@warmdrift/kgauto-compiler 2.0.0-alpha.15 → 2.0.0-alpha.17

package/dist/profiles.d.mts

@@ -1,2 +1,137 @@
-export { g as ALIASES, h as CacheStrategy, k as CliffRule, L as LoweringSpec, M as ModelProfile, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, _ as _setProfileBrainHook, t as allProfiles, x as allProfilesRaw, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-DTnIzGsA.mjs';
-import './dialect.mjs';
+import { f as Provider } from './ir-C3P4gDt0.mjs';
+import { IntentArchetypeName } from './dialect.mjs';
+
+/**
+ * Model profiles — executable knowledge about each provider/model.
+ *
+ * Unlike v1 which carried `known_failures` as prose strings, v2 makes them
+ * executable: cliffs trigger guards, lowering describes the wire format,
+ * recovery handlers describe what to do after specific failures.
+ *
+ * Each profile is the answer to "if I want to call THIS model with THIS
+ * shape of work, what does it need from me, and what should I do when it
+ * fails?"
+ */
+
+type StructuredOutputCapability = 'native' | 'grammar' | 'none';
+type SystemPromptMode = 'inline' | 'separate' | 'as_developer' | 'unsupported';
+type CacheStrategy = 'cache_control' | 'cachedContent' | 'unsupported';
+interface CliffRule {
+    /** What metric triggers this cliff. */
+    metric: 'input_tokens' | 'tool_count' | 'history_turns' | 'thinking_with_short_output';
+    /** Threshold — meaning depends on metric. */
+    threshold: number;
+    /** What action to take when triggered. */
+    action: 'downgrade_quality_warning' | 'drop_to_top_relevant' | 'force_thinking_budget_zero' | 'force_terse_output' | 'escalate_target' | 'strip_tools';
+    /**
+     * Optional: only fire this cliff when the IR's intent.archetype matches.
+     * Used for archetype-specific failure modes (e.g. Gemini Flash returns
+     * empty when summarize is offered tools).
+     */
+    whenIntent?: IntentArchetypeName;
+    /** Human-readable reason for digest reporting. */
+    reason: string;
+}
+interface RecoveryRule {
+    /** What signal triggers recovery. */
+    signal: 'empty_response_after_tool' | 'empty_response' | 'malformed_function_call' | 'rate_limit' | 'model_not_found' | 'context_overflow';
+    /** Action: retry with adjusted params, or escalate to next fallback. */
+    action: 'retry_with_params' | 'escalate' | 'log_only';
+    /** When action=retry_with_params, the param adjustments to apply. */
+    retryParams?: Record<string, unknown>;
+    /** Max retries with this rule. */
+    maxRetries?: number;
+    /** Human-readable reason for digest reporting. */
+    reason: string;
+}
+interface LoweringSpec {
+    /** Where the system prompt goes. */
+    system: {
+        mode: SystemPromptMode;
+        field?: string;
+    };
+    /** Cache strategy + parameters. */
+    cache: {
+        strategy: CacheStrategy;
+        /** Min tokens before caching is worth it (provider rules). */
+        minTokens?: number;
+        /** Discount factor on cached input (0.1 = 10% of normal price). */
+        discount?: number;
+        /** TTL hint in seconds. */
+        ttlSeconds?: number;
+    };
+    /** Tool format identifier — see lower.ts for supported formats. */
+    tools?: {
+        format: 'anthropic' | 'google' | 'openai' | 'deepseek';
+    };
+    /** Thinking config — present iff this model has a thinking knob. */
+    thinking?: {
+        /** Field path on the request. */
+        field: string;
+        /** Default value when caller hasn't specified. */
+        default?: number | 'auto' | 'off';
+    };
+}
+interface ModelProfile {
+    id: string;
+    provider: Provider;
+    status: 'current' | 'preview' | 'legacy';
+    maxContextTokens: number;
+    maxOutputTokens: number;
+    maxTools: number;
+    parallelToolCalls: boolean;
+    structuredOutput: StructuredOutputCapability;
+    systemPromptMode: SystemPromptMode;
+    streaming: boolean;
+    cliffs: CliffRule[];
+    costInputPer1m: number;
+    costOutputPer1m: number;
+    lowering: LoweringSpec;
+    recovery: RecoveryRule[];
+    strengths: string[];
+    weaknesses: string[];
+    notes?: string;
+    verifiedAgainstDocs?: string;
+    /**
+     * Hand-curated per-archetype performance score on a 0-10 scale.
+     *
+     *   10 = frontier on this archetype (e.g. Opus 4.7 on critique)
+     *    8 = strong second tier (Sonnet on plan, Pro on extract)
+     *    7 = competent (Haiku on classify, Flash on hunt)
+     *    5 = acceptable for tolerant archetypes (Flash-Lite on classify)
+     *    3 = degraded (Flash on critique, DeepSeek on hunt)
+     *
+     * Missing archetypes default to `5` (no data, neutral). Each non-default
+     * value should carry a one-line rationale in the profile's note or inline
+     * comment citing brain evidence, family prior, or "starter hypothesis —
+     * verify with telemetry."
+     *
+     * Source today: hand-curated from master plan §3.3 + §6.2 starter tables.
+     * Source tomorrow (alpha.10+): brain `archetype_model_evidence` view.
+     *
+     * Anti-hallucination guardrail (master plan §2.5): when the watcher's
+     * `--audit-fields` flag flags a profile stale (>90 days since
+     * verifiedAgainstDocs), the archetypePerf values get re-audited
+     * alongside capability fields. AI-trained intuition is NOT a valid
+     * source — only docs or brain evidence.
+     *
+     * alpha.9.
+     */
+    archetypePerf?: Partial<Record<IntentArchetypeName, number>>;
+}
+declare const ALIASES: Record<string, string>;
+interface ProfileBrainHook {
+    getProfile?: (canonicalId: string) => ModelProfile | undefined;
+    resolveAlias?: (id: string) => string | undefined;
+}
+/** @internal — called by models-brain.ts at module load. */
+declare function _setProfileBrainHook(hook: ProfileBrainHook): void;
+declare function getProfile(id: string): ModelProfile;
+declare function tryGetProfile(id: string): ModelProfile | undefined;
+declare function allProfiles(): readonly ModelProfile[];
+/** @internal — bundled-only access for adapters that need a non-brain
+ *  fallback baseline (avoids a brain → profiles → brain re-entry). */
+declare function allProfilesRaw(): readonly ModelProfile[];
+declare function profilesByProvider(provider: Provider): readonly ModelProfile[];
+
+export { ALIASES, type CacheStrategy, type CliffRule, type LoweringSpec, type ModelProfile, type RecoveryRule, type StructuredOutputCapability, type SystemPromptMode, _setProfileBrainHook, allProfiles, allProfilesRaw, getProfile, profilesByProvider, tryGetProfile };

package/dist/profiles.d.ts

@@ -1,2 +1,137 @@
-export { g as ALIASES, h as CacheStrategy, k as CliffRule, L as LoweringSpec, M as ModelProfile, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, _ as _setProfileBrainHook, t as allProfiles, x as allProfilesRaw, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-D0y6aLk0.js';
-import './dialect.js';
+import { f as Provider } from './ir-CFHU3BUT.js';
+import { IntentArchetypeName } from './dialect.js';
+
+/**
+ * Model profiles — executable knowledge about each provider/model.
+ *
+ * Unlike v1 which carried `known_failures` as prose strings, v2 makes them
+ * executable: cliffs trigger guards, lowering describes the wire format,
+ * recovery handlers describe what to do after specific failures.
+ *
+ * Each profile is the answer to "if I want to call THIS model with THIS
+ * shape of work, what does it need from me, and what should I do when it
+ * fails?"
+ */
+
+type StructuredOutputCapability = 'native' | 'grammar' | 'none';
+type SystemPromptMode = 'inline' | 'separate' | 'as_developer' | 'unsupported';
+type CacheStrategy = 'cache_control' | 'cachedContent' | 'unsupported';
+interface CliffRule {
+    /** What metric triggers this cliff. */
+    metric: 'input_tokens' | 'tool_count' | 'history_turns' | 'thinking_with_short_output';
+    /** Threshold — meaning depends on metric. */
+    threshold: number;
+    /** What action to take when triggered. */
+    action: 'downgrade_quality_warning' | 'drop_to_top_relevant' | 'force_thinking_budget_zero' | 'force_terse_output' | 'escalate_target' | 'strip_tools';
+    /**
+     * Optional: only fire this cliff when the IR's intent.archetype matches.
+     * Used for archetype-specific failure modes (e.g. Gemini Flash returns
+     * empty when summarize is offered tools).
+     */
+    whenIntent?: IntentArchetypeName;
+    /** Human-readable reason for digest reporting. */
+    reason: string;
+}
+interface RecoveryRule {
+    /** What signal triggers recovery. */
+    signal: 'empty_response_after_tool' | 'empty_response' | 'malformed_function_call' | 'rate_limit' | 'model_not_found' | 'context_overflow';
+    /** Action: retry with adjusted params, or escalate to next fallback. */
+    action: 'retry_with_params' | 'escalate' | 'log_only';
+    /** When action=retry_with_params, the param adjustments to apply. */
+    retryParams?: Record<string, unknown>;
+    /** Max retries with this rule. */
+    maxRetries?: number;
+    /** Human-readable reason for digest reporting. */
+    reason: string;
+}
+interface LoweringSpec {
+    /** Where the system prompt goes. */
+    system: {
+        mode: SystemPromptMode;
+        field?: string;
+    };
+    /** Cache strategy + parameters. */
+    cache: {
+        strategy: CacheStrategy;
+        /** Min tokens before caching is worth it (provider rules). */
+        minTokens?: number;
+        /** Discount factor on cached input (0.1 = 10% of normal price). */
+        discount?: number;
+        /** TTL hint in seconds. */
+        ttlSeconds?: number;
+    };
+    /** Tool format identifier — see lower.ts for supported formats. */
+    tools?: {
+        format: 'anthropic' | 'google' | 'openai' | 'deepseek';
+    };
+    /** Thinking config — present iff this model has a thinking knob. */
+    thinking?: {
+        /** Field path on the request. */
+        field: string;
+        /** Default value when caller hasn't specified. */
+        default?: number | 'auto' | 'off';
+    };
+}
+interface ModelProfile {
+    id: string;
+    provider: Provider;
+    status: 'current' | 'preview' | 'legacy';
+    maxContextTokens: number;
+    maxOutputTokens: number;
+    maxTools: number;
+    parallelToolCalls: boolean;
+    structuredOutput: StructuredOutputCapability;
+    systemPromptMode: SystemPromptMode;
+    streaming: boolean;
+    cliffs: CliffRule[];
+    costInputPer1m: number;
+    costOutputPer1m: number;
+    lowering: LoweringSpec;
+    recovery: RecoveryRule[];
+    strengths: string[];
+    weaknesses: string[];
+    notes?: string;
+    verifiedAgainstDocs?: string;
+    /**
+     * Hand-curated per-archetype performance score on a 0-10 scale.
+     *
+     *   10 = frontier on this archetype (e.g. Opus 4.7 on critique)
+     *    8 = strong second tier (Sonnet on plan, Pro on extract)
+     *    7 = competent (Haiku on classify, Flash on hunt)
+     *    5 = acceptable for tolerant archetypes (Flash-Lite on classify)
+     *    3 = degraded (Flash on critique, DeepSeek on hunt)
+     *
+     * Missing archetypes default to `5` (no data, neutral). Each non-default
+     * value should carry a one-line rationale in the profile's note or inline
+     * comment citing brain evidence, family prior, or "starter hypothesis —
+     * verify with telemetry."
+     *
+     * Source today: hand-curated from master plan §3.3 + §6.2 starter tables.
+     * Source tomorrow (alpha.10+): brain `archetype_model_evidence` view.
+     *
+     * Anti-hallucination guardrail (master plan §2.5): when the watcher's
+     * `--audit-fields` flag flags a profile stale (>90 days since
+     * verifiedAgainstDocs), the archetypePerf values get re-audited
+     * alongside capability fields. AI-trained intuition is NOT a valid
+     * source — only docs or brain evidence.
+     *
+     * alpha.9.
+     */
+    archetypePerf?: Partial<Record<IntentArchetypeName, number>>;
+}
+declare const ALIASES: Record<string, string>;
+interface ProfileBrainHook {
+    getProfile?: (canonicalId: string) => ModelProfile | undefined;
+    resolveAlias?: (id: string) => string | undefined;
+}
+/** @internal — called by models-brain.ts at module load. */
+declare function _setProfileBrainHook(hook: ProfileBrainHook): void;
+declare function getProfile(id: string): ModelProfile;
+declare function tryGetProfile(id: string): ModelProfile | undefined;
+declare function allProfiles(): readonly ModelProfile[];
+/** @internal — bundled-only access for adapters that need a non-brain
+ *  fallback baseline (avoids a brain → profiles → brain re-entry). */
+declare function allProfilesRaw(): readonly ModelProfile[];
+declare function profilesByProvider(provider: Provider): readonly ModelProfile[];
+
+export { ALIASES, type CacheStrategy, type CliffRule, type LoweringSpec, type ModelProfile, type RecoveryRule, type StructuredOutputCapability, type SystemPromptMode, _setProfileBrainHook, allProfiles, allProfilesRaw, getProfile, profilesByProvider, tryGetProfile };

package/dist/profiles.js

@@ -645,6 +645,215 @@ var PROFILES_RAW = [
       // sequential tools — same as V4-Flash
     }
   },
+  // ── OpenAI ──
+  // alpha.16 (2026-05-17): close the half-supported provider gap. env.ts
+  // already registered OPENAI_API_KEY + executeOpenAI + normalizeOpenAILike
+  // + lowerOpenAI all existed; profile entries were missing, so the
+  // alpha.10 auto-filter would mark openai-keyed models reachable but
+  // there were no profiles to filter IN. Half-supported is now fully
+  // supported. PB request `openai-provider-profiles` (2026-05-16).
+  //
+  // Profile data verified against developers.openai.com/api/docs/pricing
+  // + per-model pages 2026-05-17. L-049/L-081 step-zero: no AI-trained
+  // numbers — fetched live from OpenAI's docs. As of 2026-05, OpenAI's
+  // current flagship is gpt-5.5 (2025-12 cutoff); gpt-5.4-{base,mini,nano}
+  // are the workhorse family. gpt-4.1 + gpt-4o are legacy.
+  //
+  // Both 5.5 and 5.4 carry a 272K input-token pricing cliff (2x input,
+  // 1.5x output beyond that). Modeled as a `downgrade_quality_warning`
+  // cliff because it ranks the model down at large-context shapes — the
+  // semantics of "this model is now 2x more expensive" map onto the
+  // existing penalty mechanism. Cost-watcher will catch high-context
+  // spikes empirically; the cliff prevents naive routing into the doubled
+  // pricing zone.
+  {
+    id: "gpt-5.5",
+    verifiedAgainstDocs: "2026-05-17",
+    provider: "openai",
+    status: "current",
+    maxContextTokens: 105e4,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 272e3,
+        action: "downgrade_quality_warning",
+        reason: "OpenAI pricing tier shift: >272K input tokens billed at 2x input + 1.5x output rates"
+      }
+    ],
+    costInputPer1m: 5,
+    costOutputPer1m: 30,
+    lowering: {
+      system: { mode: "inline" },
+      // OpenAI caching is implicit (auto-applied to repeated prefixes
+      // ≥1024 tokens for prompt_tokens_details.cached_tokens). No
+      // wire-format marker. Discount: 10x for cached input ($0.50/$5.00).
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["reasoning", "agentic_coding", "long_context", "structured_output", "reliable_tool_use", "reasoning_effort_knob"],
+    weaknesses: ["cost", "pricing_cliff_at_272k"],
+    notes: "OpenAI frontier (2026-05). 1M context (1.05M total), 128K max output, 2025-12 cutoff. Reasoning effort knob (none/low/medium/high/xhigh). Pricing cliff at 272K input.",
+    // Frontier-tier perf hypothesis. Anchored to Opus 4.7 row (similar
+    // price/positioning). Brain evidence will refine; no telemetry yet.
+    archetypePerf: {
+      critique: 9,
+      plan: 9,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 8,
+      // parallel tool support good but cliff at 272K hurts deep multi-step
+      summarize: 7,
+      // overkill for tolerant archetype
+      classify: 7
+      // overkill; cheaper models cover this
+    }
+  },
+  {
+    id: "gpt-5.4",
+    verifiedAgainstDocs: "2026-05-17",
+    provider: "openai",
+    status: "current",
+    maxContextTokens: 105e4,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 272e3,
+        action: "downgrade_quality_warning",
+        reason: "OpenAI pricing tier shift: >272K input tokens billed at 2x input + 1.5x output rates"
+      }
+    ],
+    costInputPer1m: 2.5,
+    costOutputPer1m: 15,
+    lowering: {
+      system: { mode: "inline" },
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["reasoning", "long_context", "structured_output", "reliable_tool_use"],
+    weaknesses: ["pricing_cliff_at_272k"],
+    notes: "OpenAI workhorse (2026-05). 1M context (1.05M total), 128K max output, 2025-08 cutoff. Pricing cliff at 272K input. Pairs cleanly with Sonnet 4.6 on cost ($2.50/$15.00 vs $3.00/$15.00).",
+    // Anchored to Sonnet 4.6 row (similar price/positioning). Slight
+    // anthropic-side edge on agentic coding per master plan vibe.
+    archetypePerf: {
+      critique: 8,
+      plan: 8,
+      generate: 8,
+      ask: 8,
+      extract: 8,
+      transform: 8,
+      hunt: 7,
+      summarize: 7,
+      classify: 7
+    }
+  },
+  {
+    id: "gpt-5.4-mini",
+    verifiedAgainstDocs: "2026-05-17",
+    provider: "openai",
+    status: "current",
+    maxContextTokens: 4e5,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 0.75,
+    costOutputPer1m: 4.5,
+    lowering: {
+      system: { mode: "inline" },
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["cost", "speed", "agentic_coding", "structured_output", "reliable_tool_use"],
+    weaknesses: ["reasoning_depth"],
+    notes: "OpenAI mini-tier (2026-05). 400K context, 128K max output, 2025-08 cutoff. OpenAI describes as 'strongest mini model for coding, computer use, subagents.' Cache discount 10x ($0.075 input).",
+    // Mini-tier hypothesis. Anchored to Haiku 4.5 + Flash row pricing.
+    // Cost is slightly higher than Haiku ($0.75 vs $0.50 input) but
+    // OpenAI claims strong coding/subagent perf.
+    archetypePerf: {
+      ask: 7,
+      generate: 7,
+      extract: 7,
+      transform: 7,
+      classify: 7,
+      summarize: 7,
+      hunt: 7,
+      plan: 6,
+      critique: 5
+      // reasoning depth gap — frontier models handle this
+    }
+  },
+  {
+    id: "gpt-5.4-nano",
+    verifiedAgainstDocs: "2026-05-17",
+    provider: "openai",
+    status: "current",
+    maxContextTokens: 4e5,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 0.2,
+    costOutputPer1m: 1.25,
+    lowering: {
+      system: { mode: "inline" },
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["cost", "speed", "volume", "structured_output"],
+    weaknesses: ["reasoning_depth", "no_computer_use"],
+    notes: "OpenAI nano-tier (2026-05). 400K context, 128K max output, 2025-08 cutoff. 'Cheapest GPT-5.4-class for simple high-volume tasks.' No fine-tuning, no computer-use tools. Cache discount 10x.",
+    // Nano-tier. Anchored to Flash-Lite row ($0.10/$0.40 vs nano's
+    // $0.20/$1.25). Slightly more expensive than Flash-Lite but with
+    // OpenAI brand reliability. Good fit for classify/summarize floor.
+    archetypePerf: {
+      classify: 7,
+      summarize: 6,
+      ask: 6,
+      transform: 6,
+      extract: 6,
+      generate: 5,
+      hunt: 5,
+      plan: 4,
+      critique: 3
+      // not for reasoning archetypes
+    }
+  },
   // ── Auto-onboarded (UNVERIFIED) ──
   // Cloned by scripts/auto-onboard-models.mjs from a same-family template.
   // Each entry's pricing/context/cliffs/lowering reflects the template, NOT

package/dist/profiles.mjs

@@ -6,7 +6,7 @@ import {
   getProfile,
   profilesByProvider,
   tryGetProfile
-} from "./chunk-SFF5EVTL.mjs";
+} from "./chunk-7MTHFSNY.mjs";
 export {
   ALIASES,
   _setProfileBrainHook,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@warmdrift/kgauto-compiler",
-  "version": "2.0.0-alpha.15",
+  "version": "2.0.0-alpha.17",
   "description": "Prompt compiler + central learning brain for multi-model AI apps. Swap models without rewriting prompts.",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",
@@ -20,11 +20,16 @@
       "types": "./dist/profiles.d.ts",
       "import": "./dist/profiles.mjs",
       "require": "./dist/profiles.js"
+    },
+    "./glassbox": {
+      "types": "./dist/glassbox/index.d.ts",
+      "import": "./dist/glassbox/index.mjs",
+      "require": "./dist/glassbox/index.js"
     }
   },
   "files": ["dist", "README.md"],
   "scripts": {
-    "build": "tsup src/index.ts src/dialect.ts src/profiles.ts --format cjs,esm --dts --clean",
+    "build": "tsup src/index.ts src/dialect.ts src/profiles.ts src/glassbox/index.ts --format cjs,esm --dts --clean",
     "test": "vitest run",
     "test:watch": "vitest",
     "typecheck": "tsc --noEmit",