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

package/dist/index.js

@@ -269,33 +269,69 @@ function passCompressHistory(ir, opts = {}) {
   const summarizeAboveTokens = opts.summarizeAboveTokens;
   const historyTokensTotal = totalHistoryTokens(history);
   const countThresholdHit = history.length > summarizeOlderThan;
-  const tokenThresholdHit = summarizeAboveTokens !== void 0 && historyTokensTotal > summarizeAboveTokens && history.length > keepRecent;
+  const tokenThresholdHit = summarizeAboveTokens !== void 0 && historyTokensTotal > summarizeAboveTokens;
   if (!countThresholdHit && !tokenThresholdHit) {
     return { value: ir, mutations: [], historyTokensTotal };
   }
-  const cutIndex = history.length - keepRecent;
-  const old = history.slice(0, cutIndex);
-  const recent = history.slice(cutIndex);
-  const userTurns = old.filter((m) => m.role === "user");
-  const firstUserLine = userTurns[0]?.content.split("\n")[0]?.slice(0, 200) ?? "";
-  const oldTokens = totalHistoryTokens(old);
-  const trigger = tokenThresholdHit && !countThresholdHit ? "tokens" : "count";
-  const summary = {
-    role: "system",
-    content: `[Earlier conversation: ${old.length} turns omitted (~${oldTokens} tokens). First user message: "${firstUserLine}"]`
-  };
-  return {
-    value: { ...ir, history: [summary, ...recent] },
-    mutations: [
-      {
-        id: `compress-history-${old.length}`,
-        source: "static_pass",
-        passName: "compress_history",
-        description: trigger === "tokens" ? `Compressed ${old.length} old turns (~${oldTokens} tokens) into 1 summary \u2014 token threshold ${summarizeAboveTokens} exceeded (kept ${keepRecent} recent)` : `Compressed ${old.length} old turns into 1 summary (kept ${keepRecent} recent)`
+  if (history.length > keepRecent) {
+    const cutIndex = history.length - keepRecent;
+    const old = history.slice(0, cutIndex);
+    const recent = history.slice(cutIndex);
+    const userTurns = old.filter((m) => m.role === "user");
+    const firstUserLine = userTurns[0]?.content.split("\n")[0]?.slice(0, 200) ?? "";
+    const oldTokens = totalHistoryTokens(old);
+    const trigger = tokenThresholdHit && !countThresholdHit ? "tokens" : "count";
+    const summary = {
+      role: "system",
+      content: `[Earlier conversation: ${old.length} turns omitted (~${oldTokens} tokens). First user message: "${firstUserLine}"]`
+    };
+    return {
+      value: { ...ir, history: [summary, ...recent] },
+      mutations: [
+        {
+          id: `compress-history-${old.length}`,
+          source: "static_pass",
+          passName: "compress_history",
+          description: trigger === "tokens" ? `Compressed ${old.length} old turns (~${oldTokens} tokens) into 1 summary \u2014 token threshold ${summarizeAboveTokens} exceeded (kept ${keepRecent} recent)` : `Compressed ${old.length} old turns into 1 summary (kept ${keepRecent} recent)`
+        }
+      ],
+      historyTokensTotal
+    };
+  }
+  if (tokenThresholdHit) {
+    let fattestIdx = -1;
+    let fattestTokens = 0;
+    for (let i = 0; i < history.length; i++) {
+      const m = history[i];
+      if (!m || typeof m.content !== "string") continue;
+      const t = countTokens(m.content);
+      if (t > fattestTokens) {
+        fattestTokens = t;
+        fattestIdx = i;
       }
-    ],
-    historyTokensTotal
-  };
+    }
+    const FAT_DOMINANCE_FLOOR = 0.3;
+    const fattest = fattestIdx >= 0 ? history[fattestIdx] : void 0;
+    if (fattest && historyTokensTotal > 0 && fattestTokens / historyTokensTotal >= FAT_DOMINANCE_FLOOR) {
+      const firstLine = fattest.content.split("\n")[0]?.slice(0, 200) ?? "";
+      const newContent = `[Earlier ${fattest.role} message content omitted: ~${fattestTokens} tokens. Preview: "${firstLine}"]`;
+      const newHistory = history.slice();
+      newHistory[fattestIdx] = { ...fattest, content: newContent };
+      return {
+        value: { ...ir, history: newHistory },
+        mutations: [
+          {
+            id: `compress-fat-message-${fattestIdx}`,
+            source: "static_pass",
+            passName: "compress_history",
+            description: `Replaced fat ${fattest.role} message #${fattestIdx} content (~${fattestTokens} of ${historyTokensTotal} tokens, ${Math.round(fattestTokens / historyTokensTotal * 100)}% of history) with summary stub \u2014 token threshold ${summarizeAboveTokens} exceeded (history.length ${history.length} <= keepRecent ${keepRecent}, slice not possible)`
+          }
+        ],
+        historyTokensTotal
+      };
+    }
+  }
+  return { value: ir, mutations: [], historyTokensTotal };
 }
 function passApplyCliffs(ir, profile, estimatedInputTokens) {
   const mutations = [];
@@ -1438,140 +1474,554 @@ var PROFILES_RAW = [
       hunt: 4
       // sequential tools — same as V4-Flash
     }
-  }
-];
-var ALIASES = {
-  // DeepSeek's own model routing — both names served by V4-Flash.
-  "deepseek-chat": "deepseek-v4-flash",
-  "deepseek-reasoner": "deepseek-v4-flash",
-  // Legacy kgauto typo — actual API alias is dash-form (alpha.1 had dot).
-  "claude-haiku-4.5": "claude-haiku-4-5"
-};
-var brainHook = {};
-function _setProfileBrainHook(hook) {
-  brainHook = hook;
-}
-function canonicalId(id) {
-  return brainHook.resolveAlias?.(id) ?? ALIASES[id] ?? id;
-}
-var PROFILE_INDEX = new Map(
-  PROFILES_RAW.map((p) => [p.id, p])
-);
-function getProfile(id) {
-  const canonical = canonicalId(id);
-  const fromBrain = brainHook.getProfile?.(canonical);
-  if (fromBrain) return fromBrain;
-  const p = PROFILE_INDEX.get(canonical);
-  if (!p) {
-    const known = [...PROFILE_INDEX.keys(), ...Object.keys(ALIASES)].join(", ");
-    throw new Error(`Unknown model id: "${id}". Known: ${known}`);
-  }
-  return p;
-}
-function tryGetProfile(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);
-}
-
-// src/advisor.ts
-function runAdvisor(ir, result, profile, policy) {
-  const out = [];
-  out.push(...detectCachingOff(ir, profile));
-  out.push(...detectSingleChunkSystem(ir, profile));
-  out.push(...detectToolBloat(ir, result));
-  out.push(...detectHistoryUncached(ir, profile));
-  out.push(...detectSingleModelArray(ir, policy));
-  return out;
-}
-function detectCachingOff(ir, profile) {
-  if (profile.provider !== "anthropic") return [];
-  const totalChars = ir.sections.reduce((s, sec) => s + sec.text.length, 0);
-  if (totalChars < 2e3) return [];
-  const anyCacheable = ir.sections.some((s) => s.cacheable === true);
-  if (anyCacheable) return [];
-  return [
-    {
-      level: "warn",
-      code: "caching-off-on-claude",
-      message: `System prompt is ${totalChars} chars on Anthropic but no PromptSection has cacheable=true. Anthropic prompt caching cuts cached-prefix input cost by ~90% on subsequent calls; without it, every turn re-pays full price for the static system context.`,
-      suggestion: "Mark stable system sections (role, persona, tool policy) with `cacheable: true`. The lowering pass concatenates cacheable sections into a single cache-controlled block before the dynamic ones.",
-      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
-    }
-  ];
-}
-function detectSingleChunkSystem(ir, profile) {
-  if (profile.provider !== "anthropic") return [];
-  if (ir.sections.length !== 1) return [];
-  const only = ir.sections[0];
-  if (!only || only.text.length <= 1e3) return [];
-  return [
-    {
-      level: "info",
-      code: "single-chunk-system",
-      message: `System prompt is a single ${only.text.length}-char chunk. Splitting into NamedChunks (static role/persona vs dynamic context) gives the lowering pass a finer cache-marker boundary \u2014 only the static portion needs to be byte-stable for the cache to hit.`,
-      suggestion: "Refactor the system builder to return an array of `PromptSection` shaped { id, text, cacheable?: boolean }. Static chunks (role, persona, tool policy) get `cacheable: true`; dynamic ones (current context, today's date) don't.",
-      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+  },
+  // ── 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/$5.00).
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["reasoning", "agentic_coding", "long_context", "structured_output", "reliable_tool_use", "reasoning_effort_knob"],
+    weaknesses: ["cost", "pricing_cliff_at_272k"],
+    notes: "OpenAI frontier (2026-05). 1M context (1.05M total), 128K max output, 2025-12 cutoff. Reasoning effort knob (none/low/medium/high/xhigh). Pricing cliff at 272K input.",
+    // Frontier-tier perf hypothesis. Anchored to Opus 4.7 row (similar
+    // price/positioning). Brain evidence will refine; no telemetry yet.
+    archetypePerf: {
+      critique: 9,
+      plan: 9,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 8,
+      // parallel tool support good but cliff at 272K hurts deep multi-step
+      summarize: 7,
+      // overkill for tolerant archetype
+      classify: 7
+      // overkill; cheaper models cover this
     }
-  ];
-}
-function detectToolBloat(ir, result) {
-  const SHORT_OUTPUT = /* @__PURE__ */ new Set([
-    "classify",
-    "extract",
-    "summarize",
-    "transform",
-    "critique"
-  ]);
-  if (!ir.tools || ir.tools.length === 0) return [];
-  const toolsKept = result.diagnostics.toolsKept;
-  if (toolsKept <= 10) return [];
-  if (!SHORT_OUTPUT.has(ir.intent.archetype)) return [];
-  return [
-    {
-      level: "warn",
-      code: "tool-bloat",
-      message: `${toolsKept} tools kept after the relevance pass for archetype="${ir.intent.archetype}" (consumer declared ${ir.tools.length}). This archetype is short-output and rarely needs more than 3 tools; each tool definition eats ~350 tokens of context budget.`,
-      suggestion: "Tighten `relevanceByIntent: { [archetype]: 0..1 }` per ToolDefinition. Tools below `toolRelevanceThreshold` (default 0.2) get dropped. Without `relevanceByIntent`, every tool defaults to neutral (0.5) and stays.",
-      docsUrl: "https://github.com/stue/kgauto/blob/main/v2/README.md#tools"
+  },
+  {
+    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/$15.00).",
+    // Anchored to Sonnet 4.6 row (similar price/positioning). Slight
+    // anthropic-side edge on agentic coding per master plan vibe.
+    archetypePerf: {
+      critique: 8,
+      plan: 8,
+      generate: 8,
+      ask: 8,
+      extract: 8,
+      transform: 8,
+      hunt: 7,
+      summarize: 7,
+      classify: 7
     }
-  ];
-}
-function detectHistoryUncached(ir, profile) {
-  if (profile.provider !== "anthropic") return [];
-  if (!ir.history || ir.history.length < 2) return [];
-  if (ir.historyCachePolicy && ir.historyCachePolicy.strategy !== "none") {
-    return [];
-  }
-  return [
-    {
-      level: "warn",
-      code: "history-uncached-on-claude",
-      message: `${ir.history.length} history messages on Anthropic with no historyCachePolicy. Every turn re-pays for the full conversation context; with caching, subsequent turns hit the cache at ~10% the input cost.`,
-      suggestion: "Set `historyCachePolicy: { strategy: 'all-but-latest' }` on this IR. The lowering pass marks the message immediately preceding currentTurn with cache_control; subsequent turns whose history prefix matches byte-for-byte hit the cache.",
-      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+  },
+  {
+    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 $0.50 input) but
+    // OpenAI claims strong coding/subagent perf.
+    archetypePerf: {
+      ask: 7,
+      generate: 7,
+      extract: 7,
+      transform: 7,
+      classify: 7,
+      summarize: 7,
+      hunt: 7,
+      plan: 6,
+      critique: 5
+      // reasoning depth gap — frontier models handle this
     }
-  ];
-}
-function detectSingleModelArray(ir, policy) {
-  if (ir.models.length !== 1) return [];
-  if (policy?.posture === "locked") return [];
-  const only = ir.models[0];
-  return [
-    {
-      level: "warn",
-      code: "single-model-array",
-      message: `\`ir.models\` has length 1 (only "${only}") and posture is not 'locked'. A single-model chain has no safety net \u2014 the first 429 / 5xx / cliff hits the user as a failure. Master plan \xA71.2 closes the reliability gap with a 2-step minimum.`,
-      suggestion: "Use `getDefaultFallbackChain({ archetype: ir.intent.archetype, primary: '" + only + "', posture: 'preferred' })` for a user-anchored chain, or `getDefaultFallbackChain({ archetype, posture: 'open' })` for library-picked. If single-model is intentional (compliance/brand promise), set `policy.posture = 'locked'` to silence this rule.",
-      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#single-model-array"
+  },
+  {
+    id: "gpt-5.4-nano",
+    verifiedAgainstDocs: "2026-05-17",
+    provider: "openai",
+    status: "current",
+    maxContextTokens: 4e5,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 0.2,
+    costOutputPer1m: 1.25,
+    lowering: {
+      system: { mode: "inline" },
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["cost", "speed", "volume", "structured_output"],
+    weaknesses: ["reasoning_depth", "no_computer_use"],
+    notes: "OpenAI nano-tier (2026-05). 400K context, 128K max output, 2025-08 cutoff. 'Cheapest GPT-5.4-class for simple high-volume tasks.' No fine-tuning, no computer-use tools. Cache discount 10x.",
+    // Nano-tier. Anchored to Flash-Lite row ($0.10/$0.40 vs nano's
+    // $0.20/$1.25). Slightly more expensive than Flash-Lite but with
+    // OpenAI brand reliability. Good fit for classify/summarize floor.
+    archetypePerf: {
+      classify: 7,
+      summarize: 6,
+      ask: 6,
+      transform: 6,
+      extract: 6,
+      generate: 5,
+      hunt: 5,
+      plan: 4,
+      critique: 3
+      // not for reasoning archetypes
+    }
+  },
+  // ── Auto-onboarded (UNVERIFIED) ──
+  // Cloned by scripts/auto-onboard-models.mjs from a same-family template.
+  // Each entry's pricing/context/cliffs/lowering reflects the template, NOT
+  // provider docs. Verify before promoting status to 'current' (L-049/L-081).
+  {
+    id: "gemini-3-flash-preview",
+    verifiedAgainstDocs: "UNVERIFIED-AUTO-ONBOARD",
+    provider: "google",
+    status: "preview",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 65535,
+    maxTools: 128,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "separate",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 8e3,
+        action: "downgrade_quality_warning",
+        reason: "Quality degrades significantly above ~8K context tokens"
+      },
+      {
+        metric: "tool_count",
+        threshold: 20,
+        action: "drop_to_top_relevant",
+        reason: "Tool reliability drops above ~20 tools (despite 128 hard limit)"
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking tokens consume maxOutputTokens \u2014 empty response if drained"
+      },
+      {
+        // s11 trust artifact (2026-05-02): brain showed 5/5 empty rate on
+        // tt-intelligence/summarize/gemini-2.5-flash with tools offered.
+        // v1's disable_thinking_for_short_output already fired and didn't
+        // help — disabling thinking is necessary but not sufficient. Tools
+        // present + summarize intent confuses Flash into a no-output state
+        // (likely tool-decision purgatory). Strip tools entirely for this
+        // archetype on this model.
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Gemini Flash returns empty when summarize intent has tools offered (5/5 empty rate observed in v1 prod 2026-04-19, replayed into v2 brain 2026-04-29)"
+      }
+    ],
+    costInputPer1m: 0.3,
+    costOutputPer1m: 2.5,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      thinking: { field: "generationConfig.thinkingConfig.thinkingBudget", default: "auto" }
+    },
+    recovery: [
+      {
+        signal: "empty_response_after_tool",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Known: empty after tool result \u2014 retry with thinking off"
+      },
+      {
+        signal: "empty_response",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Empty response \u2014 try with thinking off"
+      },
+      {
+        signal: "malformed_function_call",
+        action: "escalate",
+        reason: "MALFORMED_FUNCTION_CALL maps to stop \u2014 escalate to next target"
+      }
+    ],
+    strengths: ["speed", "volume", "classification", "1m_context", "cost"],
+    weaknesses: ["complex_schemas", "large_tool_sets", "high_context_quality"],
+    notes: "AUTO-ONBOARDED 2026-05-16 from `gemini-2.5-flash`. Pricing, context, cliffs are template-cloned and UNVERIFIED \u2014 confirm against provider docs before promoting status to 'current'.",
+    // Master plan §6.2 anchor. Tier 0 for hunt (parallel tool throughput
+    // 15-75 calls/step beats Sonnet — L-040), summarize, classify.
+    archetypePerf: {
+      hunt: 9,
+      // L-040: parallel tool throughput 15-75/step
+      classify: 7,
+      // brain-validated, 218 rows
+      summarize: 7,
+      // brain-validated; cliff strips tools when present
+      transform: 7,
+      ask: 7,
+      generate: 6,
+      plan: 5,
+      extract: 6,
+      // alpha.8 MAX_TOKENS history on structured output
+      critique: 4
+      // reasoning shallower than Sonnet/Opus
+    }
+  },
+  {
+    // ── Gemini 2.5 Flash-Lite ──
+    // Onboarded 2026-05-13 (s22) after the model-release watcher surfaced
+    // it as a UNREGISTERED + NEW candidate. Released by Google July 2025,
+    // stable. Positioned BELOW Flash on the cost/perf frontier:
+    //   input  $0.10/M (Flash $0.30/M)   — 3× cheaper
+    //   output $0.40/M (Flash $2.50/M)   — 6× cheaper
+    //   cache  $0.01/M                    — 1/10 of input (vs Flash 0.25 discount)
+    // Cliffs are HYPOTHESIZED from Flash's known failure modes — Flash-Lite
+    // is a smaller sibling, so we inherit Flash's cliffs at equal-or-tighter
+    // thresholds. The brain will validate/relax these as evidence accumulates
+    // per (archetype, model) tuple. Currently ZERO brain rows for this model.
+    id: "gemini-3.1-flash-lite",
+    verifiedAgainstDocs: "UNVERIFIED-AUTO-ONBOARD",
+    provider: "google",
+    status: "preview",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 65536,
+    maxTools: 128,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "separate",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 8e3,
+        action: "downgrade_quality_warning",
+        reason: "Inherited from Flash: quality degrades above ~8K. Smaller model \u2014 likely degrades faster. Re-tune from brain after n\u226520."
+      },
+      {
+        metric: "tool_count",
+        threshold: 10,
+        action: "drop_to_top_relevant",
+        reason: "Conservative: Flash drops at 20, Flash-Lite is smaller \u2014 assume tighter ceiling until brain proves otherwise."
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking enabled per Google API (thinking: true). Same drain risk as Flash \u2014 thinking tokens consume maxOutputTokens."
+      },
+      {
+        // Strong prior: Flash hit 5/5 empty rate on summarize+tools (s11
+        // trust artifact, kgauto commit 3872832). Flash-Lite shares the
+        // same architectural family — almost certainly inherits this cliff.
+        // Ship the guard preemptively; brain telemetry confirms or relaxes.
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Inherited from Flash s11 cliff: summarize+tools \u2192 empty response. Preemptive guard until brain evidence on Flash-Lite specifically."
+      }
+    ],
+    costInputPer1m: 0.1,
+    costOutputPer1m: 0.4,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      // Cache discount 10× (vs Flash 4×) — Google's spec is $0.01/M cache vs
+      // $0.10/M input. Material for repeat-prompt workloads (classify shape).
+      cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
+      thinking: { field: "generationConfig.thinkingConfig.thinkingBudget", default: "auto" }
+    },
+    recovery: [
+      {
+        signal: "empty_response_after_tool",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Known on Flash family: empty after tool result \u2014 retry with thinking off."
+      },
+      {
+        signal: "empty_response",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Empty response \u2014 try with thinking off."
+      },
+      {
+        signal: "malformed_function_call",
+        action: "escalate",
+        reason: "MALFORMED_FUNCTION_CALL maps to stop \u2014 escalate to next target."
+      }
+    ],
+    strengths: ["lowest_cost", "speed", "volume", "classification", "summarize", "1m_context", "cache_friendly"],
+    weaknesses: ["complex_reasoning", "large_tool_sets", "complex_schemas", "structured_output_unproven", "long_context_quality"],
+    notes: "AUTO-ONBOARDED 2026-05-16 from `gemini-2.5-flash-lite`. Pricing, context, cliffs are template-cloned and UNVERIFIED \u2014 confirm against provider docs before promoting status to 'current'.",
+    // Tier 3 emergency floor for summarize/classify chains. ZERO brain
+    // rows — all values are starter hypotheses anchored to "smaller
+    // sibling of Flash, at-or-below Flash perf on every archetype." The
+    // first 50 brain rows per archetype will validate or relax these.
+    archetypePerf: {
+      classify: 6,
+      // starter hypothesis — verify (Flash is 7, lite likely ≤)
+      summarize: 6,
+      // starter hypothesis — verify; cliff strips tools
+      transform: 6,
+      // starter hypothesis — verify
+      ask: 5,
+      hunt: 5,
+      generate: 4,
+      extract: 4,
+      plan: 3,
+      critique: 3
+    }
+  }
+];
+var ALIASES = {
+  // 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 _setProfileBrainHook(hook) {
+  brainHook = hook;
+}
+function canonicalId(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(", ");
+    throw new Error(`Unknown model id: "${id}". Known: ${known}`);
+  }
+  return p;
+}
+function tryGetProfile(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);
+}
+
+// src/advisor.ts
+function runAdvisor(ir, result, profile, policy) {
+  const out = [];
+  out.push(...detectCachingOff(ir, profile));
+  out.push(...detectSingleChunkSystem(ir, profile));
+  out.push(...detectToolBloat(ir, result));
+  out.push(...detectHistoryUncached(ir, profile));
+  out.push(...detectSingleModelArray(ir, policy));
+  return out;
+}
+function detectCachingOff(ir, profile) {
+  if (profile.provider !== "anthropic") return [];
+  const totalChars = ir.sections.reduce((s, sec) => s + sec.text.length, 0);
+  if (totalChars < 2e3) return [];
+  const anyCacheable = ir.sections.some((s) => s.cacheable === true);
+  if (anyCacheable) return [];
+  return [
+    {
+      level: "warn",
+      code: "caching-off-on-claude",
+      message: `System prompt is ${totalChars} chars on Anthropic but no PromptSection has cacheable=true. Anthropic prompt caching cuts cached-prefix input cost by ~90% on subsequent calls; without it, every turn re-pays full price for the static system context.`,
+      suggestion: "Mark stable system sections (role, persona, tool policy) with `cacheable: true`. The lowering pass concatenates cacheable sections into a single cache-controlled block before the dynamic ones.",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+    }
+  ];
+}
+function detectSingleChunkSystem(ir, profile) {
+  if (profile.provider !== "anthropic") return [];
+  if (ir.sections.length !== 1) return [];
+  const only = ir.sections[0];
+  if (!only || only.text.length <= 1e3) return [];
+  return [
+    {
+      level: "info",
+      code: "single-chunk-system",
+      message: `System prompt is a single ${only.text.length}-char chunk. Splitting into NamedChunks (static role/persona vs dynamic context) gives the lowering pass a finer cache-marker boundary \u2014 only the static portion needs to be byte-stable for the cache to hit.`,
+      suggestion: "Refactor the system builder to return an array of `PromptSection` shaped { id, text, cacheable?: boolean }. Static chunks (role, persona, tool policy) get `cacheable: true`; dynamic ones (current context, today's date) don't.",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+    }
+  ];
+}
+function detectToolBloat(ir, result) {
+  const SHORT_OUTPUT = /* @__PURE__ */ new Set([
+    "classify",
+    "extract",
+    "summarize",
+    "transform",
+    "critique"
+  ]);
+  if (!ir.tools || ir.tools.length === 0) return [];
+  const toolsKept = result.diagnostics.toolsKept;
+  if (toolsKept <= 10) return [];
+  if (!SHORT_OUTPUT.has(ir.intent.archetype)) return [];
+  return [
+    {
+      level: "warn",
+      code: "tool-bloat",
+      message: `${toolsKept} tools kept after the relevance pass for archetype="${ir.intent.archetype}" (consumer declared ${ir.tools.length}). This archetype is short-output and rarely needs more than 3 tools; each tool definition eats ~350 tokens of context budget.`,
+      suggestion: "Tighten `relevanceByIntent: { [archetype]: 0..1 }` per ToolDefinition. Tools below `toolRelevanceThreshold` (default 0.2) get dropped. Without `relevanceByIntent`, every tool defaults to neutral (0.5) and stays.",
+      docsUrl: "https://github.com/stue/kgauto/blob/main/v2/README.md#tools"
+    }
+  ];
+}
+function detectHistoryUncached(ir, profile) {
+  if (profile.provider !== "anthropic") return [];
+  if (!ir.history || ir.history.length < 2) return [];
+  if (ir.historyCachePolicy && ir.historyCachePolicy.strategy !== "none") {
+    return [];
+  }
+  return [
+    {
+      level: "warn",
+      code: "history-uncached-on-claude",
+      message: `${ir.history.length} history messages on Anthropic with no historyCachePolicy. Every turn re-pays for the full conversation context; with caching, subsequent turns hit the cache at ~10% the input cost.`,
+      suggestion: "Set `historyCachePolicy: { strategy: 'all-but-latest' }` on this IR. The lowering pass marks the message immediately preceding currentTurn with cache_control; subsequent turns whose history prefix matches byte-for-byte hit the cache.",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+    }
+  ];
+}
+function detectSingleModelArray(ir, policy) {
+  if (ir.models.length !== 1) return [];
+  if (policy?.posture === "locked") return [];
+  const only = ir.models[0];
+  return [
+    {
+      level: "warn",
+      code: "single-model-array",
+      message: `\`ir.models\` has length 1 (only "${only}") and posture is not 'locked'. A single-model chain has no safety net \u2014 the first 429 / 5xx / cliff hits the user as a failure. Master plan \xA71.2 closes the reliability gap with a 2-step minimum.`,
+      suggestion: "Use `getDefaultFallbackChain({ archetype: ir.intent.archetype, primary: '" + only + "', posture: 'preferred' })` for a user-anchored chain, or `getDefaultFallbackChain({ archetype, posture: 'open' })` for library-picked. If single-model is intentional (compliance/brand promise), set `policy.posture = 'locked'` to silence this rule.",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#single-model-array"
     }
   ];
 }
@@ -2245,47 +2695,234 @@ function classifyHttpError(status, body) {
   if (status === 429) {
     return { ok: false, status, errorType: "retryable", errorCode: "rate_limit", message, raw: body };
   }
-  if (status === 408) {
-    return { ok: false, status, errorType: "retryable", errorCode: "timeout", message, raw: body };
+  if (status === 408) {
+    return { ok: false, status, errorType: "retryable", errorCode: "timeout", message, raw: body };
+  }
+  if (status >= 500) {
+    return { ok: false, status, errorType: "retryable", errorCode: "server_error", message, raw: body };
+  }
+  if (status === 404) {
+    return { ok: false, status, errorType: "retryable", errorCode: "model_not_found", message, raw: body };
+  }
+  if (status === 401 || status === 403) {
+    return { ok: false, status, errorType: "terminal", errorCode: "auth", message, raw: body };
+  }
+  if (status === 400) {
+    return { ok: false, status, errorType: "terminal", errorCode: "invalid_request", message, raw: body };
+  }
+  return { ok: false, status, errorType: "terminal", errorCode: "unknown", message, raw: body };
+}
+function extractErrorMessage(body) {
+  if (!body || typeof body !== "object") return void 0;
+  const b = body;
+  if (b.error && typeof b.error === "object") {
+    const e = b.error;
+    if (typeof e.message === "string") return e.message;
+  }
+  if (typeof b.message === "string") return b.message;
+  return void 0;
+}
+function terminalError(status, code, message) {
+  return { ok: false, status, errorType: "terminal", errorCode: code, message, raw: null };
+}
+function retryableError(status, code, message, raw) {
+  return { ok: false, status, errorType: "retryable", errorCode: code, message, raw };
+}
+function tryParseJson(s) {
+  if (typeof s !== "string" || s.length === 0) return void 0;
+  try {
+    const parsed = JSON.parse(s);
+    return typeof parsed === "object" && parsed !== null ? parsed : void 0;
+  } catch {
+    return void 0;
+  }
+}
+
+// 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 = {
+  // Reasoning floor — never degrade. Walk UP on 429 to Opus → cross-provider.
+  // alpha.16: gpt-5.5 appended as third-provider critique floor (frontier-tier,
+  // archetypePerf=9). Cross-provider-tail invariant has somewhere to land when
+  // both Anthropic + Google are unreachable (consumer adds only OpenAI key).
+  critique: [
+    "claude-opus-4-7",
+    "claude-sonnet-4-6",
+    "gemini-2.5-pro",
+    "gpt-5.5"
+  ],
+  // Reasoning matters — Sonnet primary; walk UP to Opus on 429 (rare exception
+  // to "always cheaper"); cross-provider via Pro; DeepSeek Pro as tier 3 floor.
+  plan: [
+    "claude-sonnet-4-6",
+    "claude-opus-4-7",
+    "gemini-2.5-pro",
+    "deepseek-v4-pro"
+  ],
+  // Quality + cost match. Walk Sonnet → Haiku same-provider, Pro cross,
+  // gpt-5.4-mini as third-provider tail (alpha.16 — closes the mono-Anthropic
+  // gap when consumer has only ANTHROPIC + OPENAI keys; archetypePerf=7).
+  generate: [
+    "claude-sonnet-4-6",
+    "claude-haiku-4-5",
+    "gemini-2.5-pro",
+    "gpt-5.4-mini"
+  ],
+  ask: [
+    "claude-sonnet-4-6",
+    "claude-haiku-4-5",
+    "gemini-2.5-pro",
+    "gpt-5.4-mini"
+  ],
+  // Structured-output archetype — Flash skipped (alpha.8 MAX_TOKENS cliff),
+  // DeepSeek skipped (no brain evidence). Floor at Haiku. alpha.16: gpt-5.4
+  // appended as third-provider extract floor (archetypePerf=8, native
+  // structured-output support).
+  extract: [
+    "claude-sonnet-4-6",
+    "claude-haiku-4-5",
+    "gemini-2.5-pro",
+    "gpt-5.4"
+  ],
+  // Forgiving archetype — Sonnet primary but Flash safely floors it.
+  transform: [
+    "claude-sonnet-4-6",
+    "claude-haiku-4-5",
+    "gemini-2.5-pro",
+    "gemini-2.5-flash"
+  ],
+  // Parallel-tool throughput champion (Flash, L-040). Tier 1 cross-provider
+  // Pro; tier 2 Sonnet (quality safety net for blocked-Flash case); tier 3
+  // Haiku (reduced tool budget — cliff at 16 fires).
+  hunt: [
+    "gemini-2.5-flash",
+    "gemini-2.5-pro",
+    "claude-sonnet-4-6",
+    "claude-haiku-4-5"
+  ],
+  // Cost-sensitive + tolerant. DeepSeek brain-evidence tier 1; Haiku tier 2
+  // for quality safety; Flash-Lite emergency floor (onboarded s22).
+  summarize: [
+    "gemini-2.5-flash",
+    "deepseek-v4-flash",
+    "claude-haiku-4-5",
+    "gemini-2.5-flash-lite"
+  ],
+  // Brain-validated DeepSeek tier 1 (169 rows, 0% empty); Haiku tier 2;
+  // Flash-Lite floor for repeat-prompt workloads (cache-discount 10×).
+  classify: [
+    "gemini-2.5-flash",
+    "deepseek-v4-flash",
+    "claude-haiku-4-5",
+    "gemini-2.5-flash-lite"
+  ]
+};
+function getDefaultFallbackChain(opts) {
+  const { archetype, primary, maxDepth = 3, policy, reachability } = opts;
+  if (maxDepth < 1) {
+    throw new Error(
+      `getDefaultFallbackChain: maxDepth must be >= 1, got ${maxDepth}`
+    );
+  }
+  const allChains = loadChainsFromBrain();
+  const starter = allChains[archetype];
+  if (!starter) {
+    throw new Error(
+      `getDefaultFallbackChain: unknown archetype "${archetype}". Known: ${Object.keys(allChains).join(", ")}`
+    );
   }
-  if (status >= 500) {
-    return { ok: false, status, errorType: "retryable", errorCode: "server_error", message, raw: body };
+  let chain;
+  if (primary) {
+    chain = [primary, ...starter.filter((id) => id !== primary)];
+  } else {
+    chain = [...starter];
   }
-  if (status === 404) {
-    return { ok: false, status, errorType: "retryable", errorCode: "model_not_found", message, raw: body };
+  if (policy?.blockedModels && policy.blockedModels.length > 0) {
+    const blocked = new Set(policy.blockedModels);
+    chain = chain.filter((id) => !blocked.has(id));
   }
-  if (status === 401 || status === 403) {
-    return { ok: false, status, errorType: "terminal", errorCode: "auth", message, raw: body };
+  const seen = /* @__PURE__ */ new Set();
+  const deduped = [];
+  for (const id of chain) {
+    if (!seen.has(id)) {
+      seen.add(id);
+      deduped.push(id);
+    }
   }
-  if (status === 400) {
-    return { ok: false, status, errorType: "terminal", errorCode: "invalid_request", message, raw: body };
+  let filtered = deduped;
+  if (reachability) {
+    filtered = deduped.filter((id) => isModelReachable(id, reachability));
   }
-  return { ok: false, status, errorType: "terminal", errorCode: "unknown", message, raw: body };
+  return filtered.slice(0, maxDepth);
 }
-function extractErrorMessage(body) {
-  if (!body || typeof body !== "object") return void 0;
-  const b = body;
-  if (b.error && typeof b.error === "object") {
-    const e = b.error;
-    if (typeof e.message === "string") return e.message;
+function getStarterChain(archetype) {
+  const chain = STARTER_CHAINS[archetype];
+  if (!chain) {
+    throw new Error(
+      `getStarterChain: unknown archetype "${archetype}"`
+    );
   }
-  if (typeof b.message === "string") return b.message;
-  return void 0;
-}
-function terminalError(status, code, message) {
-  return { ok: false, status, errorType: "terminal", errorCode: code, message, raw: null };
+  return [...chain];
 }
-function retryableError(status, code, message, raw) {
-  return { ok: false, status, errorType: "retryable", errorCode: code, message, raw };
+function getAllStarterChains() {
+  const out = {};
+  for (const [archetype, chain] of Object.entries(STARTER_CHAINS)) {
+    out[archetype] = [...chain];
+  }
+  return out;
 }
-function tryParseJson(s) {
-  if (typeof s !== "string" || s.length === 0) return void 0;
-  try {
-    const parsed = JSON.parse(s);
-    return typeof parsed === "object" && parsed !== null ? parsed : void 0;
-  } catch {
-    return void 0;
+function ensureCrossProviderTail(opts) {
+  const { chain, archetype, apiKeys, envSource } = opts;
+  if (chain.length < 1) return { chain };
+  const providers = /* @__PURE__ */ new Set();
+  for (const t of chain) {
+    const p = tryGetProfile(t);
+    if (p) providers.add(p.provider);
+  }
+  if (providers.size >= 2) return { chain };
+  const existingProvider = providers.values().next().value;
+  if (!existingProvider) return { chain };
+  const allChains = loadChainsFromBrain();
+  const fullChain = allChains[archetype];
+  if (!fullChain) return { chain };
+  for (const candidate of fullChain) {
+    if (chain.includes(candidate)) continue;
+    const cp = tryGetProfile(candidate);
+    if (!cp || cp.provider === existingProvider) continue;
+    if (!isModelReachable(candidate, { apiKeys, envSource })) continue;
+    return { chain: [...chain, candidate], appended: candidate };
   }
+  return { chain };
 }
 
 // src/call.ts
@@ -2333,6 +2970,58 @@ async function call(ir, opts = {}) {
         "no_reachable_models"
       );
     }
+    const archetypeName = ir.intent?.archetype;
+    if (archetypeName) {
+      const ensured = ensureCrossProviderTail({
+        chain: targetsToTry,
+        archetype: archetypeName,
+        apiKeys: opts.apiKeys
+      });
+      if (ensured.appended) {
+        targetsToTry = ensured.chain;
+      }
+    }
+  }
+  let policyBlockedFiltered;
+  if (opts.policy?.blockedModels && opts.policy.blockedModels.length > 0) {
+    const blocked = new Set(opts.policy.blockedModels);
+    const filtered = [];
+    const dropped = [];
+    for (const t of targetsToTry) {
+      if (blocked.has(t)) {
+        dropped.push(t);
+      } else {
+        filtered.push(t);
+      }
+    }
+    if (dropped.length > 0) {
+      policyBlockedFiltered = dropped;
+      targetsToTry = filtered;
+    }
+    if (targetsToTry.length === 0) {
+      const latencyMs2 = Date.now() - start;
+      await record({
+        handle: initial.handle,
+        tokensIn: 0,
+        tokensOut: 0,
+        latencyMs: latencyMs2,
+        success: false,
+        errorType: "all_blocked_by_policy",
+        promptPreview: extractPromptPreview(ir)
+      });
+      const blockedAttempts = dropped.map((m) => ({
+        model: m,
+        status: "terminal",
+        errorCode: "blocked_by_policy",
+        message: `Skipped \u2014 model ${m} is in CompilePolicy.blockedModels`
+      }));
+      throw new CallError(
+        `call(): all chain targets blocked by CompilePolicy.blockedModels: [${dropped.join(", ")}]`,
+        blockedAttempts,
+        void 0,
+        "all_blocked_by_policy"
+      );
+    }
   }
   let activeCompile = initial;
   let lastErr;
@@ -2406,7 +3095,8 @@ async function call(ir, opts = {}) {
         servedBy: targetModel,
         fellOverFrom: fellOver ? initial.target : void 0,
         fallbackReason: fellOver ? normalizeFallbackReason(attempts) : void 0,
-        unreachableFiltered
+        unreachableFiltered,
+        policyBlockedFiltered
       };
     }
     attempts.push({
@@ -2435,8 +3125,9 @@ async function call(ir, opts = {}) {
     promptPreview: extractPromptPreview(ir)
   });
   const filteredNote = unreachableFiltered && unreachableFiltered.length > 0 ? ` (also auto-filtered: [${unreachableFiltered.join(", ")}] \u2014 no API key)` : "";
+  const blockedNote = policyBlockedFiltered && policyBlockedFiltered.length > 0 ? ` (also policy-blocked: [${policyBlockedFiltered.join(", ")}])` : "";
   throw new CallError(
-    `call(): all attempts failed${lastErr ? ` \u2014 ${lastErr.errorCode}: ${lastErr.message}` : ""}${filteredNote}`,
+    `call(): all attempts failed${lastErr ? ` \u2014 ${lastErr.errorCode}: ${lastErr.message}` : ""}${filteredNote}${blockedNote}`,
     attempts,
     lastErr?.status,
     lastErr?.errorCode
@@ -2591,162 +3282,6 @@ function clamp(n) {
   return Math.max(0, Math.min(1, n));
 }
 
-// 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 = {
-  // Reasoning floor — never degrade. Walk UP on 429 to Opus → cross-provider.
-  critique: [
-    "claude-opus-4-7",
-    "claude-sonnet-4-6",
-    "gemini-2.5-pro"
-  ],
-  // Reasoning matters — Sonnet primary; walk UP to Opus on 429 (rare exception
-  // to "always cheaper"); cross-provider via Pro; DeepSeek Pro as tier 3 floor.
-  plan: [
-    "claude-sonnet-4-6",
-    "claude-opus-4-7",
-    "gemini-2.5-pro",
-    "deepseek-v4-pro"
-  ],
-  // Quality + cost match. Walk Sonnet → Haiku same-provider, Pro cross,
-  // Flash floor for the open-posture chain.
-  generate: [
-    "claude-sonnet-4-6",
-    "claude-haiku-4-5",
-    "gemini-2.5-pro",
-    "gemini-2.5-flash"
-  ],
-  ask: [
-    "claude-sonnet-4-6",
-    "claude-haiku-4-5",
-    "gemini-2.5-pro",
-    "gemini-2.5-flash"
-  ],
-  // Structured-output archetype — Flash skipped (alpha.8 MAX_TOKENS cliff),
-  // DeepSeek skipped (no brain evidence). Floor at Haiku.
-  extract: [
-    "claude-sonnet-4-6",
-    "claude-haiku-4-5",
-    "gemini-2.5-pro"
-  ],
-  // Forgiving archetype — Sonnet primary but Flash safely floors it.
-  transform: [
-    "claude-sonnet-4-6",
-    "claude-haiku-4-5",
-    "gemini-2.5-pro",
-    "gemini-2.5-flash"
-  ],
-  // Parallel-tool throughput champion (Flash, L-040). Tier 1 cross-provider
-  // Pro; tier 2 Sonnet (quality safety net for blocked-Flash case); tier 3
-  // Haiku (reduced tool budget — cliff at 16 fires).
-  hunt: [
-    "gemini-2.5-flash",
-    "gemini-2.5-pro",
-    "claude-sonnet-4-6",
-    "claude-haiku-4-5"
-  ],
-  // Cost-sensitive + tolerant. DeepSeek brain-evidence tier 1; Haiku tier 2
-  // for quality safety; Flash-Lite emergency floor (onboarded s22).
-  summarize: [
-    "gemini-2.5-flash",
-    "deepseek-v4-flash",
-    "claude-haiku-4-5",
-    "gemini-2.5-flash-lite"
-  ],
-  // Brain-validated DeepSeek tier 1 (169 rows, 0% empty); Haiku tier 2;
-  // Flash-Lite floor for repeat-prompt workloads (cache-discount 10×).
-  classify: [
-    "gemini-2.5-flash",
-    "deepseek-v4-flash",
-    "claude-haiku-4-5",
-    "gemini-2.5-flash-lite"
-  ]
-};
-function getDefaultFallbackChain(opts) {
-  const { archetype, primary, maxDepth = 3, policy, reachability } = opts;
-  if (maxDepth < 1) {
-    throw new Error(
-      `getDefaultFallbackChain: maxDepth must be >= 1, got ${maxDepth}`
-    );
-  }
-  const allChains = loadChainsFromBrain();
-  const starter = allChains[archetype];
-  if (!starter) {
-    throw new Error(
-      `getDefaultFallbackChain: unknown archetype "${archetype}". Known: ${Object.keys(allChains).join(", ")}`
-    );
-  }
-  let chain;
-  if (primary) {
-    chain = [primary, ...starter.filter((id) => id !== primary)];
-  } else {
-    chain = [...starter];
-  }
-  if (policy?.blockedModels && policy.blockedModels.length > 0) {
-    const blocked = new Set(policy.blockedModels);
-    chain = chain.filter((id) => !blocked.has(id));
-  }
-  const seen = /* @__PURE__ */ new Set();
-  const deduped = [];
-  for (const id of chain) {
-    if (!seen.has(id)) {
-      seen.add(id);
-      deduped.push(id);
-    }
-  }
-  let filtered = deduped;
-  if (reachability) {
-    filtered = deduped.filter((id) => isModelReachable(id, reachability));
-  }
-  return filtered.slice(0, maxDepth);
-}
-function getStarterChain(archetype) {
-  const chain = STARTER_CHAINS[archetype];
-  if (!chain) {
-    throw new Error(
-      `getStarterChain: unknown archetype "${archetype}"`
-    );
-  }
-  return [...chain];
-}
-function getAllStarterChains() {
-  const out = {};
-  for (const [archetype, chain] of Object.entries(STARTER_CHAINS)) {
-    out[archetype] = [...chain];
-  }
-  return out;
-}
-
 // src/archetype-perf-brain.ts
 function isPerfRow(x) {
   if (!x || typeof x !== "object") return false;