@warmdrift/kgauto-compiler 2.0.0-alpha.3 → 2.0.0-alpha.31

package/dist/profiles.js

@@ -21,7 +21,9 @@ var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: tru
 var profiles_exports = {};
 __export(profiles_exports, {
   ALIASES: () => ALIASES,
+  _setProfileBrainHook: () => _setProfileBrainHook,
   allProfiles: () => allProfiles,
+  allProfilesRaw: () => allProfilesRaw,
   getProfile: () => getProfile,
   profilesByProvider: () => profilesByProvider,
   tryGetProfile: () => tryGetProfile
@@ -79,7 +81,24 @@ var PROFILES_RAW = [
     ],
     strengths: ["reasoning", "agentic_coding", "long_context", "reliable_tool_use", "structured_output"],
     weaknesses: ["cost", "latency"],
-    notes: "Frontier (2026-05). Step-change improvement over 4.6 in agentic coding. Adaptive thinking only \u2014 no extended-thinking toggle. 1M context, 128k max output."
+    notes: "Frontier (2026-05). Step-change improvement over 4.6 in agentic coding. Adaptive thinking only \u2014 no extended-thinking toggle. 1M context, 128k max output.",
+    // Frontier perf. Drops on archetypes where parallel-tool throughput
+    // (hunt) or low-budget cost-sensitivity (classify/summarize) matters
+    // more than reasoning depth.
+    archetypePerf: {
+      critique: 10,
+      plan: 10,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 8,
+      // strong but Flash dominates parallel tool throughput
+      summarize: 8,
+      // overkill for tolerant archetype; cost-out of frontier
+      classify: 8
+      // overkill; brain-validated cheaper models cover this
+    }
   },
   {
     id: "claude-opus-4-6",
@@ -111,7 +130,20 @@ var PROFILES_RAW = [
     ],
     strengths: ["reasoning", "long_context", "reliable_tool_use", "structured_output", "extended_thinking"],
     weaknesses: ["cost", "latency"],
-    notes: "Predecessor to 4.7. Still current in Anthropic legacy table. Same pricing as 4.7 \u2014 choose 4.7 unless you need extended-thinking budget control (4.7 is adaptive-only)."
+    notes: "Predecessor to 4.7. Still current in Anthropic legacy table. Same pricing as 4.7 \u2014 choose 4.7 unless you need extended-thinking budget control (4.7 is adaptive-only).",
+    // One notch below 4.7 across the board — extended-thinking edge does
+    // not flip any archetype ranking. Legacy: chains should prefer 4.7.
+    archetypePerf: {
+      critique: 9,
+      plan: 9,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 7,
+      summarize: 8,
+      classify: 8
+    }
   },
   {
     id: "claude-sonnet-4-6",
@@ -135,7 +167,23 @@ var PROFILES_RAW = [
     ],
     strengths: ["quality", "tool_use", "long_context", "cache_friendly", "extended_thinking"],
     weaknesses: [],
-    notes: "Workhorse. Best price/quality for most multi-turn agentic work. 1M context, 64k max output."
+    notes: "Workhorse. Best price/quality for most multi-turn agentic work. 1M context, 64k max output.",
+    // Master plan §6.2 anchor. Tier 0 for plan/generate/ask/extract/transform
+    // in starter chains; tier 1 cross-provider for hunt/summarize/classify.
+    archetypePerf: {
+      ask: 9,
+      generate: 9,
+      plan: 9,
+      critique: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 7,
+      // strong but Flash beats on parallel tool throughput
+      summarize: 8,
+      // overkill for tolerant archetype
+      classify: 8
+      // overkill
+    }
   },
   {
     id: "claude-haiku-4-5",
@@ -165,7 +213,23 @@ var PROFILES_RAW = [
     ],
     strengths: ["speed", "cost", "classification", "cache_friendly", "extended_thinking"],
     weaknesses: ["complex_reasoning", "large_tool_sets"],
-    notes: "Cheapest Anthropic. Great for classify, summarize, ask shapes. 200k context, 64k max output. API alias `claude-haiku-4-5` resolves to dated snapshot `claude-haiku-4-5-20251001`."
+    notes: "Cheapest Anthropic. Great for classify, summarize, ask shapes. 200k context, 64k max output. API alias `claude-haiku-4-5` resolves to dated snapshot `claude-haiku-4-5-20251001`.",
+    // Tier 1 cross-provider anchor for short-output chains (classify/
+    // summarize/extract/transform). Falls off on plan/critique where
+    // reasoning depth matters; competes with Pro on cost+latency.
+    archetypePerf: {
+      classify: 8,
+      summarize: 8,
+      ask: 7,
+      transform: 7,
+      extract: 7,
+      hunt: 6,
+      // tool reliability drops at 16 — cliff guard fires
+      generate: 6,
+      plan: 5,
+      critique: 4
+      // reasoning depth gap vs Sonnet/Opus
+    }
   },
   // ── Google ──
   {
@@ -243,7 +307,131 @@ var PROFILES_RAW = [
     ],
     strengths: ["speed", "volume", "classification", "1m_context", "cost"],
     weaknesses: ["complex_schemas", "large_tool_sets", "high_context_quality"],
-    notes: "Fast and cheap with 1M context. Quality cliffs at 8K context and 20 tools \u2014 guard with cliffs."
+    notes: "Fast and cheap with 1M context. Quality cliffs at 8K context and 20 tools \u2014 guard with cliffs.",
+    // 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-2.5-flash-lite",
+    verifiedAgainstDocs: "2026-05-13",
+    provider: "google",
+    status: "current",
+    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: "Bottom-frontier anchor on cost: $0.10/$0.40 per 1M tokens, 1M context, 65K max output. Released July 2025 (stable). Positioned for classify / summarize / transform archetypes where quality bar is forgiving. Cliffs inherited from Flash at equal-or-tighter thresholds \u2014 re-tune per (archetype) once brain has n\u226520 rows. Alpha.8 contract layer handles MAX_TOKENS-on-structured-output via fallback chain, so structuredOutput=native is safe to declare even though Flash had alpha.8 history. Cache discount in spec: $0.01/M = 1/10 of input (richer than Flash 25%) \u2014 meaningful for repeat-prompt workloads.",
+    // 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
+    }
   },
   {
     id: "gemini-2.5-pro",
@@ -279,7 +467,21 @@ var PROFILES_RAW = [
       }
     ],
     strengths: ["reasoning", "1m_context", "structured_output", "tool_use"],
-    weaknesses: ["pricing_above_200k"]
+    weaknesses: ["pricing_above_200k"],
+    // Master plan §3.3 anchor: tier-2 cross-provider in almost every chain.
+    // Sits on the frontier at perf-9 — close to Sonnet but cheaper input.
+    archetypePerf: {
+      critique: 9,
+      plan: 9,
+      ask: 8,
+      generate: 8,
+      extract: 8,
+      transform: 8,
+      hunt: 8,
+      // tier 1 cross-provider for hunt chain
+      summarize: 7,
+      classify: 7
+    }
   },
   {
     id: "gemini-3.1-pro-preview",
@@ -317,7 +519,23 @@ var PROFILES_RAW = [
     ],
     strengths: ["reasoning", "1m_context", "agentic_coding", "structured_output", "tool_use"],
     weaknesses: ["cost", "preview_status", "pricing_above_200k"],
-    notes: "Frontier Gemini (preview, 2026-Q2). Step-change agentic coding per Google. Cache discount 10\xD7 (vs 4\xD7 for 2.5 Pro). Use status=preview to flag rollback path until GA."
+    notes: "Frontier Gemini (preview, 2026-Q2). Step-change agentic coding per Google. Cache discount 10\xD7 (vs 4\xD7 for 2.5 Pro). Use status=preview to flag rollback path until GA.",
+    // Frontier-Gemini preview — bumped one notch over 2.5 Pro on agentic
+    // coding / reasoning per Google's release notes. Preview status:
+    // chains should stay on 2.5 Pro until GA. Starter hypothesis.
+    archetypePerf: {
+      critique: 10,
+      // Google claims step-change on reasoning
+      plan: 10,
+      ask: 9,
+      generate: 9,
+      extract: 9,
+      transform: 8,
+      hunt: 9,
+      // step-change agentic per Google
+      summarize: 8,
+      classify: 7
+    }
   },
   // ── DeepSeek ──
   // 2026-05-08 audit (L-073): DeepSeek's `deepseek-chat` was silently aliased
@@ -357,7 +575,24 @@ var PROFILES_RAW = [
     ],
     strengths: ["cost", "1m_context", "json_output", "code", "reasoning"],
     weaknesses: ["parallel_tools", "large_tool_sets"],
-    notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES."
+    notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES.",
+    // Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
+    // classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
+    // Falls off on hunt (sequential tools — L-040) and reasoning depth.
+    archetypePerf: {
+      classify: 7,
+      // brain-validated, 169 rows
+      summarize: 7,
+      // archetype-tolerant, no brain evidence yet
+      ask: 6,
+      transform: 6,
+      generate: 5,
+      plan: 5,
+      extract: 5,
+      critique: 4,
+      hunt: 4
+      // sequential tool calls only — L-040
+    }
   },
   {
     id: "deepseek-v4-pro",
@@ -393,7 +628,569 @@ var PROFILES_RAW = [
     ],
     strengths: ["quality", "reasoning", "1m_context", "json_output", "code", "extended_thinking"],
     weaknesses: ["parallel_tools", "large_tool_sets"],
-    notes: "Pro tier. 1M context, 384k max output. Regular pricing $1.74/$3.48; 75% promo through 2026-05-31 ($0.435/$0.87). Default mode = thinking."
+    notes: "Pro tier. 1M context, 384k max output. Regular pricing $1.74/$3.48; 75% promo through 2026-05-31 ($0.435/$0.87). Default mode = thinking.",
+    // Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
+    // bumped one notch over V4-Flash; same parallel-tool ceiling.
+    archetypePerf: {
+      plan: 7,
+      // §3.3 tier 3 for plan
+      critique: 6,
+      ask: 7,
+      generate: 6,
+      classify: 7,
+      summarize: 7,
+      extract: 6,
+      transform: 6,
+      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).
+  {
+    // s37 (2026-05-21): UNVERIFIED-AUTO-ONBOARD → verified against
+    // ai.google.dev/gemini-api/docs/models/gemini-3-flash-preview +
+    // ai.google.dev/gemini-api/docs/pricing. L-081 catches:
+    //   maxOutputTokens 65_535 → 65_536 (off-by-one)
+    //   costInputPer1m  0.30 → 0.50  (template-cloned from 2.5-flash; actual is 1.67× more expensive)
+    //   costOutputPer1m 2.50 → 3.00  (template-cloned; actual 1.2× more expensive)
+    //   cache discount  default 0.25 → 0.10 (10× discount, $0.05/$0.50 per docs)
+    // Cliffs inherited from 2.5-flash conservatively. The 8K-context-quality
+    // cliff was a 2.5-Flash observation — Google positions Gemini 3 as
+    // sustained-frontier-on-long-context; brain evidence will validate/relax.
+    // Kept as guard for now.
+    id: "gemini-3-flash-preview",
+    verifiedAgainstDocs: "2026-05-21",
+    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 2.5-flash guard; brain evidence on Gemini 3 long-context quality will validate/relax"
+      },
+      {
+        metric: "tool_count",
+        threshold: 20,
+        action: "drop_to_top_relevant",
+        reason: "Tool reliability drops above ~20 tools (despite 128 hard limit) \u2014 inherited from Flash family"
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking tokens consume maxOutputTokens \u2014 empty response if drained"
+      },
+      {
+        // Inherited from gemini-2.5-flash s11 trust artifact. Family-likely
+        // failure mode for Flash architecture. Keep preemptively until brain
+        // evidence on Gemini 3 specifically.
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Inherited from 2.5-flash s11 cliff: summarize+tools \u2192 empty response. Preemptive guard until brain evidence on 3-flash-preview specifically."
+      }
+    ],
+    costInputPer1m: 0.5,
+    costOutputPer1m: 3,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      // 10× cache discount per Google pricing: $0.05/M cached vs $0.50/M input.
+      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: 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: "Verified s37 (2026-05-21) against Google docs. Step-change positioning vs 2.5-flash on agentic loops per Google's release notes (Dec 2025). Pricing 1.67\xD7/1.2\xD7 higher than 2.5-flash; cache discount 10\xD7 (vs 4\xD7 for 2.5). Status=preview until brain evidence accumulates.",
+    // Anchored to 2.5-flash archetypePerf as starter, with judgment adjustments
+    // for Google's "step-change on agentic" positioning. Brain evidence (zero
+    // rows today) will replace these starter values.
+    archetypePerf: {
+      hunt: 9,
+      // Inherits 2.5-flash L-040 parallel-tool tier; Google positions 3 as agentic-loop upgrade
+      classify: 7,
+      // Inherits 2.5-flash brain-validated tier (218 rows on 2.5)
+      summarize: 7,
+      // Inherits 2.5-flash; cliff strips tools when present
+      transform: 7,
+      ask: 8,
+      // +1 vs 2.5-flash — sustained-frontier positioning
+      generate: 7,
+      // +1 vs 2.5-flash — agentic coding upgrade per Google
+      plan: 6,
+      // +1 vs 2.5-flash — complex iterations per positioning
+      extract: 6,
+      critique: 5
+      // +1 vs 2.5-flash — but still below Sonnet/Opus reasoning floor
+    }
+  },
+  {
+    // ── Gemini 3.1 Flash-Lite ──
+    // Onboarded 2026-05-16 by auto-onboarder; s37 (2026-05-21) verified
+    // against ai.google.dev/gemini-api/docs/pricing.
+    //
+    // L-081 CATCHES (template clone from 2.5-flash-lite was 2.5-3.75× too cheap):
+    //   costInputPer1m  0.10 → 0.25 (template clone undervalued by 2.5×)
+    //   costOutputPer1m 0.40 → 1.50 (template clone undervalued by 3.75×)
+    //
+    // Real 3.1-flash-lite is NOT a cost-equivalent successor to 2.5-flash-lite —
+    // it sits between 2.5-flash-lite ($0.10/$0.40) and 2.5-flash ($0.30/$2.50).
+    // Cache discount 10× verified ($0.025/M cached vs $0.25/M input).
+    //
+    // Cliffs are HYPOTHESIZED from 2.5-flash family; brain evidence pending.
+    id: "gemini-3.1-flash-lite",
+    verifiedAgainstDocs: "2026-05-21",
+    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: 0.25,
+    costOutputPer1m: 1.5,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      // Cache discount 10× (vs Flash 4×) — Google docs s37: $0.025/M cached vs
+      // $0.25/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: ["low_cost", "speed", "volume", "classification", "summarize", "1m_context", "cache_friendly"],
+    weaknesses: ["complex_reasoning", "large_tool_sets", "complex_schemas", "structured_output_unproven", "long_context_quality"],
+    notes: "Verified s37 (2026-05-21) against Google docs. Sits between 2.5-flash-lite (cheaper) and 2.5-flash (more expensive) on cost frontier; 2.5\xD7 more expensive than initial template-clone. Cliffs hypothesized from Flash family \u2014 brain evidence pending.",
+    // Tier 2-3 floor for summarize/classify chains at the new (verified) price
+    // point. ZERO brain rows — values are starter hypotheses anchored to
+    // "smaller sibling of Flash at higher cost than 2.5-flash-lite." 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
+    }
+  },
+  // ─────────────────────────────────────────────────────────────────────────
+  // Gemini 3.5 Flash — hand-onboarded s37 (2026-05-21)
+  //
+  // Google positioning ("Most intelligent for sustained frontier performance
+  // on agentic and coding tasks" / "particularly effective for rapid agentic
+  // loops involving complex coding cycles and iterations") suggests this is
+  // the Flash-family upgrade specifically aimed at hunt-shape workloads.
+  // Pricing 5× input / 3.6× output vs 2.5-flash — material cost premium.
+  // archetypePerf adjusted +1 vs 2.5-flash on ask/generate/plan/critique
+  // (sustained-frontier positioning); hunt held at 9 inherited from L-040
+  // family parallel-tool tier; brain evidence will validate within 50 rows.
+  //
+  // Cliffs inherited conservatively from 2.5-flash. Google's "sustained
+  // frontier on long-context" positioning suggests the 8K cliff may not
+  // apply to 3.5 — keep as guard until brain evidence shows otherwise.
+  //
+  // Specs verified against:
+  //   ai.google.dev/gemini-api/docs/models/gemini-3.5-flash
+  //   ai.google.dev/gemini-api/docs/pricing (Standard tier)
+  // ─────────────────────────────────────────────────────────────────────────
+  {
+    id: "gemini-3.5-flash",
+    verifiedAgainstDocs: "2026-05-21",
+    provider: "google",
+    status: "current",
+    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 2.5-flash guard; Google positions 3.5 as sustained-frontier-on-long-context but brain evidence pending"
+      },
+      {
+        metric: "tool_count",
+        threshold: 20,
+        action: "drop_to_top_relevant",
+        reason: "Inherited from Flash family: tool reliability drops above ~20 (despite 128 hard limit). Validate per (archetype, model) after n\u226520."
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking mode supported per Google docs; same drain risk as 2.5-flash \u2014 thinking tokens consume maxOutputTokens"
+      },
+      {
+        // Inherited from 2.5-flash s11 trust artifact (5/5 empty rate on
+        // tt-intelligence/summarize/gemini-2.5-flash with tools offered).
+        // Family-likely failure mode for Flash architecture across versions.
+        // Keep preemptively until brain evidence on 3.5-flash specifically.
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Inherited from 2.5-flash s11 cliff (kgauto commit 3872832): summarize+tools \u2192 empty response. Preemptive guard until brain evidence on 3.5-flash specifically."
+      }
+    ],
+    costInputPer1m: 1.5,
+    costOutputPer1m: 9,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      // 10× cache discount per Google pricing: $0.15/M cached vs $1.50/M input.
+      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: "Inherited Flash-family pattern: 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: ["agentic_loops", "coding", "1m_context", "parallel_tools", "thinking_mode", "sustained_frontier"],
+    weaknesses: ["cost_vs_2_5_flash", "no_brain_evidence_yet"],
+    notes: "Hand-onboarded s37 (2026-05-21) verified against Google docs. Stable status; positioned as Flash-family upgrade for agentic loops and coding. 5\xD7/3.6\xD7 more expensive than 2.5-flash but Google claims step-change on sustained frontier work. archetypePerf adjustments are judgment-grounded starter hypotheses \u2014 brain evidence will validate within ~50 rows per archetype.",
+    // Starter hypothesis: anchored to 2.5-flash archetypePerf with +1
+    // adjustments where Google's positioning explicitly supports
+    // (agentic/coding/sustained). Hunt held at 9 inherited from L-040 family
+    // parallel-tool tier. Brain evidence will replace.
+    archetypePerf: {
+      hunt: 9,
+      // Inherited from 2.5-flash L-040 parallel-tool tier; Google positions 3.5 as agentic-loop champion
+      classify: 7,
+      // Inherited from 2.5-flash brain-validated tier (218 rows on 2.5)
+      summarize: 7,
+      // Inherited from 2.5-flash; cliff strips tools when present
+      transform: 7,
+      ask: 8,
+      // +1 vs 2.5-flash — sustained-frontier positioning
+      generate: 8,
+      // +1 vs 2.5-flash (6→8) — Google: "complex coding cycles and iterations"
+      plan: 7,
+      // +1 vs 2.5-flash (5→7) — "complex iterations" positioning
+      extract: 7,
+      // +1 vs 2.5-flash — sustained-frontier on structured tasks
+      critique: 5
+      // +1 vs 2.5-flash — but still below Sonnet/Opus reasoning floor
+    }
   }
 ];
 var ALIASES = {
@@ -403,14 +1200,20 @@ var ALIASES = {
   // Legacy kgauto typo — actual API alias is dash-form (alpha.1 had dot).
   "claude-haiku-4.5": "claude-haiku-4-5"
 };
+var brainHook = {};
+function _setProfileBrainHook(hook) {
+  brainHook = hook;
+}
 function canonicalId(id) {
-  return ALIASES[id] ?? id;
+  return brainHook.resolveAlias?.(id) ?? ALIASES[id] ?? id;
 }
 var PROFILE_INDEX = new Map(
   PROFILES_RAW.map((p) => [p.id, p])
 );
 function getProfile(id) {
   const canonical = canonicalId(id);
+  const fromBrain = brainHook.getProfile?.(canonical);
+  if (fromBrain) return fromBrain;
   const p = PROFILE_INDEX.get(canonical);
   if (!p) {
     const known = [...PROFILE_INDEX.keys(), ...Object.keys(ALIASES)].join(", ");
@@ -419,18 +1222,24 @@ function getProfile(id) {
   return p;
 }
 function tryGetProfile(id) {
-  return PROFILE_INDEX.get(canonicalId(id));
+  const canonical = canonicalId(id);
+  return brainHook.getProfile?.(canonical) ?? PROFILE_INDEX.get(canonical);
 }
 function allProfiles() {
   return PROFILES_RAW;
 }
+function allProfilesRaw() {
+  return PROFILES_RAW;
+}
 function profilesByProvider(provider) {
   return PROFILES_RAW.filter((p) => p.provider === provider);
 }
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
   ALIASES,
+  _setProfileBrainHook,
   allProfiles,
+  allProfilesRaw,
   getProfile,
   profilesByProvider,
   tryGetProfile