@warmdrift/kgauto-compiler 2.0.0-alpha.14 → 2.0.0-alpha.16

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

@@ -614,6 +614,420 @@ var PROFILES_RAW = [
       hunt: 4
       // 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
+  // provider docs. Verify before promoting status to 'current' (L-049/L-081).
+  {
+    id: "gemini-3-flash-preview",
+    verifiedAgainstDocs: "UNVERIFIED-AUTO-ONBOARD",
+    provider: "google",
+    status: "preview",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 65535,
+    maxTools: 128,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "separate",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 8e3,
+        action: "downgrade_quality_warning",
+        reason: "Quality degrades significantly above ~8K context tokens"
+      },
+      {
+        metric: "tool_count",
+        threshold: 20,
+        action: "drop_to_top_relevant",
+        reason: "Tool reliability drops above ~20 tools (despite 128 hard limit)"
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking tokens consume maxOutputTokens \u2014 empty response if drained"
+      },
+      {
+        // s11 trust artifact (2026-05-02): brain showed 5/5 empty rate on
+        // tt-intelligence/summarize/gemini-2.5-flash with tools offered.
+        // v1's disable_thinking_for_short_output already fired and didn't
+        // help — disabling thinking is necessary but not sufficient. Tools
+        // present + summarize intent confuses Flash into a no-output state
+        // (likely tool-decision purgatory). Strip tools entirely for this
+        // archetype on this model.
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Gemini Flash returns empty when summarize intent has tools offered (5/5 empty rate observed in v1 prod 2026-04-19, replayed into v2 brain 2026-04-29)"
+      }
+    ],
+    costInputPer1m: 0.3,
+    costOutputPer1m: 2.5,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      thinking: { field: "generationConfig.thinkingConfig.thinkingBudget", default: "auto" }
+    },
+    recovery: [
+      {
+        signal: "empty_response_after_tool",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Known: empty after tool result \u2014 retry with thinking off"
+      },
+      {
+        signal: "empty_response",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Empty response \u2014 try with thinking off"
+      },
+      {
+        signal: "malformed_function_call",
+        action: "escalate",
+        reason: "MALFORMED_FUNCTION_CALL maps to stop \u2014 escalate to next target"
+      }
+    ],
+    strengths: ["speed", "volume", "classification", "1m_context", "cost"],
+    weaknesses: ["complex_schemas", "large_tool_sets", "high_context_quality"],
+    notes: "AUTO-ONBOARDED 2026-05-16 from `gemini-2.5-flash`. Pricing, context, cliffs are template-cloned and UNVERIFIED \u2014 confirm against provider docs before promoting status to 'current'.",
+    // Master plan §6.2 anchor. Tier 0 for hunt (parallel tool throughput
+    // 15-75 calls/step beats Sonnet — L-040), summarize, classify.
+    archetypePerf: {
+      hunt: 9,
+      // L-040: parallel tool throughput 15-75/step
+      classify: 7,
+      // brain-validated, 218 rows
+      summarize: 7,
+      // brain-validated; cliff strips tools when present
+      transform: 7,
+      ask: 7,
+      generate: 6,
+      plan: 5,
+      extract: 6,
+      // alpha.8 MAX_TOKENS history on structured output
+      critique: 4
+      // reasoning shallower than Sonnet/Opus
+    }
+  },
+  {
+    // ── Gemini 2.5 Flash-Lite ──
+    // Onboarded 2026-05-13 (s22) after the model-release watcher surfaced
+    // it as a UNREGISTERED + NEW candidate. Released by Google July 2025,
+    // stable. Positioned BELOW Flash on the cost/perf frontier:
+    //   input  $0.10/M (Flash $0.30/M)   — 3× cheaper
+    //   output $0.40/M (Flash $2.50/M)   — 6× cheaper
+    //   cache  $0.01/M                    — 1/10 of input (vs Flash 0.25 discount)
+    // Cliffs are HYPOTHESIZED from Flash's known failure modes — Flash-Lite
+    // is a smaller sibling, so we inherit Flash's cliffs at equal-or-tighter
+    // thresholds. The brain will validate/relax these as evidence accumulates
+    // per (archetype, model) tuple. Currently ZERO brain rows for this model.
+    id: "gemini-3.1-flash-lite",
+    verifiedAgainstDocs: "UNVERIFIED-AUTO-ONBOARD",
+    provider: "google",
+    status: "preview",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 65536,
+    maxTools: 128,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "separate",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 8e3,
+        action: "downgrade_quality_warning",
+        reason: "Inherited from Flash: quality degrades above ~8K. Smaller model \u2014 likely degrades faster. Re-tune from brain after n\u226520."
+      },
+      {
+        metric: "tool_count",
+        threshold: 10,
+        action: "drop_to_top_relevant",
+        reason: "Conservative: Flash drops at 20, Flash-Lite is smaller \u2014 assume tighter ceiling until brain proves otherwise."
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking enabled per Google API (thinking: true). Same drain risk as Flash \u2014 thinking tokens consume maxOutputTokens."
+      },
+      {
+        // Strong prior: Flash hit 5/5 empty rate on summarize+tools (s11
+        // trust artifact, kgauto commit 3872832). Flash-Lite shares the
+        // same architectural family — almost certainly inherits this cliff.
+        // Ship the guard preemptively; brain telemetry confirms or relaxes.
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Inherited from Flash s11 cliff: summarize+tools \u2192 empty response. Preemptive guard until brain evidence on Flash-Lite specifically."
+      }
+    ],
+    costInputPer1m: 0.1,
+    costOutputPer1m: 0.4,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      // Cache discount 10× (vs Flash 4×) — Google's spec is $0.01/M cache vs
+      // $0.10/M input. Material for repeat-prompt workloads (classify shape).
+      cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
+      thinking: { field: "generationConfig.thinkingConfig.thinkingBudget", default: "auto" }
+    },
+    recovery: [
+      {
+        signal: "empty_response_after_tool",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Known on Flash family: empty after tool result \u2014 retry with thinking off."
+      },
+      {
+        signal: "empty_response",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Empty response \u2014 try with thinking off."
+      },
+      {
+        signal: "malformed_function_call",
+        action: "escalate",
+        reason: "MALFORMED_FUNCTION_CALL maps to stop \u2014 escalate to next target."
+      }
+    ],
+    strengths: ["lowest_cost", "speed", "volume", "classification", "summarize", "1m_context", "cache_friendly"],
+    weaknesses: ["complex_reasoning", "large_tool_sets", "complex_schemas", "structured_output_unproven", "long_context_quality"],
+    notes: "AUTO-ONBOARDED 2026-05-16 from `gemini-2.5-flash-lite`. Pricing, context, cliffs are template-cloned and UNVERIFIED \u2014 confirm against provider docs before promoting status to 'current'.",
+    // Tier 3 emergency floor for summarize/classify chains. ZERO brain
+    // rows — all values are starter hypotheses anchored to "smaller
+    // sibling of Flash, at-or-below Flash perf on every archetype." The
+    // first 50 brain rows per archetype will validate or relax these.
+    archetypePerf: {
+      classify: 6,
+      // starter hypothesis — verify (Flash is 7, lite likely ≤)
+      summarize: 6,
+      // starter hypothesis — verify; cliff strips tools
+      transform: 6,
+      // starter hypothesis — verify
+      ask: 5,
+      hunt: 5,
+      generate: 4,
+      extract: 4,
+      plan: 3,
+      critique: 3
+    }
   }
 ];
 var ALIASES = {

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';