@miller-tech/uap 1.0.0

package/dist/models/execution-profiles.js

@@ -0,0 +1,403 @@
+/**
+ * Model Execution Profiles
+ *
+ * Evidence-based feature flag presets optimized for specific model families.
+ * Each profile was derived from benchmark data (Harbor Terminal-Bench, 10 tasks).
+ *
+ * Key insight from 13 benchmark runs:
+ *   - Small models (≤7B active params): LESS is MORE. Lean instructions,
+ *     no conversation injection, domain hints only.
+ *   - Medium models (13-30B active): Can handle moderate injection.
+ *     Reflection checkpoints help. Search is neutral.
+ *   - Large models (70B+): Can handle everything. Search, reflection,
+ *     budget pressure, strategy switching all help.
+ *
+ * Usage:
+ *   const profile = getExecutionProfile('qwen3.5-35b');
+ *   // or auto-detect:
+ *   const profile = detectExecutionProfile('qwen/qwen35-a3b-iq4xs');
+ */
+/**
+ * Small MoE models: Qwen3.5 35B (3B active), Mixtral 8x7B (12B active)
+ *
+ * PROVEN in 13 benchmark runs:
+ * - Domain hints: +20% pass rate
+ * - Lean instructions: every extra token hurts
+ * - No conversation injection: regresses from 40% to 20%
+ * - Low temperature: reduces stochastic variation
+ */
+const SMALL_MOE_PROFILE = {
+    id: 'small-moe',
+    name: 'Small MoE (≤12B active)',
+    modelPatterns: [
+        /qwen.*3\.5.*35b/i,
+        /qwen.*a3b/i,
+        /mixtral.*8x7b/i,
+        /mixtral.*moe/i,
+        /deepseek.*v2.*lite/i,
+    ],
+    activeParamsBillions: 3,
+    config: {
+        domainHints: true,
+        prependNotReplaceLoopBreaker: true,
+        preExecutionHooks: true,
+        lowTemperature: true,
+        temperature: 0.15,
+        gccFlagMutation: true,
+        loopEscapeHatch: true,
+        loopEscapeThreshold: 3,
+        // Everything else OFF -- proven harmful for small models
+        webSearch: false,
+        preExecWebResearch: false,
+        verifierAwareTesting: false,
+        reflectionCheckpoints: false,
+        progressiveBudgetPressure: false,
+        outputDiffStrategySwitch: false,
+        cwdInjection: false,
+        softBudget: 35,
+        hardBudget: 50,
+        toolChoiceForce: 'required',
+    },
+    rationale: 'Proven in 13 Harbor Terminal-Bench runs. Lean instructions + domain hints ' +
+        'achieved 40% peak (4/10). Any conversation injection regresses to 20%.',
+};
+/**
+ * Small dense models: Qwen 7B, Llama 8B, Gemma 9B, Phi-3 7B
+ *
+ * Similar to small MoE but slightly more capable per-token.
+ * Can handle marginally longer instructions.
+ */
+const SMALL_DENSE_PROFILE = {
+    id: 'small-dense',
+    name: 'Small Dense (7-9B)',
+    modelPatterns: [
+        /qwen.*[^0-9]7b/i,
+        /llama.*[^0-9]8b/i,
+        /gemma.*[^0-9]9b/i,
+        /phi.*[^0-9][37]b/i,
+        /mistral.*[^0-9]7b/i,
+        /codestral.*mamba/i,
+    ],
+    activeParamsBillions: 7,
+    config: {
+        domainHints: true,
+        prependNotReplaceLoopBreaker: true,
+        preExecutionHooks: true,
+        lowTemperature: true,
+        temperature: 0.2,
+        gccFlagMutation: true,
+        loopEscapeHatch: true,
+        loopEscapeThreshold: 3,
+        webSearch: false,
+        preExecWebResearch: false,
+        verifierAwareTesting: false,
+        reflectionCheckpoints: false,
+        progressiveBudgetPressure: false,
+        outputDiffStrategySwitch: false,
+        cwdInjection: false,
+        softBudget: 35,
+        hardBudget: 50,
+        toolChoiceForce: 'required',
+    },
+    rationale: 'Similar to small MoE. Dense 7-9B models have comparable effective ' +
+        'reasoning to 3B-active MoE. Keep instructions lean.',
+};
+/**
+ * Medium models: Qwen 32B, Llama 70B, DeepSeek V3 (37B active), Gemma 27B
+ *
+ * Can handle moderate instruction length and some conversation injection.
+ * Reflection checkpoints may help. Search is neutral.
+ */
+const MEDIUM_PROFILE = {
+    id: 'medium',
+    name: 'Medium (13-37B active)',
+    modelPatterns: [
+        /qwen.*[^0-9]32b/i,
+        /qwen.*[^0-9]14b/i,
+        /llama.*[^0-9]70b/i,
+        /deepseek.*v3/i,
+        /gemma.*[^0-9]27b/i,
+        /codestral.*[^0-9]22b/i,
+        /command.*r.*plus/i,
+        /yi.*[^0-9]34b/i,
+    ],
+    activeParamsBillions: 30,
+    config: {
+        domainHints: true,
+        prependNotReplaceLoopBreaker: true,
+        preExecutionHooks: true,
+        lowTemperature: true,
+        temperature: 0.3,
+        gccFlagMutation: true,
+        loopEscapeHatch: true,
+        loopEscapeThreshold: 4,
+        // Medium models can handle some extras
+        webSearch: false, // Still off -- needs larger context window
+        preExecWebResearch: true, // Cache results, model can read if needed
+        verifierAwareTesting: true, // Model can handle the extra instruction
+        reflectionCheckpoints: true, // Helps medium models stay on track
+        reflectionInterval: 15,
+        progressiveBudgetPressure: false, // Still too noisy
+        outputDiffStrategySwitch: false, // Still too noisy
+        cwdInjection: false, // Still breaks commands
+        softBudget: 40,
+        hardBudget: 60,
+        toolChoiceForce: 'required',
+    },
+    rationale: 'Medium models can handle reflection checkpoints and verifier hints ' +
+        'without losing task context. Pre-exec research cached silently.',
+};
+/**
+ * Large models: Claude Opus 4, GPT-4o, Gemini 1.5 Pro, Llama 405B
+ *
+ * Can handle everything. Full feature set enabled.
+ * These models benefit from richer context and can synthesize search results.
+ */
+const LARGE_PROFILE = {
+    id: 'large',
+    name: 'Large (70B+ / frontier)',
+    modelPatterns: [
+        /claude.*opus/i,
+        /claude.*sonnet.*4/i,
+        /gpt.*4/i,
+        /gpt.*5/i,
+        /o[134]/i,
+        /gemini.*1\.5.*pro/i,
+        /gemini.*2/i,
+        /gemini.*ultra/i,
+        /llama.*[^0-9]405b/i,
+        /deepseek.*r1/i,
+        /qwen.*[^0-9]72b/i,
+        /qwen.*[^0-9]110b/i,
+    ],
+    activeParamsBillions: 100,
+    config: {
+        domainHints: true,
+        prependNotReplaceLoopBreaker: true,
+        preExecutionHooks: true,
+        lowTemperature: false, // Large models benefit from some creativity
+        temperature: 0.4,
+        gccFlagMutation: true,
+        loopEscapeHatch: true,
+        loopEscapeThreshold: 5,
+        // Large models can handle everything
+        webSearch: true, // Can synthesize search results
+        preExecWebResearch: true,
+        verifierAwareTesting: true,
+        reflectionCheckpoints: true,
+        reflectionInterval: 10, // More frequent -- model can handle it
+        progressiveBudgetPressure: true, // Helps with time management
+        outputDiffStrategySwitch: true, // Can process strategy alternatives
+        cwdInjection: false, // Still breaks commands regardless of model size
+        softBudget: 50,
+        hardBudget: 80,
+        toolChoiceForce: 'auto', // Large models don't need forced tool choice
+    },
+    rationale: 'Frontier models benefit from richer context, reflection, and search. ' +
+        'Higher budgets and auto tool_choice since they self-regulate well.',
+};
+/**
+ * Claude-specific profile: Opus 4, Opus 4.6, Sonnet 4
+ *
+ * Claude models have excellent instruction following and tool use.
+ * They don't need forced tool_choice and benefit from higher temperature.
+ */
+const CLAUDE_PROFILE = {
+    id: 'claude',
+    name: 'Claude (Opus/Sonnet 4+)',
+    modelPatterns: [
+        /claude.*opus.*4/i,
+        /claude.*sonnet.*4/i,
+        /anthropic.*opus/i,
+        /anthropic.*sonnet/i,
+    ],
+    activeParamsBillions: 200, // Estimated
+    config: {
+        domainHints: true,
+        prependNotReplaceLoopBreaker: true,
+        preExecutionHooks: true,
+        lowTemperature: false,
+        temperature: 0.5,
+        gccFlagMutation: false, // Claude gets gcc flags right
+        loopEscapeHatch: true,
+        loopEscapeThreshold: 5,
+        webSearch: true,
+        preExecWebResearch: true,
+        verifierAwareTesting: true,
+        reflectionCheckpoints: false, // Claude self-reflects naturally
+        progressiveBudgetPressure: true,
+        outputDiffStrategySwitch: false, // Claude changes strategy on its own
+        cwdInjection: false,
+        softBudget: 60,
+        hardBudget: 100,
+        toolChoiceForce: 'auto', // Claude uses tools reliably without forcing
+    },
+    rationale: 'Claude models have strong instruction following and self-regulation. ' +
+        'No need for forced tool_choice or reflection injection. Higher budgets.',
+};
+/**
+ * GPT-specific profile: GPT-4o, GPT-4.5, o1, o3
+ *
+ * GPT models benefit from structured prompts and explicit phase management.
+ */
+const GPT_PROFILE = {
+    id: 'gpt',
+    name: 'GPT (4o/4.5/o-series)',
+    modelPatterns: [/gpt.*4o/i, /gpt.*4\.5/i, /gpt.*5/i, /o1/i, /o3/i, /o4/i],
+    activeParamsBillions: 200, // Estimated
+    config: {
+        domainHints: true,
+        prependNotReplaceLoopBreaker: true,
+        preExecutionHooks: true,
+        lowTemperature: false,
+        temperature: 0.4,
+        gccFlagMutation: true,
+        loopEscapeHatch: true,
+        loopEscapeThreshold: 4,
+        webSearch: true,
+        preExecWebResearch: true,
+        verifierAwareTesting: true,
+        reflectionCheckpoints: true,
+        reflectionInterval: 12,
+        progressiveBudgetPressure: true, // GPT benefits from phase structure
+        outputDiffStrategySwitch: true,
+        cwdInjection: false,
+        softBudget: 50,
+        hardBudget: 80,
+        toolChoiceForce: 'auto',
+    },
+    rationale: 'GPT models benefit from structured phases and explicit budget management. ' +
+        'Strong tool use but benefits from reflection checkpoints.',
+};
+/**
+ * Gemini-specific profile: Gemini 1.5 Pro, Gemini 2.0, Gemini Ultra
+ *
+ * Gemini has very large context windows. Can handle extensive instructions.
+ */
+const GEMINI_PROFILE = {
+    id: 'gemini',
+    name: 'Gemini (1.5 Pro / 2.0)',
+    modelPatterns: [/gemini.*1\.5/i, /gemini.*2/i, /gemini.*ultra/i, /gemini.*pro/i],
+    activeParamsBillions: 150, // Estimated
+    config: {
+        domainHints: true,
+        prependNotReplaceLoopBreaker: true,
+        preExecutionHooks: true,
+        lowTemperature: false,
+        temperature: 0.4,
+        gccFlagMutation: true,
+        loopEscapeHatch: true,
+        loopEscapeThreshold: 4,
+        webSearch: true, // Gemini can synthesize well
+        preExecWebResearch: true,
+        verifierAwareTesting: true,
+        reflectionCheckpoints: true,
+        reflectionInterval: 10,
+        progressiveBudgetPressure: true,
+        outputDiffStrategySwitch: true,
+        cwdInjection: false,
+        softBudget: 50,
+        hardBudget: 80,
+        toolChoiceForce: 'auto',
+    },
+    rationale: 'Gemini has massive context windows (1M+ tokens). Can handle extensive ' +
+        'instructions, search results, and reflection without context pressure.',
+};
+// ============================================================================
+// Profile Registry
+// ============================================================================
+/** All profiles ordered from most specific to most general */
+const ALL_PROFILES = [
+    CLAUDE_PROFILE,
+    GPT_PROFILE,
+    GEMINI_PROFILE,
+    SMALL_MOE_PROFILE,
+    SMALL_DENSE_PROFILE,
+    MEDIUM_PROFILE,
+    LARGE_PROFILE,
+];
+// ============================================================================
+// Public API
+// ============================================================================
+/**
+ * Get a profile by ID.
+ */
+export function getExecutionProfile(profileId) {
+    return ALL_PROFILES.find((p) => p.id === profileId);
+}
+/**
+ * Auto-detect the best execution profile for a model name/ID.
+ *
+ * Matches against model patterns in order of specificity.
+ * Falls back to size-based heuristic if no pattern matches.
+ *
+ * @param modelName - Model name or ID (e.g., "claude-opus-4", "qwen/qwen35-a3b-iq4xs")
+ * @returns The best matching profile
+ */
+export function detectExecutionProfile(modelName) {
+    // Try pattern matching first (most specific wins)
+    for (const profile of ALL_PROFILES) {
+        for (const pattern of profile.modelPatterns) {
+            if (pattern.test(modelName)) {
+                return profile;
+            }
+        }
+    }
+    // Fallback: extract parameter count from model name
+    const paramMatch = modelName.match(/(\d+)[bB]/);
+    if (paramMatch) {
+        const params = parseInt(paramMatch[1], 10);
+        if (params <= 9)
+            return SMALL_DENSE_PROFILE;
+        if (params <= 35)
+            return MEDIUM_PROFILE;
+        return LARGE_PROFILE;
+    }
+    // Ultimate fallback: use small-dense (safest defaults)
+    return SMALL_DENSE_PROFILE;
+}
+/**
+ * Get the merged config for a model: profile defaults + user overrides.
+ *
+ * @param modelName - Model name for auto-detection
+ * @param userOverrides - User's agentExecution config from .uap.json
+ * @returns Merged config with profile defaults and user overrides
+ */
+export function getExecutionConfig(modelName, userOverrides) {
+    const profile = detectExecutionProfile(modelName);
+    // Start with schema defaults, overlay profile, then user overrides
+    const merged = {
+        domainHints: true,
+        prependNotReplaceLoopBreaker: true,
+        preExecutionHooks: true,
+        lowTemperature: true,
+        temperature: 0.15,
+        gccFlagMutation: true,
+        loopEscapeHatch: true,
+        loopEscapeThreshold: 3,
+        webSearch: false,
+        webSearchEndpoint: 'http://192.168.1.165:8888',
+        preExecWebResearch: false,
+        verifierAwareTesting: false,
+        reflectionCheckpoints: false,
+        reflectionInterval: 15,
+        progressiveBudgetPressure: false,
+        outputDiffStrategySwitch: false,
+        cwdInjection: false,
+        softBudget: 35,
+        hardBudget: 50,
+        toolChoiceForce: 'required',
+        // Apply profile
+        ...profile.config,
+        // Apply user overrides (highest priority)
+        ...userOverrides,
+    };
+    return { profile, config: merged };
+}
+/**
+ * List all available profiles.
+ */
+export function listExecutionProfiles() {
+    return [...ALL_PROFILES];
+}
+//# sourceMappingURL=execution-profiles.js.map

package/dist/models/execution-profiles.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"execution-profiles.js","sourceRoot":"","sources":["../../src/models/execution-profiles.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAuBH;;;;;;;;GAQG;AACH,MAAM,iBAAiB,GAAqB;IAC1C,EAAE,EAAE,WAAW;IACf,IAAI,EAAE,yBAAyB;IAC/B,aAAa,EAAE;QACb,kBAAkB;QAClB,YAAY;QACZ,gBAAgB;QAChB,eAAe;QACf,qBAAqB;KACtB;IACD,oBAAoB,EAAE,CAAC;IACvB,MAAM,EAAE;QACN,WAAW,EAAE,IAAI;QACjB,4BAA4B,EAAE,IAAI;QAClC,iBAAiB,EAAE,IAAI;QACvB,cAAc,EAAE,IAAI;QACpB,WAAW,EAAE,IAAI;QACjB,eAAe,EAAE,IAAI;QACrB,eAAe,EAAE,IAAI;QACrB,mBAAmB,EAAE,CAAC;QACtB,yDAAyD;QACzD,SAAS,EAAE,KAAK;QAChB,kBAAkB,EAAE,KAAK;QACzB,oBAAoB,EAAE,KAAK;QAC3B,qBAAqB,EAAE,KAAK;QAC5B,yBAAyB,EAAE,KAAK;QAChC,wBAAwB,EAAE,KAAK;QAC/B,YAAY,EAAE,KAAK;QACnB,UAAU,EAAE,EAAE;QACd,UAAU,EAAE,EAAE;QACd,eAAe,EAAE,UAAU;KAC5B;IACD,SAAS,EACP,4EAA4E;QAC5E,wEAAwE;CAC3E,CAAC;AAEF;;;;;GAKG;AACH,MAAM,mBAAmB,GAAqB;IAC5C,EAAE,EAAE,aAAa;IACjB,IAAI,EAAE,oBAAoB;IAC1B,aAAa,EAAE;QACb,iBAAiB;QACjB,kBAAkB;QAClB,kBAAkB;QAClB,mBAAmB;QACnB,oBAAoB;QACpB,mBAAmB;KACpB;IACD,oBAAoB,EAAE,CAAC;IACvB,MAAM,EAAE;QACN,WAAW,EAAE,IAAI;QACjB,4BAA4B,EAAE,IAAI;QAClC,iBAAiB,EAAE,IAAI;QACvB,cAAc,EAAE,IAAI;QACpB,WAAW,EAAE,GAAG;QAChB,eAAe,EAAE,IAAI;QACrB,eAAe,EAAE,IAAI;QACrB,mBAAmB,EAAE,CAAC;QACtB,SAAS,EAAE,KAAK;QAChB,kBAAkB,EAAE,KAAK;QACzB,oBAAoB,EAAE,KAAK;QAC3B,qBAAqB,EAAE,KAAK;QAC5B,yBAAyB,EAAE,KAAK;QAChC,wBAAwB,EAAE,KAAK;QAC/B,YAAY,EAAE,KAAK;QACnB,UAAU,EAAE,EAAE;QACd,UAAU,EAAE,EAAE;QACd,eAAe,EAAE,UAAU;KAC5B;IACD,SAAS,EACP,oEAAoE;QACpE,qDAAqD;CACxD,CAAC;AAEF;;;;;GAKG;AACH,MAAM,cAAc,GAAqB;IACvC,EAAE,EAAE,QAAQ;IACZ,IAAI,EAAE,wBAAwB;IAC9B,aAAa,EAAE;QACb,kBAAkB;QAClB,kBAAkB;QAClB,mBAAmB;QACnB,eAAe;QACf,mBAAmB;QACnB,uBAAuB;QACvB,mBAAmB;QACnB,gBAAgB;KACjB;IACD,oBAAoB,EAAE,EAAE;IACxB,MAAM,EAAE;QACN,WAAW,EAAE,IAAI;QACjB,4BAA4B,EAAE,IAAI;QAClC,iBAAiB,EAAE,IAAI;QACvB,cAAc,EAAE,IAAI;QACpB,WAAW,EAAE,GAAG;QAChB,eAAe,EAAE,IAAI;QACrB,eAAe,EAAE,IAAI;QACrB,mBAAmB,EAAE,CAAC;QACtB,uCAAuC;QACvC,SAAS,EAAE,KAAK,EAAE,2CAA2C;QAC7D,kBAAkB,EAAE,IAAI,EAAE,0CAA0C;QACpE,oBAAoB,EAAE,IAAI,EAAE,yCAAyC;QACrE,qBAAqB,EAAE,IAAI,EAAE,oCAAoC;QACjE,kBAAkB,EAAE,EAAE;QACtB,yBAAyB,EAAE,KAAK,EAAE,kBAAkB;QACpD,wBAAwB,EAAE,KAAK,EAAE,kBAAkB;QACnD,YAAY,EAAE,KAAK,EAAE,wBAAwB;QAC7C,UAAU,EAAE,EAAE;QACd,UAAU,EAAE,EAAE;QACd,eAAe,EAAE,UAAU;KAC5B;IACD,SAAS,EACP,qEAAqE;QACrE,iEAAiE;CACpE,CAAC;AAEF;;;;;GAKG;AACH,MAAM,aAAa,GAAqB;IACtC,EAAE,EAAE,OAAO;IACX,IAAI,EAAE,yBAAyB;IAC/B,aAAa,EAAE;QACb,eAAe;QACf,oBAAoB;QACpB,SAAS;QACT,SAAS;QACT,SAAS;QACT,oBAAoB;QACpB,YAAY;QACZ,gBAAgB;QAChB,oBAAoB;QACpB,eAAe;QACf,kBAAkB;QAClB,mBAAmB;KACpB;IACD,oBAAoB,EAAE,GAAG;IACzB,MAAM,EAAE;QACN,WAAW,EAAE,IAAI;QACjB,4BAA4B,EAAE,IAAI;QAClC,iBAAiB,EAAE,IAAI;QACvB,cAAc,EAAE,KAAK,EAAE,4CAA4C;QACnE,WAAW,EAAE,GAAG;QAChB,eAAe,EAAE,IAAI;QACrB,eAAe,EAAE,IAAI;QACrB,mBAAmB,EAAE,CAAC;QACtB,qCAAqC;QACrC,SAAS,EAAE,IAAI,EAAE,gCAAgC;QACjD,kBAAkB,EAAE,IAAI;QACxB,oBAAoB,EAAE,IAAI;QAC1B,qBAAqB,EAAE,IAAI;QAC3B,kBAAkB,EAAE,EAAE,EAAE,uCAAuC;QAC/D,yBAAyB,EAAE,IAAI,EAAE,6BAA6B;QAC9D,wBAAwB,EAAE,IAAI,EAAE,oCAAoC;QACpE,YAAY,EAAE,KAAK,EAAE,iDAAiD;QACtE,UAAU,EAAE,EAAE;QACd,UAAU,EAAE,EAAE;QACd,eAAe,EAAE,MAAM,EAAE,6CAA6C;KACvE;IACD,SAAS,EACP,uEAAuE;QACvE,oEAAoE;CACvE,CAAC;AAEF;;;;;GAKG;AACH,MAAM,cAAc,GAAqB;IACvC,EAAE,EAAE,QAAQ;IACZ,IAAI,EAAE,yBAAyB;IAC/B,aAAa,EAAE;QACb,kBAAkB;QAClB,oBAAoB;QACpB,kBAAkB;QAClB,oBAAoB;KACrB;IACD,oBAAoB,EAAE,GAAG,EAAE,YAAY;IACvC,MAAM,EAAE;QACN,WAAW,EAAE,IAAI;QACjB,4BAA4B,EAAE,IAAI;QAClC,iBAAiB,EAAE,IAAI;QACvB,cAAc,EAAE,KAAK;QACrB,WAAW,EAAE,GAAG;QAChB,eAAe,EAAE,KAAK,EAAE,8BAA8B;QACtD,eAAe,EAAE,IAAI;QACrB,mBAAmB,EAAE,CAAC;QACtB,SAAS,EAAE,IAAI;QACf,kBAAkB,EAAE,IAAI;QACxB,oBAAoB,EAAE,IAAI;QAC1B,qBAAqB,EAAE,KAAK,EAAE,iCAAiC;QAC/D,yBAAyB,EAAE,IAAI;QAC/B,wBAAwB,EAAE,KAAK,EAAE,qCAAqC;QACtE,YAAY,EAAE,KAAK;QACnB,UAAU,EAAE,EAAE;QACd,UAAU,EAAE,GAAG;QACf,eAAe,EAAE,MAAM,EAAE,6CAA6C;KACvE;IACD,SAAS,EACP,uEAAuE;QACvE,yEAAyE;CAC5E,CAAC;AAEF;;;;GAIG;AACH,MAAM,WAAW,GAAqB;IACpC,EAAE,EAAE,KAAK;IACT,IAAI,EAAE,uBAAuB;IAC7B,aAAa,EAAE,CAAC,UAAU,EAAE,YAAY,EAAE,SAAS,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,CAAC;IACzE,oBAAoB,EAAE,GAAG,EAAE,YAAY;IACvC,MAAM,EAAE;QACN,WAAW,EAAE,IAAI;QACjB,4BAA4B,EAAE,IAAI;QAClC,iBAAiB,EAAE,IAAI;QACvB,cAAc,EAAE,KAAK;QACrB,WAAW,EAAE,GAAG;QAChB,eAAe,EAAE,IAAI;QACrB,eAAe,EAAE,IAAI;QACrB,mBAAmB,EAAE,CAAC;QACtB,SAAS,EAAE,IAAI;QACf,kBAAkB,EAAE,IAAI;QACxB,oBAAoB,EAAE,IAAI;QAC1B,qBAAqB,EAAE,IAAI;QAC3B,kBAAkB,EAAE,EAAE;QACtB,yBAAyB,EAAE,IAAI,EAAE,oCAAoC;QACrE,wBAAwB,EAAE,IAAI;QAC9B,YAAY,EAAE,KAAK;QACnB,UAAU,EAAE,EAAE;QACd,UAAU,EAAE,EAAE;QACd,eAAe,EAAE,MAAM;KACxB;IACD,SAAS,EACP,4EAA4E;QAC5E,2DAA2D;CAC9D,CAAC;AAEF;;;;GAIG;AACH,MAAM,cAAc,GAAqB;IACvC,EAAE,EAAE,QAAQ;IACZ,IAAI,EAAE,wBAAwB;IAC9B,aAAa,EAAE,CAAC,eAAe,EAAE,YAAY,EAAE,gBAAgB,EAAE,cAAc,CAAC;IAChF,oBAAoB,EAAE,GAAG,EAAE,YAAY;IACvC,MAAM,EAAE;QACN,WAAW,EAAE,IAAI;QACjB,4BAA4B,EAAE,IAAI;QAClC,iBAAiB,EAAE,IAAI;QACvB,cAAc,EAAE,KAAK;QACrB,WAAW,EAAE,GAAG;QAChB,eAAe,EAAE,IAAI;QACrB,eAAe,EAAE,IAAI;QACrB,mBAAmB,EAAE,CAAC;QACtB,SAAS,EAAE,IAAI,EAAE,6BAA6B;QAC9C,kBAAkB,EAAE,IAAI;QACxB,oBAAoB,EAAE,IAAI;QAC1B,qBAAqB,EAAE,IAAI;QAC3B,kBAAkB,EAAE,EAAE;QACtB,yBAAyB,EAAE,IAAI;QAC/B,wBAAwB,EAAE,IAAI;QAC9B,YAAY,EAAE,KAAK;QACnB,UAAU,EAAE,EAAE;QACd,UAAU,EAAE,EAAE;QACd,eAAe,EAAE,MAAM;KACxB;IACD,SAAS,EACP,wEAAwE;QACxE,wEAAwE;CAC3E,CAAC;AAEF,+EAA+E;AAC/E,mBAAmB;AACnB,+EAA+E;AAE/E,8DAA8D;AAC9D,MAAM,YAAY,GAAuB;IACvC,cAAc;IACd,WAAW;IACX,cAAc;IACd,iBAAiB;IACjB,mBAAmB;IACnB,cAAc;IACd,aAAa;CACd,CAAC;AAEF,+EAA+E;AAC/E,aAAa;AACb,+EAA+E;AAE/E;;GAEG;AACH,MAAM,UAAU,mBAAmB,CAAC,SAAiB;IACnD,OAAO,YAAY,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,SAAS,CAAC,CAAC;AACtD,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,sBAAsB,CAAC,SAAiB;IACtD,kDAAkD;IAClD,KAAK,MAAM,OAAO,IAAI,YAAY,EAAE,CAAC;QACnC,KAAK,MAAM,OAAO,IAAI,OAAO,CAAC,aAAa,EAAE,CAAC;YAC5C,IAAI,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,CAAC;gBAC5B,OAAO,OAAO,CAAC;YACjB,CAAC;QACH,CAAC;IACH,CAAC;IAED,oDAAoD;IACpD,MAAM,UAAU,GAAG,SAAS,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC;IAChD,IAAI,UAAU,EAAE,CAAC;QACf,MAAM,MAAM,GAAG,QAAQ,CAAC,UAAU,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;QAC3C,IAAI,MAAM,IAAI,CAAC;YAAE,OAAO,mBAAmB,CAAC;QAC5C,IAAI,MAAM,IAAI,EAAE;YAAE,OAAO,cAAc,CAAC;QACxC,OAAO,aAAa,CAAC;IACvB,CAAC;IAED,uDAAuD;IACvD,OAAO,mBAAmB,CAAC;AAC7B,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,kBAAkB,CAChC,SAAiB,EACjB,aAA6C;IAE7C,MAAM,OAAO,GAAG,sBAAsB,CAAC,SAAS,CAAC,CAAC;IAElD,mEAAmE;IACnE,MAAM,MAAM,GAAyB;QACnC,WAAW,EAAE,IAAI;QACjB,4BAA4B,EAAE,IAAI;QAClC,iBAAiB,EAAE,IAAI;QACvB,cAAc,EAAE,IAAI;QACpB,WAAW,EAAE,IAAI;QACjB,eAAe,EAAE,IAAI;QACrB,eAAe,EAAE,IAAI;QACrB,mBAAmB,EAAE,CAAC;QACtB,SAAS,EAAE,KAAK;QAChB,iBAAiB,EAAE,2BAA2B;QAC9C,kBAAkB,EAAE,KAAK;QACzB,oBAAoB,EAAE,KAAK;QAC3B,qBAAqB,EAAE,KAAK;QAC5B,kBAAkB,EAAE,EAAE;QACtB,yBAAyB,EAAE,KAAK;QAChC,wBAAwB,EAAE,KAAK;QAC/B,YAAY,EAAE,KAAK;QACnB,UAAU,EAAE,EAAE;QACd,UAAU,EAAE,EAAE;QACd,eAAe,EAAE,UAAU;QAC3B,gBAAgB;QAChB,GAAG,OAAO,CAAC,MAAM;QACjB,0CAA0C;QAC1C,GAAG,aAAa;KACjB,CAAC;IAEF,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,CAAC;AACrC,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,qBAAqB;IACnC,OAAO,CAAC,GAAG,YAAY,CAAC,CAAC;AAC3B,CAAC"}

package/dist/models/executor.d.ts

@@ -0,0 +1,130 @@
+/**
+ * Task Executor
+ *
+ * Tier 2 of the Multi-Model Architecture.
+ * Responsible for:
+ * - Executing subtasks following planner specifications
+ * - Managing retries and fallbacks
+ * - Tracking execution results and costs
+ */
+import { ExecutionPlan, ExecutionResult, ModelConfig, MultiModelConfig } from './types.js';
+import { ModelRouter } from './router.js';
+import { TaskPlanner } from './planner.js';
+/**
+ * Model client interface for executing prompts
+ */
+export interface ModelClient {
+    complete(model: ModelConfig, prompt: string, options?: {
+        maxTokens?: number;
+        timeout?: number;
+        temperature?: number;
+    }): Promise<{
+        content: string;
+        tokensUsed: {
+            input: number;
+            output: number;
+        };
+        latencyMs: number;
+    }>;
+}
+/**
+ * Execution context passed to model calls
+ */
+export interface ExecutionContext {
+    planId: string;
+    subtaskId: string;
+    inputs: Map<string, string>;
+    constraints: string[];
+    previousAttempts: ExecutionResult[];
+}
+/**
+ * Executor options
+ */
+export interface ExecutorOptions {
+    maxRetries?: number;
+    retryDelayMs?: number;
+    stepTimeout?: number;
+    enableFallback?: boolean;
+    parallelExecution?: boolean;
+    maxParallel?: number;
+}
+/**
+ * Task Executor - executes plans created by the planner
+ */
+export declare class TaskExecutor {
+    private router;
+    readonly config: MultiModelConfig;
+    private client;
+    private options;
+    private results;
+    constructor(router: ModelRouter, config: MultiModelConfig, client: ModelClient, options?: ExecutorOptions);
+    /**
+     * Execute a complete plan
+     */
+    executePlan(plan: ExecutionPlan, planner: TaskPlanner, onProgress?: (result: ExecutionResult) => void): Promise<ExecutionResult[]>;
+    /**
+     * Execute multiple subtasks in parallel
+     */
+    private executeParallel;
+    /**
+     * Execute a single subtask with retry logic
+     */
+    executeSubtask(plan: ExecutionPlan, subtaskId: string, inputs: Map<string, string>): Promise<ExecutionResult>;
+    /**
+     * Attempt to execute a subtask with a specific model
+     */
+    private attemptExecution;
+    /**
+     * Build the prompt for a subtask
+     */
+    private buildPrompt;
+    /**
+     * Get max tokens based on subtask complexity
+     */
+    private getMaxTokens;
+    /**
+     * Create an error result
+     */
+    private createErrorResult;
+    /**
+     * Helper to delay
+     */
+    private delay;
+    /**
+     * Get results for a plan
+     */
+    getResults(planId: string): ExecutionResult[] | undefined;
+    /**
+     * Get total cost for a plan
+     */
+    getTotalCost(planId: string): number;
+    /**
+     * Get success rate for a plan
+     */
+    getSuccessRate(planId: string): number;
+    /**
+     * Generate execution summary
+     */
+    generateSummary(planId: string): string;
+}
+/**
+ * Create an executor instance
+ */
+export declare function createExecutor(router: ModelRouter, config: MultiModelConfig, client: ModelClient, options?: ExecutorOptions): TaskExecutor;
+/**
+ * Mock client for testing
+ */
+export declare class MockModelClient implements ModelClient {
+    private responses;
+    private latency;
+    constructor(responses?: Record<string, string>, latency?: number);
+    complete(model: ModelConfig, prompt: string): Promise<{
+        content: string;
+        tokensUsed: {
+            input: number;
+            output: number;
+        };
+        latencyMs: number;
+    }>;
+}
+//# sourceMappingURL=executor.d.ts.map

package/dist/models/executor.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"executor.d.ts","sourceRoot":"","sources":["../../src/models/executor.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,EAAE,aAAa,EAAE,eAAe,EAAW,WAAW,EAAE,gBAAgB,EAAE,MAAM,YAAY,CAAC;AACpG,OAAO,EAAE,WAAW,EAAE,MAAM,aAAa,CAAC;AAC1C,OAAO,EAAE,WAAW,EAAE,MAAM,cAAc,CAAC;AAE3C;;GAEG;AACH,MAAM,WAAW,WAAW;IAC1B,QAAQ,CACN,KAAK,EAAE,WAAW,EAClB,MAAM,EAAE,MAAM,EACd,OAAO,CAAC,EAAE;QACR,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,OAAO,CAAC,EAAE,MAAM,CAAC;QACjB,WAAW,CAAC,EAAE,MAAM,CAAC;KACtB,GACA,OAAO,CAAC;QACT,OAAO,EAAE,MAAM,CAAC;QAChB,UAAU,EAAE;YAAE,KAAK,EAAE,MAAM,CAAC;YAAC,MAAM,EAAE,MAAM,CAAA;SAAE,CAAC;QAC9C,SAAS,EAAE,MAAM,CAAC;KACnB,CAAC,CAAC;CACJ;AAED;;GAEG;AACH,MAAM,WAAW,gBAAgB;IAC/B,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC5B,WAAW,EAAE,MAAM,EAAE,CAAC;IACtB,gBAAgB,EAAE,eAAe,EAAE,CAAC;CACrC;AAED;;GAEG;AACH,MAAM,WAAW,eAAe;IAC9B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,cAAc,CAAC,EAAE,OAAO,CAAC;IACzB,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB;AAWD;;GAEG;AACH,qBAAa,YAAY;IACvB,OAAO,CAAC,MAAM,CAAc;IAC5B,QAAQ,CAAC,MAAM,EAAE,gBAAgB,CAAC;IAClC,OAAO,CAAC,MAAM,CAAc;IAC5B,OAAO,CAAC,OAAO,CAAkB;IACjC,OAAO,CAAC,OAAO,CAAiC;gBAG9C,MAAM,EAAE,WAAW,EACnB,MAAM,EAAE,gBAAgB,EACxB,MAAM,EAAE,WAAW,EACnB,OAAO,GAAE,eAAoB;IAS/B;;OAEG;IACG,WAAW,CACf,IAAI,EAAE,aAAa,EACnB,OAAO,EAAE,WAAW,EACpB,UAAU,CAAC,EAAE,CAAC,MAAM,EAAE,eAAe,KAAK,IAAI,GAC7C,OAAO,CAAC,eAAe,EAAE,CAAC;IAuC7B;;OAEG;YACW,eAAe;IAyB7B;;OAEG;IACG,cAAc,CAClB,IAAI,EAAE,aAAa,EACnB,SAAS,EAAE,MAAM,EACjB,MAAM,EAAE,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,GAC1B,OAAO,CAAC,eAAe,CAAC;IAiE3B;;OAEG;YACW,gBAAgB;IAsC9B;;OAEG;IACH,OAAO,CAAC,WAAW;IAsEnB;;OAEG;IACH,OAAO,CAAC,YAAY;IAUpB;;OAEG;IACH,OAAO,CAAC,iBAAiB;IAoBzB;;OAEG;IACH,OAAO,CAAC,KAAK;IAIb;;OAEG;IACH,UAAU,CAAC,MAAM,EAAE,MAAM,GAAG,eAAe,EAAE,GAAG,SAAS;IAIzD;;OAEG;IACH,YAAY,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM;IAMpC;;OAEG;IACH,cAAc,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM;IAOtC;;OAEG;IACH,eAAe,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM;CAsDxC;AAED;;GAEG;AACH,wBAAgB,cAAc,CAC5B,MAAM,EAAE,WAAW,EACnB,MAAM,EAAE,gBAAgB,EACxB,MAAM,EAAE,WAAW,EACnB,OAAO,CAAC,EAAE,eAAe,GACxB,YAAY,CAEd;AAED;;GAEG;AACH,qBAAa,eAAgB,YAAW,WAAW;IACjD,OAAO,CAAC,SAAS,CAAsB;IACvC,OAAO,CAAC,OAAO,CAAS;gBAEZ,SAAS,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAAE,OAAO,GAAE,MAAa;IAKhE,QAAQ,CACZ,KAAK,EAAE,WAAW,EAClB,MAAM,EAAE,MAAM,GACb,OAAO,CAAC;QACT,OAAO,EAAE,MAAM,CAAC;QAChB,UAAU,EAAE;YAAE,KAAK,EAAE,MAAM,CAAC;YAAC,MAAM,EAAE,MAAM,CAAA;SAAE,CAAC;QAC9C,SAAS,EAAE,MAAM,CAAC;KACnB,CAAC;CAqBH"}