@warmdrift/kgauto-compiler 2.0.0-alpha.26 → 2.0.0-alpha.28

package/dist/glassbox-routes/index.js

@@ -58,6 +58,1658 @@ function checkAuth(req, config) {
   return null;
 }
 
+// src/profiles.ts
+var ANTHROPIC_LOWERING_BASE = {
+  system: { mode: "inline" },
+  cache: {
+    strategy: "cache_control",
+    minTokens: 1024,
+    discount: 0.1,
+    ttlSeconds: 300
+  },
+  tools: { format: "anthropic" }
+};
+var GOOGLE_LOWERING_BASE = {
+  system: { mode: "separate", field: "systemInstruction" },
+  cache: {
+    strategy: "cachedContent",
+    minTokens: 4096,
+    discount: 0.25,
+    ttlSeconds: 3600
+  },
+  tools: { format: "google" }
+};
+var PROFILES_RAW = [
+  // ── Anthropic ──
+  {
+    id: "claude-opus-4-7",
+    verifiedAgainstDocs: "2026-05-08",
+    provider: "anthropic",
+    status: "current",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 5,
+    costOutputPer1m: 25,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      {
+        signal: "rate_limit",
+        action: "escalate",
+        reason: "429 from Anthropic \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", "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.",
+    // 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",
+    verifiedAgainstDocs: "2026-05-08",
+    provider: "anthropic",
+    status: "legacy",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 5,
+    costOutputPer1m: 25,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      {
+        signal: "rate_limit",
+        action: "escalate",
+        reason: "429 from Anthropic \u2014 escalate to fallback chain"
+      },
+      {
+        signal: "model_not_found",
+        action: "escalate",
+        reason: "Model deprecated/renamed \u2014 escalate (L-061)"
+      }
+    ],
+    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).",
+    // 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",
+    verifiedAgainstDocs: "2026-05-08",
+    provider: "anthropic",
+    status: "current",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 64e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 3,
+    costOutputPer1m: 15,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    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.",
+    // 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",
+    verifiedAgainstDocs: "2026-05-08",
+    provider: "anthropic",
+    status: "current",
+    maxContextTokens: 2e5,
+    maxOutputTokens: 64e3,
+    maxTools: 32,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "tool_count",
+        threshold: 16,
+        action: "drop_to_top_relevant",
+        reason: "Haiku reliability degrades above ~16 tools"
+      }
+    ],
+    costInputPer1m: 1,
+    costOutputPer1m: 5,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to Sonnet" }
+    ],
+    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`.",
+    // 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 ──
+  {
+    id: "gemini-2.5-flash",
+    verifiedAgainstDocs: "2026-05-08",
+    provider: "google",
+    status: "current",
+    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: "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",
+    verifiedAgainstDocs: "2026-05-08",
+    provider: "google",
+    status: "current",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 65535,
+    maxTools: 128,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "separate",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 2e5,
+        action: "downgrade_quality_warning",
+        reason: "Pricing doubles above 200K: input $1.25\u2192$2.50/M, output $10\u2192$15/M"
+      }
+    ],
+    costInputPer1m: 1.25,
+    costOutputPer1m: 10,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      thinking: { field: "generationConfig.thinkingConfig.thinkingBudget", default: "auto" }
+    },
+    recovery: [
+      {
+        signal: "malformed_function_call",
+        action: "escalate",
+        reason: "MALFORMED_FUNCTION_CALL \u2014 escalate"
+      }
+    ],
+    strengths: ["reasoning", "1m_context", "structured_output", "tool_use"],
+    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",
+    verifiedAgainstDocs: "2026-05-08",
+    provider: "google",
+    status: "preview",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 65535,
+    maxTools: 128,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "separate",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 2e5,
+        action: "downgrade_quality_warning",
+        reason: "Pricing doubles above 200K: input $2\u2192$4/M, output $12\u2192$18/M"
+      }
+    ],
+    costInputPer1m: 2,
+    costOutputPer1m: 12,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
+      thinking: { field: "generationConfig.thinkingConfig.thinkingBudget", default: "auto" }
+    },
+    recovery: [
+      {
+        signal: "malformed_function_call",
+        action: "escalate",
+        reason: "MALFORMED_FUNCTION_CALL \u2014 escalate"
+      }
+    ],
+    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.",
+    // 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
+  // to `deepseek-v4-flash` non-thinking mode. Old kgauto profile claimed 64k
+  // context + $0.27/$1.10 — actual is 1M context + $0.14/$0.28. Now modeled
+  // as: V4-Flash + V4-Pro as canonical profiles; deepseek-chat and
+  // deepseek-reasoner registered as aliases (see ALIASES below).
+  {
+    id: "deepseek-v4-flash",
+    verifiedAgainstDocs: "2026-05-08",
+    provider: "deepseek",
+    status: "current",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 384e3,
+    maxTools: 16,
+    parallelToolCalls: false,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "tool_count",
+        threshold: 1,
+        action: "drop_to_top_relevant",
+        reason: "Sequential tool calls only \u2014 L-040"
+      }
+    ],
+    costInputPer1m: 0.14,
+    costOutputPer1m: 0.28,
+    lowering: {
+      system: { mode: "inline" },
+      cache: { strategy: "unsupported" },
+      tools: { format: "deepseek" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" }
+    ],
+    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.",
+    // 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",
+    verifiedAgainstDocs: "2026-05-08",
+    provider: "deepseek",
+    status: "current",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 384e3,
+    maxTools: 16,
+    parallelToolCalls: false,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "tool_count",
+        threshold: 1,
+        action: "drop_to_top_relevant",
+        reason: "Sequential tool calls only \u2014 L-040"
+      }
+    ],
+    // Profile carries REGULAR pricing, not the 75%-off promo (ends 2026-05-31).
+    // Under-estimating cost is worse than over-estimating for budget caps.
+    costInputPer1m: 1.74,
+    costOutputPer1m: 3.48,
+    lowering: {
+      system: { mode: "inline" },
+      cache: { strategy: "unsupported" },
+      tools: { format: "deepseek" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" }
+    ],
+    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.",
+    // 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 = {
+  // DeepSeek's own model routing — both names served by V4-Flash.
+  "deepseek-chat": "deepseek-v4-flash",
+  "deepseek-reasoner": "deepseek-v4-flash",
+  // Legacy kgauto typo — actual API alias is dash-form (alpha.1 had dot).
+  "claude-haiku-4.5": "claude-haiku-4-5"
+};
+var brainHook = {};
+function canonicalId(id) {
+  return brainHook.resolveAlias?.(id) ?? ALIASES[id] ?? id;
+}
+var PROFILE_INDEX = new Map(
+  PROFILES_RAW.map((p) => [p.id, p])
+);
+function tryGetProfile(id) {
+  const canonical = canonicalId(id);
+  return brainHook.getProfile?.(canonical) ?? PROFILE_INDEX.get(canonical);
+}
+
+// src/env.ts
+var SUPPORTED_PROVIDERS = Object.freeze([
+  "anthropic",
+  "google",
+  "openai",
+  "deepseek"
+]);
+function isSupportedProvider(p) {
+  return SUPPORTED_PROVIDERS.includes(p);
+}
+var PROVIDER_ENV_KEYS = Object.freeze({
+  anthropic: Object.freeze(["ANTHROPIC_API_KEY"]),
+  google: Object.freeze([
+    "GOOGLE_API_KEY",
+    "GEMINI_API_KEY",
+    "GOOGLE_GENERATIVE_AI_API_KEY"
+  ]),
+  openai: Object.freeze(["OPENAI_API_KEY"]),
+  deepseek: Object.freeze(["DEEPSEEK_API_KEY"])
+});
+function defaultEnv() {
+  return typeof process !== "undefined" && process.env ? process.env : {};
+}
+function readKeyValue(raw) {
+  if (raw === void 0) return void 0;
+  const trimmed = raw.trim();
+  return trimmed.length > 0 ? trimmed : void 0;
+}
+function resolveProviderKey(provider, opts = {}) {
+  if (!isSupportedProvider(provider)) return void 0;
+  const explicit = readKeyValue(opts.apiKeys?.[provider]);
+  if (explicit) return explicit;
+  const env = opts.envSource ?? defaultEnv();
+  for (const name of PROVIDER_ENV_KEYS[provider]) {
+    const v = readKeyValue(env[name]);
+    if (v) return v;
+  }
+  return void 0;
+}
+function isProviderReachable(provider, opts = {}) {
+  return resolveProviderKey(provider, opts) !== void 0;
+}
+function isModelReachable(modelId, opts = {}) {
+  const profile = tryGetProfile(modelId);
+  if (!profile) return false;
+  return isProviderReachable(profile.provider, opts);
+}
+
+// src/brain-query.ts
+var FRESH_SNAPSHOT = {
+  data: null,
+  expiresAt: 0,
+  refreshing: false,
+  warned: false
+};
+var snapshot = { ...FRESH_SNAPSHOT };
+var runtime;
+function createBrainQueryCache(opts) {
+  return () => {
+    const rt = runtime;
+    if (!rt || !rt.enabledTables.has(opts.table)) {
+      return opts.bundledFallback();
+    }
+    const now = Date.now();
+    const stale = snapshot.expiresAt <= now;
+    if (stale && !snapshot.refreshing) {
+      snapshot.refreshing = true;
+      void asyncRefresh(rt);
+    }
+    if (snapshot.data) {
+      const rows = snapshot.data[opts.table];
+      if (Array.isArray(rows) && rows.length > 0) {
+        try {
+          return opts.mapRows(rows);
+        } catch {
+          return opts.bundledFallback();
+        }
+      }
+    }
+    return opts.bundledFallback();
+  };
+}
+var pendingRefresh;
+async function asyncRefresh(rt) {
+  const promise = doRefresh(rt);
+  pendingRefresh = promise;
+  try {
+    await promise;
+  } finally {
+    if (pendingRefresh === promise) pendingRefresh = void 0;
+  }
+}
+var DEFAULT_CONFIG_URL = "https://kgauto-dashboard.vercel.app/api/kgauto-v2/config";
+async function doRefresh(rt) {
+  const url = rt.configEndpoint ?? DEFAULT_CONFIG_URL;
+  try {
+    const res = await rt.fetchImpl(url, { method: "GET" });
+    if (!res.ok) {
+      throw new Error(`brain-query ${res.status}: ${res.statusText}`);
+    }
+    const body = await res.json();
+    if (runtime !== rt) return;
+    snapshot = {
+      data: body,
+      expiresAt: Date.now() + rt.ttlMs,
+      refreshing: false,
+      warned: snapshot.warned
+    };
+  } catch (err) {
+    if (runtime !== rt) return;
+    snapshot.refreshing = false;
+    snapshot.expiresAt = Date.now() + rt.ttlMs;
+    if (!snapshot.warned) {
+      snapshot.warned = true;
+      (rt.onError ?? defaultOnError)(err);
+    }
+  }
+}
+function defaultOnError(err) {
+  console.warn("[kgauto] brain-query failed (using bundled fallback):", err);
+}
+
+// src/chains-brain.ts
+function isChainsRow(x) {
+  if (!x || typeof x !== "object") return false;
+  const r = x;
+  return typeof r.archetype === "string" && typeof r.tier === "number" && typeof r.model_id === "string";
+}
+function mapRowsToChains(rows) {
+  const grouped = /* @__PURE__ */ new Map();
+  for (const row of rows) {
+    if (!isChainsRow(row)) continue;
+    const list = grouped.get(row.archetype) ?? [];
+    list.push(row);
+    grouped.set(row.archetype, list);
+  }
+  const out = {};
+  for (const [archetype, group] of grouped.entries()) {
+    group.sort((a, b) => a.tier - b.tier);
+    out[archetype] = group.map((r) => r.model_id);
+  }
+  const bundled = getAllStarterChains();
+  for (const archetype of Object.keys(bundled)) {
+    if (!out[archetype]) out[archetype] = bundled[archetype];
+  }
+  return out;
+}
+var loadChainsFromBrain = createBrainQueryCache({
+  table: "kgauto_chains",
+  mapRows: mapRowsToChains,
+  bundledFallback: getAllStarterChains
+});
+
+// src/fallback.ts
+var STARTER_CHAINS_GROUNDED = {
+  // Reasoning floor — never degrade. Walk UP on 429 to Opus → cross-provider.
+  critique: [
+    { id: "claude-opus-4-7", grounding: "judgment", reason: "Highest reasoning bar, no degradation tier \u2014 engineer pick, awaiting measured backing" },
+    { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Same-provider walk-down from Opus on 429" },
+    { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor in similar quality bracket" },
+    { id: "gpt-5.5", grounding: "judgment", reason: "alpha.16: third-provider frontier-tier floor (archetypePerf=9)" }
+  ],
+  // Reasoning matters — Sonnet primary; walk UP to Opus on 429.
+  plan: [
+    { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Reasoning + cost balance \u2014 engineer pick" },
+    { id: "claude-opus-4-7", grounding: "judgment", reason: 'Same-provider walk-UP on 429 (rare exception to "always cheaper")' },
+    { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor" },
+    { id: "deepseek-v4-pro", grounding: "judgment", reason: "Tier 3 cost floor \u2014 no brain evidence yet" }
+  ],
+  // Quality + cost match.
+  generate: [
+    { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Quality + cost match \u2014 engineer pick" },
+    { id: "claude-haiku-4-5", grounding: "judgment", reason: "Same-provider step-down" },
+    { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor" },
+    { id: "gpt-5.4-mini", grounding: "judgment", reason: "alpha.16: third-provider tail (archetypePerf=7) \u2014 closes mono-Anthropic gap" }
+  ],
+  // ask::sonnet — STARTER_CHAINS calls this "Quality + cost match" but
+  // tt-intel s78 prod data showed 27% empty rate. Labeled 'judgment' until
+  // evidence either validates or refutes the placement.
+  ask: [
+    { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Quality + cost match \u2014 engineer pick. NOTE: tt-intel s78 prod showed 27% empty rate; placement awaits measurement validation" },
+    { id: "claude-haiku-4-5", grounding: "judgment", reason: "Same-provider step-down" },
+    { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor" },
+    { id: "gpt-5.4-mini", grounding: "judgment", reason: "alpha.16: third-provider tail (archetypePerf=7)" }
+  ],
+  // Structured-output archetype — Flash skipped (alpha.8 MAX_TOKENS cliff,
+  // capability-fact); DeepSeek skipped (no brain evidence).
+  extract: [
+    { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Reliable structured-output anchor \u2014 engineer pick" },
+    { id: "claude-haiku-4-5", grounding: "judgment", reason: "Same-provider step-down with native structured output" },
+    { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor with structured-output support" },
+    { id: "gpt-5.4", grounding: "capability-fact", reason: "alpha.16: third-provider floor \u2014 native structured-output capability (archetypePerf=8)" }
+  ],
+  // Forgiving archetype — Sonnet primary but Flash safely floors it.
+  transform: [
+    { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Quality anchor \u2014 engineer pick" },
+    { id: "claude-haiku-4-5", grounding: "judgment", reason: "Same-provider step-down" },
+    { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor" },
+    { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost floor \u2014 forgiving archetype tolerates Flash" }
+  ],
+  // Parallel-tool throughput champion — Flash leads on the L-040 cliff
+  // (capability-fact: Flash 15-75 parallel calls/step vs DeepSeek 7-8).
+  hunt: [
+    { id: "gemini-2.5-flash", grounding: "capability-fact", reason: "L-040 parallel-tool throughput champion (15-75 calls/step)" },
+    { id: "gemini-2.5-pro", grounding: "capability-fact", reason: "Cross-provider tier 1 with strong parallel-tool support" },
+    { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Quality safety net for blocked-Flash case" },
+    { id: "claude-haiku-4-5", grounding: "judgment", reason: "Reduced tool budget \u2014 cliff at 16 fires" }
+  ],
+  // Cost-sensitive + tolerant. DeepSeek brain-evidence tier 1.
+  summarize: [
+    { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost-sensitive primary \u2014 engineer pick" },
+    { id: "deepseek-v4-flash", grounding: "measured", reason: "Brain-validated tier 1 for cost-sensitive summarize workloads", n: 169 },
+    { id: "claude-haiku-4-5", grounding: "judgment", reason: "Quality safety net" },
+    { id: "gemini-2.5-flash-lite", grounding: "judgment", reason: "Emergency floor \u2014 onboarded s22, no brain evidence yet" }
+  ],
+  // Brain-validated DeepSeek tier 1 (169 rows, 0% empty rate).
+  classify: [
+    { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost-sensitive primary \u2014 engineer pick" },
+    { id: "deepseek-v4-flash", grounding: "measured", reason: "Brain-validated tier 1 (169 rows, 0% empty rate)", n: 169 },
+    { id: "claude-haiku-4-5", grounding: "judgment", reason: "Quality safety net" },
+    { id: "gemini-2.5-flash-lite", grounding: "judgment", reason: "Cache-discount 10\xD7 floor for repeat-prompt workloads" }
+  ]
+};
+var STARTER_CHAINS = (() => {
+  const out = {};
+  for (const [archetype, entries] of Object.entries(STARTER_CHAINS_GROUNDED)) {
+    out[archetype] = entries.map((e) => e.id);
+  }
+  return out;
+})();
+var STARTER_CHAINS_BY_MODE_GROUNDED = {
+  hunt: {
+    sequential: [
+      {
+        id: "deepseek-v4-pro",
+        grounding: "judgment",
+        reason: "alpha.20 E3: cheap + good reasoning at single-step granularity; L-040 cliff silenced when sequential \u2014 hypothesis not yet measured"
+      },
+      {
+        id: "deepseek-v4-flash",
+        grounding: "judgment",
+        reason: "Cheapest viable; sibling-provider fallback"
+      },
+      {
+        id: "claude-sonnet-4-6",
+        grounding: "judgment",
+        reason: "Cross-provider safety net \u2014 Sonnet handles sequential agentic loops cleanly"
+      },
+      {
+        id: "gemini-2.5-pro",
+        grounding: "judgment",
+        reason: "Third-provider tail when no DeepSeek key reachable"
+      }
+    ]
+  }
+};
+var STARTER_CHAINS_BY_MODE = (() => {
+  const out = {};
+  for (const [archetype, modes] of Object.entries(STARTER_CHAINS_BY_MODE_GROUNDED)) {
+    if (modes?.sequential) {
+      out[archetype] = {
+        sequential: modes.sequential.map((e) => e.id)
+      };
+    }
+  }
+  return out;
+})();
+function resolveStarterForMode(archetype, toolOrchestration, allChains) {
+  if (toolOrchestration === "sequential") {
+    const overlay = STARTER_CHAINS_BY_MODE[archetype]?.sequential;
+    if (overlay) return [...overlay];
+  }
+  return allChains[archetype];
+}
+function getDefaultFallbackChain(opts) {
+  const { archetype, primary, maxDepth = 3, policy, reachability, toolOrchestration } = opts;
+  if (maxDepth < 1) {
+    throw new Error(
+      `getDefaultFallbackChain: maxDepth must be >= 1, got ${maxDepth}`
+    );
+  }
+  const allChains = loadChainsFromBrain();
+  const starter = resolveStarterForMode(archetype, toolOrchestration, allChains);
+  if (!starter) {
+    throw new Error(
+      `getDefaultFallbackChain: unknown archetype "${archetype}". Known: ${Object.keys(allChains).join(", ")}`
+    );
+  }
+  let chain;
+  if (primary) {
+    chain = [primary, ...starter.filter((id) => id !== primary)];
+  } else {
+    chain = [...starter];
+  }
+  if (policy?.blockedModels && policy.blockedModels.length > 0) {
+    const blocked = new Set(policy.blockedModels);
+    chain = chain.filter((id) => !blocked.has(id));
+  }
+  const seen = /* @__PURE__ */ new Set();
+  const deduped = [];
+  for (const id of chain) {
+    if (!seen.has(id)) {
+      seen.add(id);
+      deduped.push(id);
+    }
+  }
+  let filtered = deduped;
+  if (reachability) {
+    filtered = deduped.filter((id) => isModelReachable(id, reachability));
+  }
+  return filtered.slice(0, maxDepth);
+}
+function getAllStarterChains() {
+  const out = {};
+  for (const [archetype, chain] of Object.entries(STARTER_CHAINS)) {
+    out[archetype] = [...chain];
+  }
+  return out;
+}
+
+// src/compatibility.ts
+var ARCHETYPE_FLOOR_DEFAULT = 6;
+
+// src/glassbox-routes/counterfactuals.ts
+var COUNTERFACTUAL_MIN_SAVINGS_RATIO = 0.1;
+var COUNTERFACTUAL_MAX_RESULTS = 2;
+function computeCounterfactuals(args) {
+  const {
+    servedModel,
+    servedCostUsd,
+    archetype,
+    tokensIn,
+    tokensOut,
+    cacheReadInputTokens = 0,
+    toolOrchestration
+  } = args;
+  if (tokensIn <= 0) return [];
+  if (servedCostUsd <= 0) return [];
+  let chain;
+  try {
+    chain = getDefaultFallbackChain({
+      archetype,
+      posture: "open",
+      maxDepth: 10,
+      toolOrchestration
+    });
+  } catch {
+    return [];
+  }
+  const candidates = [];
+  const minSavings = servedCostUsd * COUNTERFACTUAL_MIN_SAVINGS_RATIO;
+  for (const modelId of chain) {
+    if (modelId === servedModel) continue;
+    const profile = tryGetProfile(modelId);
+    if (!profile) continue;
+    const perf = profile.archetypePerf?.[archetype] ?? 5;
+    if (perf < ARCHETYPE_FLOOR_DEFAULT) continue;
+    const estimated = estimateCostUsd({
+      profile,
+      tokensIn,
+      tokensOut,
+      cacheReadInputTokens
+    });
+    if (estimated === void 0) continue;
+    const savings = servedCostUsd - estimated;
+    if (savings < minSavings) continue;
+    const savingsPercent = Math.round(savings / servedCostUsd * 100);
+    const reason = buildReason({
+      modelId,
+      archetype,
+      perf,
+      profile
+    });
+    candidates.push({
+      modelId,
+      estimatedCostUsd: round6(estimated),
+      savingsUsd: round6(savings),
+      savingsPercent,
+      reason
+    });
+  }
+  candidates.sort((a, b) => a.estimatedCostUsd - b.estimatedCostUsd);
+  return candidates.slice(0, COUNTERFACTUAL_MAX_RESULTS);
+}
+function estimateCostUsd(args) {
+  const { profile, tokensIn, tokensOut, cacheReadInputTokens } = args;
+  const cacheableIn = Math.min(cacheReadInputTokens, tokensIn);
+  const nonCachedIn = Math.max(tokensIn - cacheableIn, 0);
+  const discount = profile.lowering.cache.discount ?? 1;
+  const inUsd = nonCachedIn / 1e6 * profile.costInputPer1m + cacheableIn / 1e6 * profile.costInputPer1m * discount;
+  const outUsd = tokensOut / 1e6 * profile.costOutputPer1m;
+  const total = inUsd + outUsd;
+  if (!Number.isFinite(total)) return void 0;
+  if (total < 0) return void 0;
+  return total;
+}
+function round6(n) {
+  return Math.round(n * 1e6) / 1e6;
+}
+function buildReason(args) {
+  const { modelId, archetype, perf, profile } = args;
+  const hook = profile.strengths?.[0];
+  const suffix = hook ? `, ${hook.replace(/_/g, " ")}` : "";
+  return `${modelId} on ${archetype}: archetypePerf=${perf}${suffix}`;
+}
+
+// src/glassbox-routes/projected-cost.ts
+var INSUFFICIENT_VOLUME_THRESHOLD = 5;
+var WINDOW_DAYS = 7;
+async function computeProjectedDailyCost(args) {
+  const {
+    appId,
+    archetype,
+    servedCostUsd,
+    brainEndpoint,
+    brainJwt,
+    brainAnonKey,
+    fetch: fetchImpl
+  } = args;
+  if (!appId || !archetype) return void 0;
+  if (!Number.isFinite(servedCostUsd) || servedCostUsd <= 0) return void 0;
+  const doFetch = fetchImpl ?? ((...a) => globalThis.fetch(...a));
+  const base = brainEndpoint.replace(/\/+$/, "");
+  const cutoffIso = new Date(
+    Date.now() - WINDOW_DAYS * 24 * 60 * 60 * 1e3
+  ).toISOString();
+  const qs = new URLSearchParams();
+  qs.set("app_id", `eq.${appId}`);
+  qs.set("intent_archetype", `eq.${archetype}`);
+  qs.set("created_at", `gte.${cutoffIso}`);
+  qs.set("select", "handle");
+  qs.set("limit", "0");
+  const url = `${base}/rest/v1/compile_outcomes?${qs.toString()}`;
+  let res;
+  try {
+    res = await doFetch(url, {
+      method: "GET",
+      headers: {
+        Authorization: `Bearer ${brainJwt}`,
+        apikey: brainAnonKey,
+        Accept: "application/json",
+        // Triggers PostgREST exact count in Content-Range header.
+        Prefer: "count=exact"
+      }
+    });
+  } catch {
+    return void 0;
+  }
+  if (!res.ok) return void 0;
+  const contentRange = res.headers.get("content-range");
+  const count = parseContentRangeCount(contentRange);
+  if (count === void 0) return void 0;
+  const avgPerDay = count / WINDOW_DAYS;
+  if (avgPerDay < INSUFFICIENT_VOLUME_THRESHOLD) return void 0;
+  const projected = avgPerDay * servedCostUsd;
+  return Math.round(projected * 1e6) / 1e6;
+}
+function parseContentRangeCount(header) {
+  if (!header) return void 0;
+  const slash = header.lastIndexOf("/");
+  if (slash < 0) return void 0;
+  const tail = header.slice(slash + 1).trim();
+  if (tail === "*" || tail === "") return void 0;
+  const n = Number.parseInt(tail, 10);
+  if (!Number.isFinite(n) || n < 0) return void 0;
+  return n;
+}
+
 // src/glassbox-routes/proxy.ts
 var JSON_HEADERS2 = {
   "Content-Type": "application/json",
@@ -97,15 +1749,151 @@ function rowToSummary(row) {
     estimatedCostUsd: typeof row.cost_usd_actual === "number" ? row.cost_usd_actual : 0
   };
 }
+var INPUT_RATIO_YELLOW = 0.65;
+var INPUT_RATIO_RED = 0.85;
+var CACHE_HEALTH_MIN_TOKENS = 1e3;
+var CACHE_RATIO_GREEN = 0.5;
+var CACHE_RATIO_YELLOW = 0.1;
+var FALLBACK_REASONS = /* @__PURE__ */ new Set([
+  "rate_limit",
+  "provider_auth_failed",
+  "provider_error",
+  "cliff",
+  "cost_cap",
+  "contract_violation"
+]);
+function asString(v) {
+  return typeof v === "string" && v.length > 0 ? v : void 0;
+}
+function asNumber(v) {
+  return typeof v === "number" && Number.isFinite(v) ? v : void 0;
+}
+function asNumberOrZero(v) {
+  return typeof v === "number" && Number.isFinite(v) ? v : 0;
+}
+function asStringArray(v) {
+  if (!Array.isArray(v)) return [];
+  const out = [];
+  for (const e of v) {
+    if (typeof e === "string") out.push(e);
+  }
+  return out;
+}
+function asFallbackReason(v) {
+  if (typeof v !== "string") return void 0;
+  const candidate = v;
+  if (candidate && FALLBACK_REASONS.has(candidate)) return candidate;
+  return "provider_error";
+}
+function rowToAdvisory(raw) {
+  if (!raw || typeof raw !== "object") return void 0;
+  const r = raw;
+  const level = r.level;
+  const code = r.code;
+  const message = r.message;
+  if (level !== "info" && level !== "warn" && level !== "critical" || typeof code !== "string" || typeof message !== "string") {
+    return void 0;
+  }
+  const out = { level, code, message };
+  const suggestion = asString(r.suggestion);
+  if (suggestion) out.suggestion = suggestion;
+  const docsUrl = asString(r.docs_url ?? r.docsUrl);
+  if (docsUrl) out.docsUrl = docsUrl;
+  const adapter = toAdapter(r.suggested_adaptation ?? r.suggestedAdaptation);
+  if (adapter) out.suggestedAdaptation = adapter;
+  return out;
+}
+function toAdapter(raw) {
+  if (!raw || typeof raw !== "object") return void 0;
+  const a = raw;
+  if (a.parameter === "toolOrchestration" && a.value === "sequential" && typeof a.consequence === "string") {
+    return {
+      parameter: "toolOrchestration",
+      value: "sequential",
+      consequence: a.consequence
+    };
+  }
+  return void 0;
+}
+function computeHealth(args) {
+  const {
+    tokensIn,
+    tokensOut,
+    historyCacheableTokens,
+    inputCacheHitRatio,
+    fellOverFrom,
+    target
+  } = args;
+  const total = tokensIn + tokensOut;
+  const ratio = total > 0 ? tokensIn / total : 0;
+  let inputRatioStatus;
+  if (ratio > INPUT_RATIO_RED) inputRatioStatus = "red";
+  else if (ratio > INPUT_RATIO_YELLOW) inputRatioStatus = "yellow";
+  else inputRatioStatus = "green";
+  let cacheStatus;
+  if (historyCacheableTokens <= CACHE_HEALTH_MIN_TOKENS) {
+    cacheStatus = "na";
+  } else if (inputCacheHitRatio >= CACHE_RATIO_GREEN) {
+    cacheStatus = "green";
+  } else if (inputCacheHitRatio >= CACHE_RATIO_YELLOW) {
+    cacheStatus = "yellow";
+  } else {
+    cacheStatus = "red";
+  }
+  const fallbackStatus = fellOverFrom !== void 0 && fellOverFrom !== target ? "red" : "green";
+  return { inputRatioStatus, cacheStatus, fallbackStatus };
+}
 function rowToDetail(row) {
-  return {
-    ...rowToSummary(row),
-    mutationsApplied: Array.isArray(row.mutations_applied) ? row.mutations_applied : [],
-    advisories: [],
-    // compile_outcomes has no advisories column today
-    rawRequest: row.prompt_preview ?? void 0,
-    rawResponse: row.response_preview ?? void 0
+  const summary = rowToSummary(row);
+  const tokensIn = summary.tokensIn;
+  const tokensOut = summary.tokensOut;
+  const cacheReadInputTokens = asNumberOrZero(row.cache_read_input_tokens);
+  const cacheCreationInputTokens = asNumberOrZero(
+    row.cache_creation_input_tokens
+  );
+  const historyCacheableTokens = asNumberOrZero(row.history_cacheable_tokens);
+  const inputCacheHitRatio = tokensIn > 0 ? cacheReadInputTokens / tokensIn : 0;
+  const fellOverFrom = asString(row.fell_over_from);
+  const fallbackReasonRaw = row.fallback_reason;
+  const fallbackReason = fellOverFrom ? asFallbackReason(fallbackReasonRaw) : void 0;
+  const requestedModel = asString(row.requested_model) ?? fellOverFrom;
+  const advisoriesRaw = Array.isArray(row.advisories) ? row.advisories : [];
+  const advisories = [];
+  for (const a of advisoriesRaw) {
+    const rec = rowToAdvisory(a);
+    if (rec) advisories.push(rec);
+  }
+  const health = computeHealth({
+    tokensIn,
+    tokensOut,
+    cacheReadInputTokens,
+    historyCacheableTokens,
+    inputCacheHitRatio,
+    fellOverFrom,
+    target: summary.target
+  });
+  const detail = {
+    ...summary,
+    mutationsApplied: asStringArray(row.mutations_applied),
+    advisories,
+    rawRequest: asString(row.prompt_preview),
+    rawResponse: asString(row.response_preview),
+    requestedModel,
+    finishReason: asString(row.finish_reason),
+    ttftMs: asNumber(row.ttft_ms),
+    totalMs: asNumber(row.total_ms) ?? asNumber(row.latency_ms),
+    toolsCount: asNumber(row.tools_count),
+    historyDepth: asNumber(row.history_depth),
+    systemPromptChars: asNumber(row.system_prompt_chars),
+    cacheReadInputTokens,
+    cacheCreationInputTokens,
+    historyCacheableTokens,
+    inputCacheHitRatio,
+    fellOverFrom,
+    fallbackReason,
+    health
   };
+  return detail;
 }
 function createProxyHandler(config) {
   const {
@@ -177,7 +1965,31 @@ function createProxyHandler(config) {
     if (traceId) {
       const first = scrubbed[0];
       if (!first) return jsonError2(404, "not_found");
-      return jsonResponse(200, rowToDetail(first));
+      const detail = rowToDetail(first);
+      const counterfactuals = computeCounterfactuals({
+        servedModel: detail.target,
+        servedCostUsd: detail.estimatedCostUsd,
+        archetype: detail.archetype,
+        tokensIn: detail.tokensIn,
+        tokensOut: detail.tokensOut,
+        cacheReadInputTokens: detail.cacheReadInputTokens
+      });
+      detail.counterfactuals = counterfactuals;
+      if (detail.estimatedCostUsd > 0) {
+        const projected = await computeProjectedDailyCost({
+          appId: detail.appId,
+          archetype: detail.archetype,
+          servedCostUsd: detail.estimatedCostUsd,
+          brainEndpoint: base,
+          brainJwt,
+          brainAnonKey,
+          fetch: doFetch
+        });
+        if (projected !== void 0) {
+          detail.projectedDailyCostUsd = projected;
+        }
+      }
+      return jsonResponse(200, detail);
     }
     return jsonResponse(200, { traces: scrubbed.map(rowToSummary) });
   };