQLite-backed ledger. Cross-process because every process opens its own
+ * connection to the same WAL-mode `memory.db`.
+ */
+export class SqliteSpendLedger {
+    db;
+    constructor(db) {
+        this.db = db;
+        // Additive schema only — CREATE TABLE/INDEX IF NOT EXISTS never drops data.
+        this.db.exec(CREATE_TABLE_SQL);
+    }
+    record(entry) {
+        try {
+            this.db
+                .prepare(`INSERT INTO llm_spend
+             (ts, provider, model, cost_usd, cost_source, prompt_tokens, completion_tokens, request_id)
+           VALUES (@ts, @provider, @model, @cost_usd, @cost_source, @prompt_tokens, @completion_tokens, @request_id)`)
+                .run({
+                ts: Date.now(),
+                provider: entry.provider,
+                model: entry.model,
+                cost_usd: entry.costUsd,
+                cost_source: entry.costSource,
+                prompt_tokens: entry.promptTokens,
+                completion_tokens: entry.completionTokens,
+                request_id: entry.requestId,
+            });
+        }
+        catch {
+            // A ledger write must never break a generation call.
+        }
+    }
+    spentSince(windowMs, nowMs = Date.now()) {
+        try {
+            const row = this.db
+                .prepare(`SELECT COALESCE(SUM(cost_usd), 0) AS total FROM llm_spend WHERE ts >= ?`)
+                .get(nowMs - windowMs);
+            return row?.total ?? 0;
+        }
+        catch {
+            return 0;
+        }
+    }
+}
+/**
+ * In-process fallback ledger. Used when no unified DB is available. Enforces
+ * within the current process only — no worse than the pre-ADR-123 behavior.
+ */
+export class InMemorySpendLedger {
+    entries = [];
+    record(entry) {
+        this.entries.push({ ts: Date.now(), costUsd: entry.costUsd });
+    }
+    spentSince(windowMs, nowMs = Date.now()) {
+        const cutoff = nowMs - windowMs;
+        let total = 0;
+        for (const e of this.entries) {
+            if (e.ts >= cutoff)
+                total += e.costUsd;
+        }
+        return total;
+    }
+}
+/**
+ * Build the default ledger. Prefers the unified `memory.db` (cross-process);
+ * falls back to an in-memory ledger if the kernel DB can't be reached. Uses a
+ * dynamic import so the LLM layer stays statically decoupled from the kernel.
+ */
+export async function createDefaultSpendLedger() {
+    try {
+        const mod = await import('../../kernel/unified-memory.js');
+        const mem = mod.getUnifiedMemory();
+        if (!mem.isInitialized()) {
+            await mem.initialize();
+        }
+        return new SqliteSpendLedger(mem.getDatabase());
+    }
+    catch {
+        return new InMemorySpendLedger();
+    }
+}
+//# sourceMappingURL=spend-ledger.js.map

package/dist/shared/llm/translation/message-formatter.js

@@ -15,12 +15,14 @@
  */
 const SYSTEM_PROMPT_STRATEGIES = {
     claude: 'native', // Anthropic supports separate system param
+    'claude-code': 'native', // ADR-123: passed via --append-system-prompt
     openai: 'native', // OpenAI supports system role messages
     ollama: 'native', // Ollama supports system role
     openrouter: 'native', // OpenRouter passes through to underlying provider
     gemini: 'native', // Gemini supports systemInstruction
     'azure-openai': 'native', // Azure OpenAI same as OpenAI
     bedrock: 'native', // Bedrock supports system for Claude
+    cognitum: 'native', // ADR-123: OpenAI-compatible system role
     onnx: 'first-message', // ONNX models typically don't have system prompt support
 };
 /**

package/dist/shared/llm/translation/prompt-translator.js

@@ -37,10 +37,12 @@ export function detectMessageFormat(messages) {
 export function getTargetFormat(provider) {
     switch (provider) {
         case 'claude':
+        case 'claude-code': // ADR-123: Anthropic-native (runs via `claude -p`)
         case 'bedrock':
             return 'anthropic';
         case 'openai':
         case 'azure-openai':
+        case 'cognitum': // ADR-123: OpenAI-compatible gateway
             return 'openai';
         case 'gemini':
             return 'gemini';

package/dist/shared/llm/translation/tool-translator.js

@@ -16,6 +16,7 @@
 export function getToolSchemaFormat(provider) {
     switch (provider) {
         case 'claude':
+        case 'claude-code': // ADR-123: Anthropic-native
         case 'bedrock':
             return 'anthropic';
         case 'openai':
@@ -23,6 +24,7 @@ export function getToolSchemaFormat(provider) {
         case 'openrouter':
         case 'ollama':
         case 'onnx':
+        case 'cognitum': // ADR-123: OpenAI-compatible
             return 'openai';
         case 'gemini':
             return 'gemini';

package/dist/validation/index.d.ts

@@ -13,6 +13,7 @@
 export { SwarmSkillValidator, createSwarmSkillValidator, DEFAULT_SWARM_VALIDATION_CONFIG, P0_SKILLS, DEFAULT_VALIDATION_MODELS, } from './swarm-skill-validator.js';
 export type { SwarmValidationConfig, SwarmTopology, SwarmValidationResult, SwarmValidationSummary, SkillValidatorFn, } from './swarm-skill-validator.js';
 export { ParallelEvalRunner, createParallelEvalRunner, DEFAULT_PARALLEL_EVAL_CONFIG, MockLLMExecutor, } from './parallel-eval-runner.js';
+export { ProviderLLMExecutor, resolveEvalExecutor, type EvalExecutorResolution, } from './provider-llm-executor.js';
 export type { ParallelEvalConfig, EvalTestCase, EvalTestCaseInput, EvalTestCaseExpectedOutput, EvalTestCaseValidation, EvalSuite, TestCaseTask, ParallelEvalResult, WorkerMessage, WorkerProgress, EvalProgress, LLMExecutor, } from './parallel-eval-runner.js';
 export { CommandEvalRunner, createCommandEvalRunner, evalJsonPath, severityAtLeast, isCommandEvalSuite, validateCommandResult, } from './command-eval-runner.js';
 export type { CommandEvalSuite, CommandEvalTestCase, CommandEvalExpected, CommandEvalResult, CommandEvalTestResult, CommandEvalRunnerOptions, CommandRunner, Severity, TestCasePriority, } from './command-eval-runner.js';

package/dist/validation/index.js

@@ -24,6 +24,8 @@ P0_SKILLS, DEFAULT_VALIDATION_MODELS, } from './swarm-skill-validator.js';
 // Parallel Eval Runner (ADR-056 Phase 5)
 // ============================================================================
 export { ParallelEvalRunner, createParallelEvalRunner, DEFAULT_PARALLEL_EVAL_CONFIG, MockLLMExecutor, } from './parallel-eval-runner.js';
+// Issue #557 follow-up: real (non-mock) eval executor.
+export { ProviderLLMExecutor, resolveEvalExecutor, } from './provider-llm-executor.js';
 // ============================================================================
 // Command Eval Runner (ADR-091: qe-browser and other shell-based skills)
 // ============================================================================

package/dist/validation/provider-llm-executor.d.ts

@@ -0,0 +1,45 @@
+/**
+ * Agentic QE v3 — Real LLM executor for the eval runner (issue #557 follow-up).
+ *
+ * `aqe eval run` historically defaulted to MockLLMExecutor (canned, keyword-
+ * matched responses), so LLM-mode eval suites scored fabricated output. This
+ * executor makes REAL model calls through the shared LLM layer, so eval
+ * inherits the ADR-123 guardrails automatically: enforced budget caps, the
+ * `AQE_LLM_PROVIDER` selector (including the claude-code subscription), the
+ * billing notice, and the cross-process spend ledger.
+ */
+import type { LLMExecutor } from './parallel-eval-runner';
+import { ProviderManager } from '../shared/llm/provider-manager';
+import type { LLMProviderType } from '../shared/llm/interfaces';
+/** Real executor: delegates to ProviderManager.generate (budget-gated). */
+export declare class ProviderLLMExecutor implements LLMExecutor {
+    private readonly manager;
+    readonly providerType: LLMProviderType;
+    constructor(manager: ProviderManager, providerType: LLMProviderType);
+    execute(prompt: string, model: string, options?: {
+        timeout?: number;
+    }): Promise<{
+        output: string;
+        tokensUsed: number;
+        durationMs: number;
+    }>;
+}
+export interface EvalExecutorResolution {
+    /** The real executor, or undefined when no provider is configured. */
+    executor?: ProviderLLMExecutor;
+    /** The provider chosen (for display), when resolved. */
+    providerType?: LLMProviderType;
+    /** Billing mode of the chosen provider (for the pre-run notice). */
+    billingMode?: string;
+    /** Manager to dispose when done (caller owns lifecycle). */
+    manager?: ProviderManager;
+    /** Why no executor could be built (for a clear CLI error). */
+    reason?: string;
+}
+/**
+ * Resolve a real eval executor from the current environment/config. Honors
+ * `AQE_LLM_PROVIDER`. Returns `{ reason }` (no executor) when nothing is
+ * configured, so the CLI can hard-error instead of silently mocking.
+ */
+export declare function resolveEvalExecutor(env?: NodeJS.ProcessEnv): Promise<EvalExecutorResolution>;
+//# sourceMappingURL=provider-llm-executor.d.ts.map

package/dist/validation/provider-llm-executor.js

@@ -0,0 +1,98 @@
+/**
+ * Agentic QE v3 — Real LLM executor for the eval runner (issue #557 follow-up).
+ *
+ * `aqe eval run` historically defaulted to MockLLMExecutor (canned, keyword-
+ * matched responses), so LLM-mode eval suites scored fabricated output. This
+ * executor makes REAL model calls through the shared LLM layer, so eval
+ * inherits the ADR-123 guardrails automatically: enforced budget caps, the
+ * `AQE_LLM_PROVIDER` selector (including the claude-code subscription), the
+ * billing notice, and the cross-process spend ledger.
+ */
+import { ProviderManager } from '../shared/llm/provider-manager';
+import { billingModeForType } from '../shared/llm/billing-modes';
+import { loadRouterConfig, detectAvailableProvidersFromEnv, } from '../shared/llm/router/config-store';
+import { ClaudeCodeProvider } from '../shared/llm/providers/claude-code';
+import { OllamaProvider } from '../shared/llm/providers/ollama';
+/** Real executor: delegates to ProviderManager.generate (budget-gated). */
+export class ProviderLLMExecutor {
+    manager;
+    providerType;
+    constructor(manager, providerType) {
+        this.manager = manager;
+        this.providerType = providerType;
+    }
+    async execute(prompt, model, options) {
+        const response = await this.manager.generate(prompt, {
+            model,
+            timeoutMs: options?.timeout,
+            preferredProvider: this.providerType,
+        });
+        return {
+            output: response.content,
+            tokensUsed: response.usage.totalTokens,
+            durationMs: response.latencyMs,
+        };
+    }
+}
+/**
+ * Non-billing readiness check: is this provider usable WITHOUT making a paid
+ * probe call? API providers need a key in env; claude-code needs the CLI on
+ * PATH; local providers (ollama) are treated as available when selected.
+ */
+async function isConfiguredCheaply(type, env) {
+    if (type === 'claude-code') {
+        // Free, local: just runs `claude --version`.
+        return new ClaudeCodeProvider().isAvailable();
+    }
+    if (type === 'ollama') {
+        // Local + free probe: only "configured" if the server actually answers,
+        // so a machine with no Ollama running falls through to the hard error
+        // (issue #557 choice: never silently pick a dead provider).
+        return new OllamaProvider().isAvailable();
+    }
+    // API providers: key presence only — never a billing probe.
+    return detectAvailableProvidersFromEnv(env).has(type);
+}
+/**
+ * Resolve a real eval executor from the current environment/config. Honors
+ * `AQE_LLM_PROVIDER`. Returns `{ reason }` (no executor) when nothing is
+ * configured, so the CLI can hard-error instead of silently mocking.
+ */
+export async function resolveEvalExecutor(env = process.env) {
+    const config = loadRouterConfig({ env });
+    // Candidate order: the resolved default provider first, then any other
+    // enabled provider, so an AQE_LLM_PROVIDER override wins.
+    const enabled = Object.entries(config.providers ?? {})
+        .filter(([, cfg]) => cfg?.enabled)
+        .map(([type]) => type);
+    const candidates = [
+        config.defaultProvider,
+        ...enabled.filter((t) => t !== config.defaultProvider),
+    ];
+    for (const type of candidates) {
+        if (await isConfiguredCheaply(type, env)) {
+            const pmConfig = {
+                primary: type,
+                fallbacks: [],
+                loadBalancing: 'round-robin',
+                providers: { [type]: config.providers?.[type] },
+                global: { enableCostTracking: true },
+            };
+            const manager = new ProviderManager(pmConfig);
+            await manager.initialize();
+            return {
+                executor: new ProviderLLMExecutor(manager, type),
+                providerType: type,
+                billingMode: billingModeForType(type),
+                manager,
+            };
+        }
+    }
+    return {
+        reason: 'No LLM provider is configured for real eval runs. Set a provider key ' +
+            '(e.g. ANTHROPIC_API_KEY, OPENAI_API_KEY, COGNITUM_API_KEY), or run on your ' +
+            'Claude subscription with AQE_LLM_PROVIDER=claude-code, or pass --mock to use ' +
+            'canned responses (offline testing only — results are not real).',
+    };
+}
+//# sourceMappingURL=provider-llm-executor.js.map

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentic-qe",
-  "version": "3.12.1",
+  "version": "3.12.2",
   "description": "Agentic Quality Engineering V3 - Domain-Driven Design Architecture with 13 Bounded Contexts, O(log n) coverage analysis, ReasoningBank learning, 60 specialized QE agents, mathematical Coherence verification, deep Claude Flow integration",
   "type": "module",
   "main": "./dist/index.js",

package/dist/cli/chunks/adapter-X4UEC6UJ.js

@@ -1,2 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.1");process.exit(0)}
-import{a,b,c}from"./chunk-QOAI6ETC.js";import"./chunk-C7HGYLNF.js";import"./chunk-RJ7TU2MP.js";import"./chunk-O65OPAZH.js";import"./chunk-UBPJQKYS.js";import"./chunk-U5ITD7ZD.js";export{a as AgentBoosterAdapter,b as createAgentBoosterAdapter,c as createAgentBoosterAdapterSync};

package/dist/cli/chunks/aqe-learning-engine-5FG7JCY7.js

@@ -1,2 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.1");process.exit(0)}
-import{a,b,c,d,e,f}from"./chunk-OJORNYN3.js";import"./chunk-QJXUK54J.js";import"./chunk-PZZDAECU.js";import"./chunk-YGB3MT3R.js";import"./chunk-BPM7NEHS.js";import"./chunk-RIASOMB2.js";import"./chunk-42PYEEYQ.js";import"./chunk-3W2FZXFD.js";import"./chunk-JDZIBSMY.js";import"./chunk-CJTXZI37.js";import"./chunk-CYYT6YEV.js";import"./chunk-OXVAGR64.js";import"./chunk-QORWBION.js";import"./chunk-XS6BIOPQ.js";import"./chunk-EOBZJDJR.js";import"./chunk-XDRAU3W6.js";import"./chunk-PQ4NSXZM.js";import"./chunk-FNMEVLSN.js";import"./chunk-NOCPO55P.js";import"./chunk-IE6FDICX.js";import"./chunk-O65OPAZH.js";import"./chunk-M5P5WNXA.js";import"./chunk-NHRBKOYD.js";import"./chunk-QXZOUETY.js";import"./chunk-BLU7OLQZ.js";import"./chunk-4LKQZW2H.js";import"./chunk-UBPJQKYS.js";import"./chunk-75AVUS5U.js";import"./chunk-TMPRR3FZ.js";import"./chunk-R4A6PY5D.js";import"./chunk-JRNDFC23.js";import"./chunk-6BCVKCG5.js";import"./chunk-UR655D4K.js";import"./chunk-PZ3Q34OT.js";import"./chunk-CXUHD2O5.js";import"./chunk-IC6SOVGO.js";import"./chunk-USTUWUCE.js";import"./chunk-U5ITD7ZD.js";export{b as AQELearningEngine,a as DEFAULT_ENGINE_CONFIG,c as DEFAULT_META_LEARNING_CONFIG,d as MetaLearningEngine,e as createAQELearningEngine,f as createDefaultLearningEngine};

package/dist/cli/chunks/base-CWPMFSZG.js

@@ -1,2 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.1");process.exit(0)}
-import{a,b,c,d,e,f,g,h,i,j,k,l}from"./chunk-BALFQQB7.js";import"./chunk-LRWBOJ5Y.js";import"./chunk-XDRAU3W6.js";import"./chunk-PQ4NSXZM.js";import"./chunk-FNMEVLSN.js";import"./chunk-NOCPO55P.js";import"./chunk-IE6FDICX.js";import"./chunk-O65OPAZH.js";import"./chunk-BLU7OLQZ.js";import"./chunk-UBPJQKYS.js";import"./chunk-75AVUS5U.js";import"./chunk-4U5MJX2H.js";import"./chunk-TMPRR3FZ.js";import"./chunk-R4A6PY5D.js";import"./chunk-JRNDFC23.js";import"./chunk-6BCVKCG5.js";import"./chunk-UR655D4K.js";import"./chunk-PZ3Q34OT.js";import"./chunk-CXUHD2O5.js";import"./chunk-IC6SOVGO.js";import"./chunk-USTUWUCE.js";import"./chunk-U5ITD7ZD.js";export{l as MCPToolBase,k as defaultToolLogger,j as getKernel,i as getLLMRouter,e as getMemoryBackend,g as getSharedLLMRouter,c as getSharedMemoryBackend,a as isMemoryOnlyMode,b as registerRvfResetFn,h as resetSharedLLMRouter,d as resetSharedMemoryBackend,f as setSharedLLMRouter};

package/dist/cli/chunks/browser-workflow-A77OFBVY.js

@@ -1,2 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.1");process.exit(0)}
-import{a}from"./chunk-774X6OJY.js";import"./chunk-BALFQQB7.js";import"./chunk-LRWBOJ5Y.js";import"./chunk-XDRAU3W6.js";import"./chunk-PQ4NSXZM.js";import"./chunk-FNMEVLSN.js";import"./chunk-NOCPO55P.js";import"./chunk-IE6FDICX.js";import"./chunk-O65OPAZH.js";import"./chunk-BLU7OLQZ.js";import"./chunk-UBPJQKYS.js";import"./chunk-75AVUS5U.js";import"./chunk-4U5MJX2H.js";import"./chunk-TMPRR3FZ.js";import"./chunk-R4A6PY5D.js";import"./chunk-JRNDFC23.js";import"./chunk-6BCVKCG5.js";import"./chunk-UR655D4K.js";import"./chunk-PZ3Q34OT.js";import"./chunk-CXUHD2O5.js";import"./chunk-IC6SOVGO.js";import"./chunk-USTUWUCE.js";import"./chunk-U5ITD7ZD.js";export{a as BrowserWorkflowTool};

package/dist/cli/chunks/chunk-CDVKFAYP.js

@@ -1,18 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.1");process.exit(0)}
-import{a as T,c as N}from"./chunk-UBPJQKYS.js";import{b as D,c as L,d as V}from"./chunk-6BCVKCG5.js";N();V();import*as C from"fs";import*as v from"path";import*as W from"yaml";var q=5,H=4,Q=3e4,K=".claude/skills",B=5e3,J=500,$=100,Y=1e3,F=500,z=.8,X=.2,Z=.1,w=.1,ee=.1,te=.1,se=.1,re=200,ne=500,ae=300,ie=1,oe=.9,U={maxWorkers:q,batchSize:H,retryFailedTests:!0,timeout:Q,skillsDir:K,progressIntervalMs:B},S=class{async execute(e,r,t){let s=L($,J+$);return await new Promise(o=>setTimeout(o,s)),{output:this.generateMockResponse(e),tokensUsed:D(F,Y+F),durationMs:s}}generateMockResponse(e){let r=[];return e.includes("alt")&&r.push("alt","1.1.1","perceivable"),e.includes("contrast")&&r.push("contrast","1.4.3","4.5:1"),e.includes("keyboard")&&r.push("keyboard","2.1.1","operable","button"),e.includes("label")&&r.push("label","3.3.2","understandable"),e.includes("ARIA")&&r.push("ARIA","4.1.2","robust"),e.includes("focus")&&r.push("focus","2.4.7","outline"),e.includes("lang")&&r.push("lang","3.1.1","language"),e.includes("caption")&&r.push("caption","1.2.2","track"),e.includes("accessible")&&r.push("accessible","compliant"),r.length===0&&r.push("accessibility","WCAG","finding"),`Analysis complete. Found issues related to: ${r.join(", ")}.
-    Recommendations: Implement proper ${r[0]} attributes for better accessibility.
-    Severity: ${e.includes("critical")?"critical":"serious"}`}},M=class{constructor(e,r,t){this.executor=r;this.config=t;this.id=e}executor;config;id;status="idle";tasksCompleted=0;tasksFailed=0;lastHeartbeat=Date.now();async executeBatch(e){this.status="running";let r=[];for(let t of e){this.lastHeartbeat=Date.now();try{let s=await this.executeTestCase(t);if(r.push(s),s.passed)this.tasksCompleted++;else if(this.tasksFailed++,this.config.retryFailedTests&&!s.passed){let n=await this.executeTestCase(t,!0);n.passed&&(r[r.length-1]=n,this.tasksCompleted++,this.tasksFailed--)}}catch(s){this.tasksFailed++,r.push({testId:t.testCaseId,passed:!1,expectedPatterns:t.testCase.expected_output.must_contain||[],actualPatterns:[],reasoningQuality:0,category:t.testCase.category,priority:t.testCase.priority,error:T(s)})}}return this.status="idle",r}async executeTestCase(e,r=!1){let t=Date.now(),s=this.buildPrompt(e),n=await this.executor.execute(s,e.model,{timeout:this.config.timeout}),o=this.validateResponse(n.output,e.testCase.expected_output,e.testCase.validation);return{testId:e.testCaseId,passed:o.passed,expectedPatterns:e.testCase.expected_output.must_contain||[],actualPatterns:o.foundPatterns,reasoningQuality:o.reasoningQuality,executionTimeMs:Date.now()-t,category:e.testCase.category,priority:e.testCase.priority,error:o.error}}buildPrompt(e){let{testCase:r}=e,t=r.input.context;return`Analyze the following ${t.language||"HTML"} code for accessibility issues.
-Context: ${t.description||r.description}
-WCAG Level: ${t.wcagLevel||"AA"}
-
-Code:
-\`\`\`${t.language||"html"}
-${r.input.code}
-\`\`\`
-
-Provide a detailed analysis including:
-1. All accessibility issues found
-2. WCAG success criteria violated
-3. Severity classification (critical/serious/moderate/minor)
-4. Remediation recommendations with code examples`}validateResponse(e,r,t){let s=[],n=[],o=r.must_contain||[],i=0;for(let p of o)new RegExp(p,"i").test(e)&&(s.push(p),i++);let l=r.must_not_contain||[];for(let p of l)new RegExp(p,"i").test(e)&&n.push(`Output should not contain: ${p}`);let u=r.must_match_regex||[];for(let p of u)new RegExp(p,"i").test(e)||n.push(`Output should match regex: ${p}`);let f=o.length>0?i/o.length:1,m=t?.keyword_match_threshold??z,g=f>=m,c=this.calculateReasoningQuality(e,t),_=t?.reasoning_quality_min??0,h=c>=_;return{passed:g&&h&&n.length===0,foundPatterns:s,reasoningQuality:c,error:n.length>0?n.join("; "):void 0}}calculateReasoningQuality(e,r){let t=0;if(e.length>re&&(t+=X),e.length>ne&&(t+=Z),e.includes(`
-`)&&(t+=w),(e.includes("1.")||e.includes("-"))&&(t+=w),e.includes("```")&&(t+=w),e.includes("WCAG")&&(t+=ee),(e.includes("remediation")||e.includes("recommendation"))&&(t+=te),(e.includes("severity")||e.includes("critical"))&&(t+=se),r?.grading_rubric){let s=r.grading_rubric,n=0;s.completeness&&e.length>ae&&(n+=s.completeness),s.accuracy&&e.includes("WCAG")&&(n+=s.accuracy),s.actionability&&(e.includes("fix")||e.includes("change"))&&(n+=s.actionability),t=Math.max(t,n)}return Math.min(t,1)}getProgress(e){return{workerId:this.id,tasksCompleted:this.tasksCompleted,tasksTotal:e,elapsedMs:Date.now()-this.lastHeartbeat}}},P=class{constructor(e=U,r,t=new S){this.config=e;this.skillValidationLearner=r;this.executor=t;for(let s=0;s<e.maxWorkers;s++)this.workers.push(new M(s,t,e))}config;skillValidationLearner;executor;workers=[];progressCallback;onProgress(e){this.progressCallback=e}async runEvalParallel(e,r){let t=Date.now(),s=this.loadEvalSuite(e);if(!s)throw new Error(`Failed to load eval suite for skill: ${e}`);let n=this.createTasks(s,r),o=this.partitionTestCases(n,this.config.batchSize),i=n.length*(this.config.timeout/2),l=[],u=[];for(let d=0;d<o.length;d++){let b=this.workers[d%this.workers.length];u.push(b.executeBatch(o[d]))}let f=this.startProgressReporting(e,r,n.length,t);try{let d=await Promise.all(u);for(let b of d)l.push(...b)}finally{f&&clearInterval(f)}let m=Date.now()-t,g=l.filter(d=>d.passed).length,c=l.filter(d=>!d.passed).length,_=n.length-l.length,h=l.length>0?g/l.length:0,x=l.length>0?l.reduce((d,b)=>d+b.reasoningQuality,0)/l.length:0,p=m>0?i/m:1,y=h>=s.success_criteria.pass_rate&&x>=(s.success_criteria.avg_reasoning_quality??0);return await this.recordOutcome({skillName:e,trustTier:this.determineTrustTier(s),validationLevel:"eval",model:r,passed:y,score:h,testCaseResults:l,timestamp:new Date,runId:`${e}-${r}-${Date.now()}`,metadata:{duration:m,parallelSpeedup:p,workersUsed:this.config.maxWorkers,version:s.version}}),{skill:e,model:r,totalTests:n.length,passedTests:g,failedTests:c,skippedTests:_,passRate:h,testResults:l,totalDurationMs:m,parallelSpeedup:p,avgReasoningQuality:x,passed:y,workersUsed:this.config.maxWorkers,timestamp:new Date}}async runMultipleEvalsParallel(e,r){let t=new Map,s=[];for(let i of e){t.set(i,[]);for(let l of r)s.push({skill:i,model:l})}let n=s.map(async({skill:i,model:l})=>{let u=await this.runEvalParallel(i,l);return{skill:i,result:u}}),o=await Promise.all(n);for(let{skill:i,result:l}of o)t.get(i).push(l);return t}loadEvalSuite(e){let r=v.isAbsolute(this.config.skillsDir)?this.config.skillsDir:v.join(process.cwd(),this.config.skillsDir),t=v.join(r,e,"evals",`${e}.yaml`);if(!C.existsSync(t))return console.error(`Eval suite not found: ${t}`),null;try{let s=C.readFileSync(t,"utf-8");return W.parse(s)}catch(s){return console.error(`Failed to parse eval suite: ${s instanceof Error?s.message:s}`),null}}createTasks(e,r){return e.test_cases.map((t,s)=>({skillName:e.skill,testCaseId:t.id,testCase:t,model:r,batchId:Math.floor(s/this.config.batchSize),indexInBatch:s%this.config.batchSize}))}partitionTestCases(e,r){let t=[];for(let s=0;s<e.length;s+=r)t.push(e.slice(s,s+r));return t}determineTrustTier(e){let r=e.mcp_integration?.enabled??!1,t=e.learning?.cross_model_comparison??!1,s=e.success_criteria.critical_pass_rate??0;return r&&t&&s>=ie?3:r||s>=oe?2:1}startProgressReporting(e,r,t,s){return this.progressCallback?setInterval(()=>{let n=Date.now()-s,o=this.workers.reduce((c,_)=>c+_.tasksCompleted,0),i=this.workers.reduce((c,_)=>c+_.tasksFailed,0),l=this.workers.filter(c=>c.status==="running").length,u=n>0?o/n:0,f=t-o,m=u>0?f/u:0,g={skill:e,model:r,totalTasks:t,completedTasks:o,failedTasks:i,activeWorkers:l,elapsedMs:n,estimatedRemainingMs:m,workerProgress:this.workers.map(c=>c.getProgress(t))};this.progressCallback(g)},this.config.progressIntervalMs):null}async recordOutcome(e){await this.skillValidationLearner.recordValidationOutcome(e)}getWorkerStatus(){return this.workers.map(e=>({id:e.id,status:e.status,tasksCompleted:e.tasksCompleted,tasksFailed:e.tasksFailed}))}resetWorkers(){for(let e of this.workers)e.tasksCompleted=0,e.tasksFailed=0,e.status="idle"}};function fe(a,e={},r){let t={...U,...e};return new P(t,a,r)}N();import*as R from"node:fs";import*as E from"node:path";import*as j from"yaml";import{spawnSync as le}from"node:child_process";var G=["none","low","medium","high","critical"];function ce(a,e){if(typeof a!="string")return!1;let r=G.indexOf(a),t=G.indexOf(e);return r>=0&&t>=0&&r>=t}function A(a,e){if(typeof e!="string"||!e.startsWith("."))return;let r=e==="."?[]:e.slice(1).split("."),t=a;for(let s of r){if(t==null||typeof t!="object")return;t=t[s]}return t}var ue=(a,e)=>{let r=le("bash",["-c",a],{timeout:e.timeoutMs,cwd:e.cwd,encoding:"utf8",maxBuffer:16777216});return{status:r.status,stdout:typeof r.stdout=="string"?r.stdout:"",stderr:typeof r.stderr=="string"?r.stderr:"",error:r.error}};function de(a){if(!a||typeof a!="object")return!1;let e=a;return!Array.isArray(e.test_cases)||e.test_cases.length===0?!1:typeof e.test_cases[0]?.input?.command=="string"}function me(a,e,r){let t=[],s=a.expected||{};s.exit_code!==void 0&&e!==s.exit_code&&t.push(`exit_code: expected ${s.exit_code}, got ${e}`);let n=!!s.json_fields||s.severity_at_least!==void 0||s.candidate_count_at_least!==void 0,o=null;if(n)try{o=JSON.parse(r)}catch(i){return t.push(`stdout is not valid JSON: ${T(i)}`),{passed:!1,failures:t}}if(s.json_fields)for(let[i,l]of Object.entries(s.json_fields)){let u=A(o,i);I(u,l)||t.push(`json_fields ${i}: expected ${JSON.stringify(l)}, got ${JSON.stringify(u)}`)}if(s.severity_at_least!==void 0){let i=A(o,".output.checkInjection.severity");ce(i,s.severity_at_least)||t.push(`severity_at_least: expected >= ${s.severity_at_least}, got ${JSON.stringify(i)}`)}if(s.candidate_count_at_least!==void 0){let i=A(o,".output.intentScore.candidateCount");(typeof i=="number"?i:Number.NaN)>=s.candidate_count_at_least||t.push(`candidate_count_at_least: expected >= ${s.candidate_count_at_least}, got ${JSON.stringify(i)}`)}return{passed:t.length===0,failures:t}}function I(a,e){if(a===e)return!0;if(a==null||e==null)return a===e;if(typeof a!=typeof e)return!1;if(Array.isArray(a))return!Array.isArray(e)||a.length!==e.length?!1:a.every((r,t)=>I(r,e[t]));if(typeof a=="object"){if(typeof e!="object"||Array.isArray(e))return!1;let r=a,t=e,s=Object.keys(r),n=Object.keys(t);return s.length!==n.length?!1:s.every(o=>I(r[o],t[o]))}return!1}var O=class{skillsDir;timeoutMs;cwd;runner;constructor(e={}){this.skillsDir=e.skillsDir??".claude/skills",this.timeoutMs=e.timeoutMs??6e4,this.cwd=e.cwd??process.cwd(),this.runner=e.runner??ue}loadSuite(e){let r=E.isAbsolute(this.skillsDir)?this.skillsDir:E.join(this.cwd,this.skillsDir),t=E.join(r,e,"evals",`${e}.yaml`);if(!R.existsSync(t))return null;try{let s=j.parse(R.readFileSync(t,"utf-8"));return de(s)?s:null}catch{return null}}async run(e){let r=Date.now(),t=this.loadSuite(e);if(!t)throw new Error(`Command-eval suite not found or not command-mode at ${this.skillsDir}/${e}/evals/${e}.yaml`);let s=[];for(let c of t.test_cases)s.push(this.executeCase(c));let n=s.filter(c=>c.passed).length,o=s.length-n,i=s.length>0?n/s.length:0,l=t.validation?.required_pass_rate??.9,u=t.validation?.critical_must_pass??!0,m=s.filter(c=>c.priority==="critical").every(c=>c.passed),g=i>=l&&(!u||m);return{skill:e,passed:g,totalTests:s.length,passedTests:n,failedTests:o,passRate:i,criticalPassed:m,testResults:s,totalDurationMs:Date.now()-r,timestamp:new Date}}executeCase(e){let r=Date.now(),t=e.input?.setup??[];for(let o of t){let i=this.runner(o,{timeoutMs:this.timeoutMs,cwd:this.cwd});if(i.status!==0)return{testId:e.id,passed:!1,category:e.category,priority:e.priority,exitCode:i.status,stdoutSnippet:k(i.stdout),stderrSnippet:k(i.stderr),failures:[],setupFailure:`setup step failed (exit ${i.status}): ${o}`,durationMs:Date.now()-r}}let s=this.runner(e.input.command,{timeoutMs:this.timeoutMs,cwd:this.cwd}),n=me(e,s.status,s.stdout);return{testId:e.id,passed:n.passed,category:e.category,priority:e.priority,exitCode:s.status,stdoutSnippet:k(s.stdout),stderrSnippet:k(s.stderr),failures:n.failures,durationMs:Date.now()-r}}};function k(a,e=500){return a.length<=e?a:`${a.slice(0,e)}\u2026[truncated ${a.length-e}b]`}function be(a={}){return new O(a)}export{U as a,fe as b,de as c,be as d};

package/dist/cli/chunks/chunk-FNMEVLSN.js

@@ -1,2 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.1");process.exit(0)}
-import{a as e,b as g,e as r}from"./chunk-NOCPO55P.js";import{c as o}from"./chunk-U5ITD7ZD.js";var L=o(()=>{"use strict";e();g();r()});export{L as a};

package/dist/cli/chunks/chunk-P7G5LGUE.js

@@ -1,3 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.1");process.exit(0)}
-import{a,b as u}from"./chunk-R7XEACKR.js";import{m as P}from"./chunk-XDRAU3W6.js";import{b as p}from"./chunk-IE6FDICX.js";P();import*as i from"fs";import*as d from"path";var v="llm-config.json",f={claude:["ANTHROPIC_API_KEY","CLAUDE_API_KEY"],openai:["OPENAI_API_KEY"],ollama:[],openrouter:["OPENROUTER_API_KEY"],gemini:["GOOGLE_AI_API_KEY","GEMINI_API_KEY","GOOGLE_API_KEY"],"azure-openai":["AZURE_OPENAI_API_KEY"],bedrock:["AWS_ACCESS_KEY_ID"],onnx:[]},R=new Set(["claude","openai","ollama","openrouter","gemini","azure-openai","bedrock"]);function E(e){let r=e??p();return d.join(r,".agentic-qe",v)}function _(e=process.env){let r=new Set;for(let n of a){if(!R.has(n))continue;let o=f[n];if(o.length===0){r.add(n);continue}o.some(t=>(e[t]??"").trim().length>0)&&r.add(n)}return r}function g(e){let r=E(e);if(!i.existsSync(r))return{};let n=i.readFileSync(r,"utf8");if(!n.trim())return{};let o=JSON.parse(n);if(typeof o!="object"||o===null)throw new Error(`Invalid ${v}: expected object, got ${typeof o}`);return o}function C(e,r){let n=E(r),o=d.dirname(n);i.existsSync(o)||i.mkdirSync(o,{recursive:!0});let t=y(e),s=JSON.stringify(t,null,2)+`
-`,c=n+".tmp-"+process.pid;i.writeFileSync(c,s,{encoding:"utf8",mode:384}),i.renameSync(c,n)}function y(e){if(!e.providers)return JSON.parse(JSON.stringify(e));let r=JSON.parse(JSON.stringify(e)),n=[];if(r.providers)for(let o of Object.keys(r.providers)){let t=r.providers[o];t&&"apiKey"in t&&t.apiKey&&(delete t.apiKey,n.push(o))}return n.length>0&&console.warn(`[router-config] Refused to persist apiKey for: ${n.join(", ")}. API keys belong in environment variables (e.g. ANTHROPIC_API_KEY, GEMINI_API_KEY, OPENAI_API_KEY, OPENROUTER_API_KEY), NOT in .agentic-qe/llm-config.json which may be checked into source control.`),r}function l(e,r){let n={...e,...r,providers:{...e.providers??{}}};if(r.providers)for(let[o,t]of Object.entries(r.providers)){if(!t)continue;let s=o;n.providers[s]={...e.providers?.[s]??{},...t}}return n}function m(e,r=process.env){let n=_(r),o={...e,providers:{...e.providers??{}}};for(let t of n){if(f[t].length===0)continue;let c=o.providers[t]??{enabled:!1};c.enabled||(o.providers[t]={...c,enabled:!0})}return o}function A(e={}){let r=g(e.projectRoot),n=l(u,r),o=m(n,e.env);return e.override?l(o,e.override):o}function S(e={}){let r=e.env??process.env;for(let o of a){let t=f[o];if(t.length!==0&&t.some(s=>(r[s]??"").trim().length>0))return!0}let n=g(e.projectRoot);if(n.providers){for(let o of Object.values(n.providers))if(o?.enabled)return!0}return!1}export{_ as a,C as b,A as c,S as d};

package/dist/cli/chunks/chunk-R7XEACKR.js

@@ -1,2 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.1");process.exit(0)}
-var t=["claude","openai","ollama","openrouter","gemini","azure-openai","bedrock","onnx"];var e={maxAttempts:3,delayMs:100,trySameProviderAlternateModel:!0,skipOpenCircuits:!0,retryableErrors:["RATE_LIMITED","TIMEOUT","NETWORK_ERROR","PROVIDER_UNAVAILABLE"]},r=[{provider:"claude",models:["claude-sonnet-4-6","claude-haiku-4-5-20251001"],enabled:!0,priority:100,maxAttempts:2,timeoutMs:3e4},{provider:"openai",models:["gpt-4o","gpt-4o-mini"],enabled:!0,priority:90,maxAttempts:2,timeoutMs:3e4},{provider:"openrouter",models:["anthropic/claude-sonnet-4","openai/gpt-4o"],enabled:!0,priority:85,maxAttempts:2,timeoutMs:3e4},{provider:"ollama",models:["llama3.1","mistral"],enabled:!0,priority:80,maxAttempts:2,timeoutMs:6e4},{provider:"onnx",models:["phi-4","all-MiniLM-L6-v2"],enabled:!0,priority:70,maxAttempts:1,timeoutMs:12e4}],o={id:"default-fallback-chain",entries:r,maxRetries:3,retryDelayMs:100,backoffMultiplier:2,maxDelayMs:5e3},n={mode:"rule-based",rules:[],defaultProvider:"claude",defaultModel:"claude-sonnet-4-6",fallbackChain:o,fallbackBehavior:e,providers:{claude:{enabled:!0,defaultModel:"claude-sonnet-4-6"},openai:{enabled:!0,defaultModel:"gpt-4o"},ollama:{enabled:!0,defaultModel:"llama3.1"},openrouter:{enabled:!0,defaultModel:"anthropic/claude-sonnet-4"},gemini:{enabled:!1,defaultModel:"gemini-pro"},"azure-openai":{enabled:!1},bedrock:{enabled:!1},onnx:{enabled:!0,defaultModel:"phi-4"}},enableMetrics:!0,metricsConfig:{enabled:!0,collectionIntervalMs:6e4,retentionMs:864e5,trackPerRequest:!0},maxDecisionTimeMs:10,cacheDecisions:!0,decisionCacheTtlMs:6e4,cacheConfig:{enabled:!0,ttlMs:6e4,maxEntries:1e3,keyStrategy:"agent-domain"},auditConfig:{enabled:!0,maxEntries:1e4,logContent:!1,logResponse:!1}},a=[{canonicalId:"claude-opus-4-7",canonicalName:"Claude Opus 4.7",providerIds:{claude:"claude-opus-4-7",openrouter:"anthropic/claude-opus-4.7",bedrock:"anthropic.claude-opus-4-7-v1:0"},capabilities:{supportsStreaming:!0,supportsTools:!0,supportsMCP:!0,supportsVision:!0,supportsExtendedThinking:!0,supportsJsonMode:!0,maxContextTokens:1e6,maxOutputTokens:128e3},tier:"flagship",family:"claude",inputCostPer1M:5,outputCostPer1M:25},{canonicalId:"claude-sonnet-4-6",canonicalName:"Claude Sonnet 4.6",providerIds:{claude:"claude-sonnet-4-6",openrouter:"anthropic/claude-sonnet-4.6",bedrock:"anthropic.claude-sonnet-4-6-v1:0"},capabilities:{supportsStreaming:!0,supportsTools:!0,supportsMCP:!0,supportsVision:!0,supportsExtendedThinking:!0,supportsJsonMode:!0,maxContextTokens:2e5,maxOutputTokens:64e3},tier:"advanced",family:"claude",inputCostPer1M:3,outputCostPer1M:15},{canonicalId:"claude-haiku-4-5",canonicalName:"Claude Haiku 4.5",providerIds:{claude:"claude-haiku-4-5-20251001",openrouter:"anthropic/claude-haiku-4.5",bedrock:"anthropic.claude-haiku-4-5-v1:0"},capabilities:{supportsStreaming:!0,supportsTools:!0,supportsMCP:!1,supportsVision:!0,supportsExtendedThinking:!0,supportsJsonMode:!0,maxContextTokens:2e5,maxOutputTokens:16384},tier:"standard",family:"claude",inputCostPer1M:1,outputCostPer1M:5},{canonicalId:"gpt-4o",canonicalName:"GPT-4o",providerIds:{openai:"gpt-4o","azure-openai":"gpt-4o",openrouter:"openai/gpt-4o"},capabilities:{supportsStreaming:!0,supportsTools:!0,supportsMCP:!1,supportsVision:!0,supportsExtendedThinking:!1,supportsJsonMode:!0,maxContextTokens:128e3,maxOutputTokens:16384},tier:"advanced",family:"gpt",inputCostPer1M:2.5,outputCostPer1M:10},{canonicalId:"gpt-4o-mini",canonicalName:"GPT-4o Mini",providerIds:{openai:"gpt-4o-mini","azure-openai":"gpt-4o-mini",openrouter:"openai/gpt-4o-mini"},capabilities:{supportsStreaming:!0,supportsTools:!0,supportsMCP:!1,supportsVision:!0,supportsExtendedThinking:!1,supportsJsonMode:!0,maxContextTokens:128e3,maxOutputTokens:16384},tier:"standard",family:"gpt",inputCostPer1M:.15,outputCostPer1M:.6},{canonicalId:"gemini-pro",canonicalName:"Gemini 2.0 Pro",providerIds:{gemini:"gemini-2.0-pro",openrouter:"google/gemini-2.0-pro"},capabilities:{supportsStreaming:!0,supportsTools:!0,supportsMCP:!1,supportsVision:!0,supportsExtendedThinking:!1,supportsJsonMode:!0,maxContextTokens:2e6,maxOutputTokens:8192},tier:"advanced",family:"gemini",inputCostPer1M:1.25,outputCostPer1M:5},{canonicalId:"llama-3.1-70b",canonicalName:"LLaMA 3.1 70B",providerIds:{ollama:"llama3.1:70b",openrouter:"meta-llama/llama-3.1-70b",bedrock:"meta.llama3-1-70b-instruct-v1:0"},capabilities:{supportsStreaming:!0,supportsTools:!0,supportsMCP:!1,supportsVision:!1,supportsExtendedThinking:!1,supportsJsonMode:!0,maxContextTokens:128e3,maxOutputTokens:4096},tier:"advanced",family:"llama",inputCostPer1M:.9,outputCostPer1M:.9},{canonicalId:"phi-4",canonicalName:"Microsoft Phi-4",providerIds:{ollama:"phi4",onnx:"phi-4",openrouter:"microsoft/phi-4"},capabilities:{supportsStreaming:!0,supportsTools:!1,supportsMCP:!1,supportsVision:!1,supportsExtendedThinking:!1,supportsJsonMode:!1,maxContextTokens:16e3,maxOutputTokens:4096,requiresNetwork:!1},tier:"efficient",family:"phi",inputCostPer1M:0,outputCostPer1M:0}];export{t as a,n as b,a as c};