@warmdrift/kgauto-compiler 2.0.0-alpha.15 → 2.0.0-alpha.16

package/dist/{chunk-SFF5EVTL.mjs → chunk-7MTHFSNY.mjs}

@@ -615,6 +615,215 @@ 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/index.d.mts

@@ -1,5 +1,5 @@
-import { M as ModelProfile, C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, R as RecordInput, O as OracleScore, e as CompileResult, B as BestPracticeAdvisory, f as Provider } from './profiles-DTnIzGsA.mjs';
-export { g as ALIASES, h as CacheStrategy, i as CallAttempt, j as CallError, k as CliffRule, l as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, L as LoweringSpec, m as Message, n as MutationApplied, o as NormalizedTokens, p as PromptSection, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, T as ToolCall, s as ToolDefinition, t as allProfiles, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-DTnIzGsA.mjs';
+import { M as ModelProfile, C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, R as RecordInput, O as OracleScore, e as CompileResult, B as BestPracticeAdvisory, f as Provider } from './profiles-BoLYdl7F.mjs';
+export { g as ALIASES, h as CacheStrategy, i as CallAttempt, j as CallError, k as CliffRule, l as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, L as LoweringSpec, m as Message, n as MutationApplied, o as NormalizedTokens, p as PromptSection, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, T as ToolCall, s as ToolDefinition, t as allProfiles, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-BoLYdl7F.mjs';
 import { IntentArchetypeName } from './dialect.mjs';
 export { ALL_ARCHETYPES, ContextBucket, DIALECT_VERSION, HistoryDepth, INTENT_ARCHETYPES, OutputMode, ShapeSignature, ToolCountBucket, bucketContext, bucketHistory, bucketToolCount, hashShape, isArchetype, learningKey } from './dialect.mjs';
 

package/dist/index.d.ts

@@ -1,5 +1,5 @@
-import { M as ModelProfile, C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, R as RecordInput, O as OracleScore, e as CompileResult, B as BestPracticeAdvisory, f as Provider } from './profiles-D0y6aLk0.js';
-export { g as ALIASES, h as CacheStrategy, i as CallAttempt, j as CallError, k as CliffRule, l as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, L as LoweringSpec, m as Message, n as MutationApplied, o as NormalizedTokens, p as PromptSection, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, T as ToolCall, s as ToolDefinition, t as allProfiles, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-D0y6aLk0.js';
+import { M as ModelProfile, C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, R as RecordInput, O as OracleScore, e as CompileResult, B as BestPracticeAdvisory, f as Provider } from './profiles-CVB2_5C8.js';
+export { g as ALIASES, h as CacheStrategy, i as CallAttempt, j as CallError, k as CliffRule, l as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, L as LoweringSpec, m as Message, n as MutationApplied, o as NormalizedTokens, p as PromptSection, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, T as ToolCall, s as ToolDefinition, t as allProfiles, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-CVB2_5C8.js';
 import { IntentArchetypeName } from './dialect.js';
 export { ALL_ARCHETYPES, ContextBucket, DIALECT_VERSION, HistoryDepth, INTENT_ARCHETYPES, OutputMode, ShapeSignature, ToolCountBucket, bucketContext, bucketHistory, bucketToolCount, hashShape, isArchetype, learningKey } from './dialect.js';
 

package/dist/index.js

@@ -1475,6 +1475,215 @@ 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
@@ -2563,10 +2772,14 @@ var loadChainsFromBrain = createBrainQueryCache({
 // src/fallback.ts
 var STARTER_CHAINS = {
   // Reasoning floor — never degrade. Walk UP on 429 to Opus → cross-provider.
+  // alpha.16: gpt-5.5 appended as third-provider critique floor (frontier-tier,
+  // archetypePerf=9). Cross-provider-tail invariant has somewhere to land when
+  // both Anthropic + Google are unreachable (consumer adds only OpenAI key).
   critique: [
     "claude-opus-4-7",
     "claude-sonnet-4-6",
-    "gemini-2.5-pro"
+    "gemini-2.5-pro",
+    "gpt-5.5"
   ],
   // Reasoning matters — Sonnet primary; walk UP to Opus on 429 (rare exception
   // to "always cheaper"); cross-provider via Pro; DeepSeek Pro as tier 3 floor.
@@ -2577,25 +2790,29 @@ var STARTER_CHAINS = {
     "deepseek-v4-pro"
   ],
   // Quality + cost match. Walk Sonnet → Haiku same-provider, Pro cross,
-  // Flash floor for the open-posture chain.
+  // gpt-5.4-mini as third-provider tail (alpha.16 — closes the mono-Anthropic
+  // gap when consumer has only ANTHROPIC + OPENAI keys; archetypePerf=7).
   generate: [
     "claude-sonnet-4-6",
     "claude-haiku-4-5",
     "gemini-2.5-pro",
-    "gemini-2.5-flash"
+    "gpt-5.4-mini"
   ],
   ask: [
     "claude-sonnet-4-6",
     "claude-haiku-4-5",
     "gemini-2.5-pro",
-    "gemini-2.5-flash"
+    "gpt-5.4-mini"
   ],
   // Structured-output archetype — Flash skipped (alpha.8 MAX_TOKENS cliff),
-  // DeepSeek skipped (no brain evidence). Floor at Haiku.
+  // DeepSeek skipped (no brain evidence). Floor at Haiku. alpha.16: gpt-5.4
+  // appended as third-provider extract floor (archetypePerf=8, native
+  // structured-output support).
   extract: [
     "claude-sonnet-4-6",
     "claude-haiku-4-5",
-    "gemini-2.5-pro"
+    "gemini-2.5-pro",
+    "gpt-5.4"
   ],
   // Forgiving archetype — Sonnet primary but Flash safely floors it.
   transform: [
@@ -2765,6 +2982,47 @@ async function call(ir, opts = {}) {
       }
     }
   }
+  let policyBlockedFiltered;
+  if (opts.policy?.blockedModels && opts.policy.blockedModels.length > 0) {
+    const blocked = new Set(opts.policy.blockedModels);
+    const filtered = [];
+    const dropped = [];
+    for (const t of targetsToTry) {
+      if (blocked.has(t)) {
+        dropped.push(t);
+      } else {
+        filtered.push(t);
+      }
+    }
+    if (dropped.length > 0) {
+      policyBlockedFiltered = dropped;
+      targetsToTry = filtered;
+    }
+    if (targetsToTry.length === 0) {
+      const latencyMs2 = Date.now() - start;
+      await record({
+        handle: initial.handle,
+        tokensIn: 0,
+        tokensOut: 0,
+        latencyMs: latencyMs2,
+        success: false,
+        errorType: "all_blocked_by_policy",
+        promptPreview: extractPromptPreview(ir)
+      });
+      const blockedAttempts = dropped.map((m) => ({
+        model: m,
+        status: "terminal",
+        errorCode: "blocked_by_policy",
+        message: `Skipped \u2014 model ${m} is in CompilePolicy.blockedModels`
+      }));
+      throw new CallError(
+        `call(): all chain targets blocked by CompilePolicy.blockedModels: [${dropped.join(", ")}]`,
+        blockedAttempts,
+        void 0,
+        "all_blocked_by_policy"
+      );
+    }
+  }
   let activeCompile = initial;
   let lastErr;
   const failedProviders = /* @__PURE__ */ new Set();
@@ -2837,7 +3095,8 @@ async function call(ir, opts = {}) {
         servedBy: targetModel,
         fellOverFrom: fellOver ? initial.target : void 0,
         fallbackReason: fellOver ? normalizeFallbackReason(attempts) : void 0,
-        unreachableFiltered
+        unreachableFiltered,
+        policyBlockedFiltered
       };
     }
     attempts.push({
@@ -2866,8 +3125,9 @@ async function call(ir, opts = {}) {
     promptPreview: extractPromptPreview(ir)
   });
   const filteredNote = unreachableFiltered && unreachableFiltered.length > 0 ? ` (also auto-filtered: [${unreachableFiltered.join(", ")}] \u2014 no API key)` : "";
+  const blockedNote = policyBlockedFiltered && policyBlockedFiltered.length > 0 ? ` (also policy-blocked: [${policyBlockedFiltered.join(", ")}])` : "";
   throw new CallError(
-    `call(): all attempts failed${lastErr ? ` \u2014 ${lastErr.errorCode}: ${lastErr.message}` : ""}${filteredNote}`,
+    `call(): all attempts failed${lastErr ? ` \u2014 ${lastErr.errorCode}: ${lastErr.message}` : ""}${filteredNote}${blockedNote}`,
     attempts,
     lastErr?.status,
     lastErr?.errorCode

package/dist/index.mjs

@@ -17,7 +17,7 @@ import {
   getProfile,
   profilesByProvider,
   tryGetProfile
-} from "./chunk-SFF5EVTL.mjs";
+} from "./chunk-7MTHFSNY.mjs";
 
 // src/tokenizer.ts
 var tokenizerImpl = defaultCharBasedCounter;
@@ -1568,10 +1568,14 @@ var loadChainsFromBrain = createBrainQueryCache({
 // src/fallback.ts
 var STARTER_CHAINS = {
   // Reasoning floor — never degrade. Walk UP on 429 to Opus → cross-provider.
+  // alpha.16: gpt-5.5 appended as third-provider critique floor (frontier-tier,
+  // archetypePerf=9). Cross-provider-tail invariant has somewhere to land when
+  // both Anthropic + Google are unreachable (consumer adds only OpenAI key).
   critique: [
     "claude-opus-4-7",
     "claude-sonnet-4-6",
-    "gemini-2.5-pro"
+    "gemini-2.5-pro",
+    "gpt-5.5"
   ],
   // Reasoning matters — Sonnet primary; walk UP to Opus on 429 (rare exception
   // to "always cheaper"); cross-provider via Pro; DeepSeek Pro as tier 3 floor.
@@ -1582,25 +1586,29 @@ var STARTER_CHAINS = {
     "deepseek-v4-pro"
   ],
   // Quality + cost match. Walk Sonnet → Haiku same-provider, Pro cross,
-  // Flash floor for the open-posture chain.
+  // gpt-5.4-mini as third-provider tail (alpha.16 — closes the mono-Anthropic
+  // gap when consumer has only ANTHROPIC + OPENAI keys; archetypePerf=7).
   generate: [
     "claude-sonnet-4-6",
     "claude-haiku-4-5",
     "gemini-2.5-pro",
-    "gemini-2.5-flash"
+    "gpt-5.4-mini"
   ],
   ask: [
     "claude-sonnet-4-6",
     "claude-haiku-4-5",
     "gemini-2.5-pro",
-    "gemini-2.5-flash"
+    "gpt-5.4-mini"
   ],
   // Structured-output archetype — Flash skipped (alpha.8 MAX_TOKENS cliff),
-  // DeepSeek skipped (no brain evidence). Floor at Haiku.
+  // DeepSeek skipped (no brain evidence). Floor at Haiku. alpha.16: gpt-5.4
+  // appended as third-provider extract floor (archetypePerf=8, native
+  // structured-output support).
   extract: [
     "claude-sonnet-4-6",
     "claude-haiku-4-5",
-    "gemini-2.5-pro"
+    "gemini-2.5-pro",
+    "gpt-5.4"
   ],
   // Forgiving archetype — Sonnet primary but Flash safely floors it.
   transform: [
@@ -1770,6 +1778,47 @@ async function call(ir, opts = {}) {
       }
     }
   }
+  let policyBlockedFiltered;
+  if (opts.policy?.blockedModels && opts.policy.blockedModels.length > 0) {
+    const blocked = new Set(opts.policy.blockedModels);
+    const filtered = [];
+    const dropped = [];
+    for (const t of targetsToTry) {
+      if (blocked.has(t)) {
+        dropped.push(t);
+      } else {
+        filtered.push(t);
+      }
+    }
+    if (dropped.length > 0) {
+      policyBlockedFiltered = dropped;
+      targetsToTry = filtered;
+    }
+    if (targetsToTry.length === 0) {
+      const latencyMs2 = Date.now() - start;
+      await record({
+        handle: initial.handle,
+        tokensIn: 0,
+        tokensOut: 0,
+        latencyMs: latencyMs2,
+        success: false,
+        errorType: "all_blocked_by_policy",
+        promptPreview: extractPromptPreview(ir)
+      });
+      const blockedAttempts = dropped.map((m) => ({
+        model: m,
+        status: "terminal",
+        errorCode: "blocked_by_policy",
+        message: `Skipped \u2014 model ${m} is in CompilePolicy.blockedModels`
+      }));
+      throw new CallError(
+        `call(): all chain targets blocked by CompilePolicy.blockedModels: [${dropped.join(", ")}]`,
+        blockedAttempts,
+        void 0,
+        "all_blocked_by_policy"
+      );
+    }
+  }
   let activeCompile = initial;
   let lastErr;
   const failedProviders = /* @__PURE__ */ new Set();
@@ -1842,7 +1891,8 @@ async function call(ir, opts = {}) {
         servedBy: targetModel,
         fellOverFrom: fellOver ? initial.target : void 0,
         fallbackReason: fellOver ? normalizeFallbackReason(attempts) : void 0,
-        unreachableFiltered
+        unreachableFiltered,
+        policyBlockedFiltered
       };
     }
     attempts.push({
@@ -1871,8 +1921,9 @@ async function call(ir, opts = {}) {
     promptPreview: extractPromptPreview(ir)
   });
   const filteredNote = unreachableFiltered && unreachableFiltered.length > 0 ? ` (also auto-filtered: [${unreachableFiltered.join(", ")}] \u2014 no API key)` : "";
+  const blockedNote = policyBlockedFiltered && policyBlockedFiltered.length > 0 ? ` (also policy-blocked: [${policyBlockedFiltered.join(", ")}])` : "";
   throw new CallError(
-    `call(): all attempts failed${lastErr ? ` \u2014 ${lastErr.errorCode}: ${lastErr.message}` : ""}${filteredNote}`,
+    `call(): all attempts failed${lastErr ? ` \u2014 ${lastErr.errorCode}: ${lastErr.message}` : ""}${filteredNote}${blockedNote}`,
     attempts,
     lastErr?.status,
     lastErr?.errorCode

package/dist/{profiles-DTnIzGsA.d.mts → profiles-BoLYdl7F.d.mts}

@@ -521,6 +521,23 @@ interface CallResult {
      * when `noAutoFilter: true`).
      */
     unreachableFiltered?: string[];
+    /**
+     * alpha.16. Models that policy.blockedModels filtering dropped from the
+     * fallback walk. Defense-in-depth at the call() boundary — compile()'s
+     * passScoreTargets already excludes blocked entries from the initial
+     * target + fallbackChain, but if a consumer re-shapes the chain and
+     * threads policy through only partially, this filter catches the gap.
+     *
+     * Resolves TT-40 follow-on `policy-block-not-enforced-on-fallback-chain`
+     * (2026-05-15) where mutations_applied recorded the block intent but
+     * the call walker landed on the blocked model anyway.
+     *
+     * Undefined when no filter ran (no blockedModels set). Populated only
+     * when filter ran AND dropped at least one entry — empty drops are
+     * stored as `undefined` to keep brain telemetry quiet on the common
+     * case.
+     */
+    policyBlockedFiltered?: string[];
 }
 /**
  * Thrown when call() exhausts the fallback chain without success.

package/dist/{profiles-D0y6aLk0.d.ts → profiles-CVB2_5C8.d.ts}

@@ -521,6 +521,23 @@ interface CallResult {
      * when `noAutoFilter: true`).
      */
     unreachableFiltered?: string[];
+    /**
+     * alpha.16. Models that policy.blockedModels filtering dropped from the
+     * fallback walk. Defense-in-depth at the call() boundary — compile()'s
+     * passScoreTargets already excludes blocked entries from the initial
+     * target + fallbackChain, but if a consumer re-shapes the chain and
+     * threads policy through only partially, this filter catches the gap.
+     *
+     * Resolves TT-40 follow-on `policy-block-not-enforced-on-fallback-chain`
+     * (2026-05-15) where mutations_applied recorded the block intent but
+     * the call walker landed on the blocked model anyway.
+     *
+     * Undefined when no filter ran (no blockedModels set). Populated only
+     * when filter ran AND dropped at least one entry — empty drops are
+     * stored as `undefined` to keep brain telemetry quiet on the common
+     * case.
+     */
+    policyBlockedFiltered?: string[];
 }
 /**
  * Thrown when call() exhausts the fallback chain without success.

package/dist/profiles.d.mts

@@ -1,2 +1,2 @@
-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';
+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-BoLYdl7F.mjs';
 import './dialect.mjs';

package/dist/profiles.d.ts

@@ -1,2 +1,2 @@
-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';
+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-CVB2_5C8.js';
 import './dialect.js';

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.16",
   "description": "Prompt compiler + central learning brain for multi-model AI apps. Swap models without rewriting prompts.",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",