@warmdrift/kgauto-compiler 2.0.0-alpha.13 → 2.0.0-alpha.15

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 = [];
@@ -1438,6 +1474,211 @@ var PROFILES_RAW = [
       hunt: 4
       // sequential tools — same as V4-Flash
     }
+  },
+  // ── 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 = {
@@ -2288,56 +2529,257 @@ function tryParseJson(s) {
   }
 }
 
-// src/call.ts
-async function call(ir, opts = {}) {
-  const initial = compileAndRegister(ir, opts);
-  const start = Date.now();
-  const attempts = [];
-  const rawTargets = [initial.target, ...initial.fallbackChain];
-  let unreachableFiltered;
-  let targetsToTry;
-  if (opts.noAutoFilter) {
-    targetsToTry = rawTargets;
-  } else {
-    const dropped = [];
-    targetsToTry = [];
-    for (const t of rawTargets) {
-      if (isModelReachable(t, { apiKeys: opts.apiKeys })) {
-        targetsToTry.push(t);
-      } else {
-        dropped.push(t);
-      }
-    }
-    unreachableFiltered = dropped;
-    if (targetsToTry.length === 0) {
-      const latencyMs2 = Date.now() - start;
-      await record({
-        handle: initial.handle,
-        tokensIn: 0,
-        tokensOut: 0,
-        latencyMs: latencyMs2,
-        success: false,
-        errorType: "no_reachable_models",
-        promptPreview: extractPromptPreview(ir)
-      });
-      const noReachableAttempts = dropped.map((m) => ({
-        model: m,
-        status: "terminal",
-        errorCode: "unreachable_provider",
-        message: `No API key for ${m}'s provider \u2014 set one of PROVIDER_ENV_KEYS or pass apiKeys`
-      }));
-      throw new CallError(
-        `call(): no reachable models in chain. Filtered: [${dropped.join(", ")}]. Add a key for one provider, or pass apiKeys.`,
-        noReachableAttempts,
-        void 0,
-        "no_reachable_models"
+// 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;
+}
+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
+async function call(ir, opts = {}) {
+  const initial = compileAndRegister(ir, opts);
+  const start = Date.now();
+  const attempts = [];
+  const rawTargets = [initial.target, ...initial.fallbackChain];
+  let unreachableFiltered;
+  let targetsToTry;
+  if (opts.noAutoFilter) {
+    targetsToTry = rawTargets;
+  } else {
+    const dropped = [];
+    targetsToTry = [];
+    for (const t of rawTargets) {
+      if (isModelReachable(t, { apiKeys: opts.apiKeys })) {
+        targetsToTry.push(t);
+      } else {
+        dropped.push(t);
+      }
+    }
+    unreachableFiltered = dropped;
+    if (targetsToTry.length === 0) {
+      const latencyMs2 = Date.now() - start;
+      await record({
+        handle: initial.handle,
+        tokensIn: 0,
+        tokensOut: 0,
+        latencyMs: latencyMs2,
+        success: false,
+        errorType: "no_reachable_models",
+        promptPreview: extractPromptPreview(ir)
+      });
+      const noReachableAttempts = dropped.map((m) => ({
+        model: m,
+        status: "terminal",
+        errorCode: "unreachable_provider",
+        message: `No API key for ${m}'s provider \u2014 set one of PROVIDER_ENV_KEYS or pass apiKeys`
+      }));
+      throw new CallError(
+        `call(): no reachable models in chain. Filtered: [${dropped.join(", ")}]. Add a key for one provider, or pass apiKeys.`,
+        noReachableAttempts,
+        void 0,
+        "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 activeCompile = initial;
   let lastErr;
+  const failedProviders = /* @__PURE__ */ new Set();
   for (let i = 0; i < targetsToTry.length; i++) {
     const targetModel = targetsToTry[i];
+    const targetProfile = tryGetProfile(targetModel);
+    if (targetProfile && failedProviders.has(targetProfile.provider) && !opts.noFallback) {
+      attempts.push({
+        model: targetModel,
+        status: "terminal",
+        errorCode: "auth_inferred",
+        message: `Skipped \u2014 provider ${targetProfile.provider} returned 401/403 earlier in this call; same key inferred to fail`
+      });
+      continue;
+    }
     if (targetModel !== initial.target) {
       try {
         activeCompile = compileAndRegister(
@@ -2406,6 +2848,10 @@ async function call(ir, opts = {}) {
     });
     lastErr = validated;
     if (validated.errorType === "terminal" || opts.noFallback) {
+      if (validated.errorCode === "auth" && !opts.noFallback && activeCompile.provider) {
+        failedProviders.add(activeCompile.provider);
+        continue;
+      }
       break;
     }
   }
@@ -2484,6 +2930,7 @@ function normalizeFallbackReason(attempts) {
     return "cliff";
   }
   if (code === "cost_cap_exceeded") return "cost_cap";
+  if (code === "auth" || code === "auth_inferred") return "provider_auth_failed";
   return "provider_error";
 }
 
@@ -2575,162 +3022,6 @@ 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;