@warmdrift/kgauto-compiler 2.0.0-alpha.64 → 2.0.0-alpha.66

package/dist/{chunk-QKXTMVCT.mjs → chunk-4UO4CCSP.mjs}

@@ -1413,6 +1413,156 @@ var PROFILES_RAW = [
       classify: 8
       // overkill
     }
+  },
+  // ── Z.ai (GLM) ──
+  // alpha.65 (2026-07-18): first Z.ai profile, provider `zai` added same
+  // release. Step-zero sources (L-049/L-081, all fetched 2026-07-18):
+  // docs.z.ai/guides/llm/glm-5.2 (capabilities), docs.z.ai api-reference
+  // chat-completion (wire params: thinking.type enabled|disabled default
+  // enabled, reasoning_effort max..none, tool_stream, max output 131072),
+  // docs.z.ai/guides/overview/pricing ($1.4 in / $4.4 out, cached input
+  // $0.26, cache storage limited-time free), vLLM recipe (native ctx
+  // 1,048,576). NO AI-trained numbers.
+  {
+    id: "glm-5.2",
+    verifiedAgainstDocs: "2026-07-18",
+    provider: "zai",
+    status: "preview",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 131072,
+    // No documented tool-count cap — conservative default pending measured
+    // evidence; matches the deepseek posture, not a doc number.
+    maxTools: 16,
+    // Hosted-API docs document tools + tool_choice + tool_stream but say
+    // NOTHING about multiple tool_calls per response. Conservative false so
+    // hunt-shaped ranking doesn't over-promise; flip on doc or probe
+    // evidence (L-081: undocumented ≠ true).
+    parallelToolCalls: false,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 1.4,
+    costOutputPer1m: 4.4,
+    lowering: {
+      system: { mode: "inline" },
+      // Automatic prefix caching ("intelligent caching mechanism") — no wire
+      // marker to emit, DeepSeek-style. Cache-hit input $0.26/M (~0.19× of
+      // miss) tracked in notes; storage limited-time free as of 2026-07-18.
+      cache: { strategy: "unsupported" },
+      tools: { format: "openai" },
+      // thinking.type: 'enabled' (server default) | 'disabled'. lowerZai
+      // emits an explicit disable only on forceThinkingZero cliffs.
+      thinking: { field: "thinking.type", default: "auto" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    strengths: ["cost", "1m_context", "code", "tool_use", "json_output", "reasoning"],
+    weaknesses: ["parallel_tools"],
+    notes: "GLM-5.2 (Z.ai, 2026-06): agentic-coding flagship. 1,048,576 ctx / 131,072 max out. $1.40/$4.40 per 1M; cached input $0.26/M (automatic caching, no marker; storage limited-time free as of 2026-07-18). Thinking on by default (`thinking.type`), `reasoning_effort` max..none. Parallel tool calls UNDOCUMENTED on the hosted API \u2014 profile says false until doc or probe evidence. status:preview \u2014 no brain evidence yet; earns placement via the machinery.",
+    // Starter hypotheses — verify with telemetry/probes; NO brain evidence
+    // yet. Anchored on the vendor's agentic-coding positioning at a
+    // mid-cost tier, discounted for the unproven parallel-tool story.
+    archetypePerf: {
+      generate: 7,
+      // agentic-coding positioning — starter hypothesis
+      transform: 7,
+      // starter hypothesis
+      plan: 7,
+      // starter hypothesis
+      summarize: 7,
+      // tolerant archetype — starter hypothesis
+      ask: 6,
+      extract: 6,
+      critique: 6,
+      classify: 6,
+      hunt: 5
+      // tools work; parallel unproven → discounted
+    }
+  },
+  // ── Moonshot (Kimi) ──
+  // alpha.65 (2026-07-18): first Moonshot profile, provider `moonshot` added
+  // same release. Step-zero sources (all fetched 2026-07-18):
+  // platform.kimi.ai/docs/pricing/chat-k3 ($3.00 cache-miss in / $0.30
+  // cache-hit in / $15.00 out; 1,048,576 ctx; ToolCalls + JSON Mode +
+  // Partial Mode), kimi-k3-quickstart (max_completion_tokens default
+  // 131072, up to 1,048,576; base https://api.moonshot.ai/v1; vision via
+  // base64 or ms://file-id, NO public image URLs), models-overview
+  // (temperature FIXED 1.0, top_p FIXED 0.95, n FIXED 1, always reasons,
+  // reasoning_effort 'max' only), tool-calls guide ("can choose to call
+  // multiple tools at once … will tend to call them in parallel").
+  {
+    id: "kimi-k3",
+    verifiedAgainstDocs: "2026-07-18",
+    provider: "moonshot",
+    status: "preview",
+    maxContextTokens: 1048576,
+    // Documented ceiling; default is 131,072 when unset.
+    maxOutputTokens: 1048576,
+    // No hard doc cap; docs recommend dynamic tool loading at "dozens or
+    // hundreds" — 64 matches the frontier-agentic posture (sonnet-5 tier).
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 3,
+    costOutputPer1m: 15,
+    lowering: {
+      system: { mode: "inline" },
+      // Automatic context caching — no wire marker. Cache-hit input $0.30/M
+      // (0.1× of miss) tracked in notes.
+      cache: { strategy: "unsupported" },
+      tools: { format: "openai" }
+      // No thinking knob: reasoning is always-on and reasoning_effort
+      // accepts only 'max' (= the server default). Nothing to lower.
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    strengths: ["quality", "reasoning", "1m_context", "parallel_tools", "tool_use", "code", "vision"],
+    // 'latency' → latencyTierOf derives 'slow': always-on max-effort
+    // reasoner with no off switch; no measured p50 yet, the tag is the
+    // honest prior (reasoner family precedent: deepseek-v4-pro ~48s).
+    weaknesses: ["cost", "latency"],
+    notes: "Kimi K3 (Moonshot, API live 2026-07-16): 2.8T-param open-weight flagship, native vision. 1,048,576 ctx; max_completion_tokens default 131,072, ceiling 1,048,576. $3.00/$15.00 per 1M, cache-hit input $0.30/M (automatic caching). Reasoning ALWAYS ON (reasoning_effort max only, cannot disable; responses may carry reasoning_content). Wire quirks: temperature fixed 1.0 / top_p 0.95 / n 1 (kgauto never emits those); vision input must be base64 or ms://file-id \u2014 public image URLs rejected. status:preview \u2014 earns placement via the machinery.",
+    // Starter hypotheses — verify with telemetry/probes; NO brain evidence
+    // yet. Anchored on documented long-horizon-agentic + parallel-tool
+    // strengths; discounted on terse archetypes where always-on reasoning
+    // bills $15/M output for thinking the archetype doesn't need.
+    archetypePerf: {
+      hunt: 8,
+      // parallel tools + long-horizon agentic — starter hypothesis
+      plan: 8,
+      // starter hypothesis
+      generate: 8,
+      // starter hypothesis
+      critique: 7,
+      ask: 7,
+      extract: 7,
+      transform: 7,
+      summarize: 6,
+      // works, but $15/M output reasoning tax on a tolerant archetype
+      classify: 5
+      // always-on reasoning cost + reasoner hedge risk
+    },
+    // Diagnostics-only (no prompt rewrite): reasoner-family hedge risk on
+    // decisive archetypes, carried as a warning until K3-specific probe
+    // evidence exists. The deepseek-reasoner family showed 8/10 judge
+    // rationales citing hedging on classify (exclusion-finding ID 20,
+    // 2026-05-28); K3 is an always-on reasoner and inherits the risk class,
+    // not the conviction.
+    archetypeConventions: [
+      {
+        archetype: "classify",
+        cliffWarning: "kimi-k3 is an always-on reasoner: decisive archetypes (classify) may hedge, and reasoning tokens bill at $15/M output. Family-prior risk (deepseek-reasoner precedent), not K3-measured \u2014 probe before relying on it for classify.",
+        reason: "Reasoner-family prior (exclusion-finding ID 20, 2026-05-28, deepseek-v4-pro on tt-intel/classify). No K3-specific evidence yet \u2014 warning only, no prompt rewrite."
+      }
+    ]
   }
 ];
 var ALIASES = {

package/dist/{chunk-DGPYLFDG.mjs → chunk-CMB6PZWS.mjs}

@@ -1,5 +1,5 @@
 // src/version.ts
-var LIBRARY_VERSION = "2.0.0-alpha.64";
+var LIBRARY_VERSION = "2.0.0-alpha.66";
 
 // src/key-health.ts
 var JSON_HEADERS = { "Content-Type": "application/json" };
@@ -48,6 +48,14 @@ var PROBE_SPECS = [
       return Number.isFinite(n) ? n : void 0;
     }
   },
+  {
+    provider: "moonshot",
+    canonicalEnvName: "MOONSHOT_API_KEY",
+    buildRequest: (key) => ({
+      url: "https://api.moonshot.ai/v1/models",
+      headers: { Authorization: `Bearer ${key}` }
+    })
+  },
   {
     provider: "google",
     canonicalEnvName: "GEMINI_API_KEY",
@@ -79,6 +87,11 @@ function createKeyHealthRoute(config) {
       const aiSdk = envKey(env, "GOOGLE_GENERATIVE_AI_API_KEY");
       key = gemini ?? google ?? aiSdk;
       envName = gemini ? "GEMINI_API_KEY" : google ? "GOOGLE_API_KEY" : aiSdk ? "GOOGLE_GENERATIVE_AI_API_KEY" : "GEMINI_API_KEY";
+    } else if (spec.provider === "moonshot") {
+      const moonshot = envKey(env, "MOONSHOT_API_KEY");
+      const kimi = envKey(env, "KIMI_API_KEY");
+      key = moonshot ?? kimi;
+      envName = moonshot ? "MOONSHOT_API_KEY" : kimi ? "KIMI_API_KEY" : "MOONSHOT_API_KEY";
     } else {
       key = envKey(env, spec.canonicalEnvName);
     }

package/dist/{chunk-YZRPNSSQ.mjs → chunk-SBFSYCQG.mjs}

@@ -1,6 +1,6 @@
 import {
   tryGetProfile
-} from "./chunk-QKXTMVCT.mjs";
+} from "./chunk-4UO4CCSP.mjs";
 
 // src/brain-query.ts
 var FRESH_SNAPSHOT = {
@@ -258,7 +258,9 @@ var SUPPORTED_PROVIDERS = Object.freeze([
   "anthropic",
   "google",
   "openai",
-  "deepseek"
+  "deepseek",
+  "zai",
+  "moonshot"
 ]);
 function isSupportedProvider(p) {
   return SUPPORTED_PROVIDERS.includes(p);
@@ -271,7 +273,15 @@ var PROVIDER_ENV_KEYS = Object.freeze({
     "GOOGLE_GENERATIVE_AI_API_KEY"
   ]),
   openai: Object.freeze(["OPENAI_API_KEY"]),
-  deepseek: Object.freeze(["DEEPSEEK_API_KEY"])
+  deepseek: Object.freeze(["DEEPSEEK_API_KEY"]),
+  // alpha.65 — Z.ai (GLM family). ZAI_API_KEY is the canonical name; the
+  // Z_AI_API_KEY variant covers the underscore convention some tooling uses
+  // for the brand's "Z.ai" spelling.
+  zai: Object.freeze(["ZAI_API_KEY", "Z_AI_API_KEY"]),
+  // alpha.65 — Moonshot AI (Kimi family). MOONSHOT_API_KEY is canonical
+  // (api.moonshot.ai); KIMI_API_KEY covers the platform.kimi.ai rebrand
+  // surface so consumers keyed under either name resolve.
+  moonshot: Object.freeze(["MOONSHOT_API_KEY", "KIMI_API_KEY"])
 });
 function defaultEnv() {
   return typeof process !== "undefined" && process.env ? process.env : {};
@@ -323,6 +333,10 @@ function readBrainReadEnv(envSource = defaultEnv()) {
   if (!anonKey) missingEnv.push("KGAUTO_V2_BRAIN_ANON_KEY");
   return { endpoint, jwt, anonKey, missingEnv };
 }
+function isSameModelRetryEnabledFromEnv(envSource = defaultEnv()) {
+  const raw = (envSource.KGAUTO_SAME_MODEL_RETRY ?? "").trim().toLowerCase();
+  return raw === "1" || raw === "true";
+}
 
 // src/fallback.ts
 var STARTER_CHAINS_GROUNDED = {
@@ -691,6 +705,7 @@ export {
   isModelReachable,
   getReachabilityDiagnostic,
   readBrainReadEnv,
+  isSameModelRetryEnabledFromEnv,
   loadChainsFromBrain,
   getDefaultFallbackChain,
   getStarterChain,

package/dist/glassbox/index.d.mts

@@ -1,6 +1,6 @@
-import { G as GlassboxEvent } from '../types-axAX1Bg0.mjs';
-export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-axAX1Bg0.mjs';
-import '../ir-B2dRyJXO.mjs';
+import { G as GlassboxEvent } from '../types-8pWRDbUg.mjs';
+export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-8pWRDbUg.mjs';
+import '../ir-BTvyl8-w.mjs';
 import '../dialect.mjs';
 
 /**

package/dist/glassbox/index.d.ts

@@ -1,6 +1,6 @@
-import { G as GlassboxEvent } from '../types-C-Wp7HpI.js';
-export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-C-Wp7HpI.js';
-import '../ir-ClU56aBc.js';
+import { G as GlassboxEvent } from '../types-BdYlJz0Z.js';
+export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-BdYlJz0Z.js';
+import '../ir-Qnw6L265.js';
 import '../dialect.js';
 
 /**

package/dist/glassbox-routes/format.d.mts

@@ -1,5 +1,5 @@
-import { T as TraceHealth } from '../types-CSDweZsl.mjs';
-import '../ir-B2dRyJXO.mjs';
+import { T as TraceHealth } from '../types-Cya0jPAF.mjs';
+import '../ir-BTvyl8-w.mjs';
 import '../dialect.mjs';
 
 /**

package/dist/glassbox-routes/format.d.ts

@@ -1,5 +1,5 @@
-import { T as TraceHealth } from '../types-ByN3r0_7.js';
-import '../ir-ClU56aBc.js';
+import { T as TraceHealth } from '../types-DxYXuFxj.js';
+import '../ir-Qnw6L265.js';
 import '../dialect.js';
 
 /**

package/dist/glassbox-routes/index.d.mts

@@ -1,7 +1,7 @@
-import { G as GlassboxEvent } from '../types-axAX1Bg0.mjs';
-import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-CSDweZsl.mjs';
-export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-CSDweZsl.mjs';
-import '../ir-B2dRyJXO.mjs';
+import { G as GlassboxEvent } from '../types-8pWRDbUg.mjs';
+import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-Cya0jPAF.mjs';
+export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-Cya0jPAF.mjs';
+import '../ir-BTvyl8-w.mjs';
 import '../dialect.mjs';
 
 /**

package/dist/glassbox-routes/index.d.ts

@@ -1,7 +1,7 @@
-import { G as GlassboxEvent } from '../types-C-Wp7HpI.js';
-import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-ByN3r0_7.js';
-export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-ByN3r0_7.js';
-import '../ir-ClU56aBc.js';
+import { G as GlassboxEvent } from '../types-BdYlJz0Z.js';
+import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-DxYXuFxj.js';
+export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-DxYXuFxj.js';
+import '../ir-Qnw6L265.js';
 import '../dialect.js';
 
 /**

package/dist/glassbox-routes/index.js

@@ -1467,6 +1467,156 @@ var PROFILES_RAW = [
       classify: 8
       // overkill
     }
+  },
+  // ── Z.ai (GLM) ──
+  // alpha.65 (2026-07-18): first Z.ai profile, provider `zai` added same
+  // release. Step-zero sources (L-049/L-081, all fetched 2026-07-18):
+  // docs.z.ai/guides/llm/glm-5.2 (capabilities), docs.z.ai api-reference
+  // chat-completion (wire params: thinking.type enabled|disabled default
+  // enabled, reasoning_effort max..none, tool_stream, max output 131072),
+  // docs.z.ai/guides/overview/pricing ($1.4 in / $4.4 out, cached input
+  // $0.26, cache storage limited-time free), vLLM recipe (native ctx
+  // 1,048,576). NO AI-trained numbers.
+  {
+    id: "glm-5.2",
+    verifiedAgainstDocs: "2026-07-18",
+    provider: "zai",
+    status: "preview",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 131072,
+    // No documented tool-count cap — conservative default pending measured
+    // evidence; matches the deepseek posture, not a doc number.
+    maxTools: 16,
+    // Hosted-API docs document tools + tool_choice + tool_stream but say
+    // NOTHING about multiple tool_calls per response. Conservative false so
+    // hunt-shaped ranking doesn't over-promise; flip on doc or probe
+    // evidence (L-081: undocumented ≠ true).
+    parallelToolCalls: false,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 1.4,
+    costOutputPer1m: 4.4,
+    lowering: {
+      system: { mode: "inline" },
+      // Automatic prefix caching ("intelligent caching mechanism") — no wire
+      // marker to emit, DeepSeek-style. Cache-hit input $0.26/M (~0.19× of
+      // miss) tracked in notes; storage limited-time free as of 2026-07-18.
+      cache: { strategy: "unsupported" },
+      tools: { format: "openai" },
+      // thinking.type: 'enabled' (server default) | 'disabled'. lowerZai
+      // emits an explicit disable only on forceThinkingZero cliffs.
+      thinking: { field: "thinking.type", default: "auto" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    strengths: ["cost", "1m_context", "code", "tool_use", "json_output", "reasoning"],
+    weaknesses: ["parallel_tools"],
+    notes: "GLM-5.2 (Z.ai, 2026-06): agentic-coding flagship. 1,048,576 ctx / 131,072 max out. $1.40/$4.40 per 1M; cached input $0.26/M (automatic caching, no marker; storage limited-time free as of 2026-07-18). Thinking on by default (`thinking.type`), `reasoning_effort` max..none. Parallel tool calls UNDOCUMENTED on the hosted API \u2014 profile says false until doc or probe evidence. status:preview \u2014 no brain evidence yet; earns placement via the machinery.",
+    // Starter hypotheses — verify with telemetry/probes; NO brain evidence
+    // yet. Anchored on the vendor's agentic-coding positioning at a
+    // mid-cost tier, discounted for the unproven parallel-tool story.
+    archetypePerf: {
+      generate: 7,
+      // agentic-coding positioning — starter hypothesis
+      transform: 7,
+      // starter hypothesis
+      plan: 7,
+      // starter hypothesis
+      summarize: 7,
+      // tolerant archetype — starter hypothesis
+      ask: 6,
+      extract: 6,
+      critique: 6,
+      classify: 6,
+      hunt: 5
+      // tools work; parallel unproven → discounted
+    }
+  },
+  // ── Moonshot (Kimi) ──
+  // alpha.65 (2026-07-18): first Moonshot profile, provider `moonshot` added
+  // same release. Step-zero sources (all fetched 2026-07-18):
+  // platform.kimi.ai/docs/pricing/chat-k3 ($3.00 cache-miss in / $0.30
+  // cache-hit in / $15.00 out; 1,048,576 ctx; ToolCalls + JSON Mode +
+  // Partial Mode), kimi-k3-quickstart (max_completion_tokens default
+  // 131072, up to 1,048,576; base https://api.moonshot.ai/v1; vision via
+  // base64 or ms://file-id, NO public image URLs), models-overview
+  // (temperature FIXED 1.0, top_p FIXED 0.95, n FIXED 1, always reasons,
+  // reasoning_effort 'max' only), tool-calls guide ("can choose to call
+  // multiple tools at once … will tend to call them in parallel").
+  {
+    id: "kimi-k3",
+    verifiedAgainstDocs: "2026-07-18",
+    provider: "moonshot",
+    status: "preview",
+    maxContextTokens: 1048576,
+    // Documented ceiling; default is 131,072 when unset.
+    maxOutputTokens: 1048576,
+    // No hard doc cap; docs recommend dynamic tool loading at "dozens or
+    // hundreds" — 64 matches the frontier-agentic posture (sonnet-5 tier).
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 3,
+    costOutputPer1m: 15,
+    lowering: {
+      system: { mode: "inline" },
+      // Automatic context caching — no wire marker. Cache-hit input $0.30/M
+      // (0.1× of miss) tracked in notes.
+      cache: { strategy: "unsupported" },
+      tools: { format: "openai" }
+      // No thinking knob: reasoning is always-on and reasoning_effort
+      // accepts only 'max' (= the server default). Nothing to lower.
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    strengths: ["quality", "reasoning", "1m_context", "parallel_tools", "tool_use", "code", "vision"],
+    // 'latency' → latencyTierOf derives 'slow': always-on max-effort
+    // reasoner with no off switch; no measured p50 yet, the tag is the
+    // honest prior (reasoner family precedent: deepseek-v4-pro ~48s).
+    weaknesses: ["cost", "latency"],
+    notes: "Kimi K3 (Moonshot, API live 2026-07-16): 2.8T-param open-weight flagship, native vision. 1,048,576 ctx; max_completion_tokens default 131,072, ceiling 1,048,576. $3.00/$15.00 per 1M, cache-hit input $0.30/M (automatic caching). Reasoning ALWAYS ON (reasoning_effort max only, cannot disable; responses may carry reasoning_content). Wire quirks: temperature fixed 1.0 / top_p 0.95 / n 1 (kgauto never emits those); vision input must be base64 or ms://file-id \u2014 public image URLs rejected. status:preview \u2014 earns placement via the machinery.",
+    // Starter hypotheses — verify with telemetry/probes; NO brain evidence
+    // yet. Anchored on documented long-horizon-agentic + parallel-tool
+    // strengths; discounted on terse archetypes where always-on reasoning
+    // bills $15/M output for thinking the archetype doesn't need.
+    archetypePerf: {
+      hunt: 8,
+      // parallel tools + long-horizon agentic — starter hypothesis
+      plan: 8,
+      // starter hypothesis
+      generate: 8,
+      // starter hypothesis
+      critique: 7,
+      ask: 7,
+      extract: 7,
+      transform: 7,
+      summarize: 6,
+      // works, but $15/M output reasoning tax on a tolerant archetype
+      classify: 5
+      // always-on reasoning cost + reasoner hedge risk
+    },
+    // Diagnostics-only (no prompt rewrite): reasoner-family hedge risk on
+    // decisive archetypes, carried as a warning until K3-specific probe
+    // evidence exists. The deepseek-reasoner family showed 8/10 judge
+    // rationales citing hedging on classify (exclusion-finding ID 20,
+    // 2026-05-28); K3 is an always-on reasoner and inherits the risk class,
+    // not the conviction.
+    archetypeConventions: [
+      {
+        archetype: "classify",
+        cliffWarning: "kimi-k3 is an always-on reasoner: decisive archetypes (classify) may hedge, and reasoning tokens bill at $15/M output. Family-prior risk (deepseek-reasoner precedent), not K3-measured \u2014 probe before relying on it for classify.",
+        reason: "Reasoner-family prior (exclusion-finding ID 20, 2026-05-28, deepseek-v4-pro on tt-intel/classify). No K3-specific evidence yet \u2014 warning only, no prompt rewrite."
+      }
+    ]
   }
 ];
 var ALIASES = {
@@ -1493,7 +1643,9 @@ var SUPPORTED_PROVIDERS = Object.freeze([
   "anthropic",
   "google",
   "openai",
-  "deepseek"
+  "deepseek",
+  "zai",
+  "moonshot"
 ]);
 function isSupportedProvider(p) {
   return SUPPORTED_PROVIDERS.includes(p);
@@ -1506,7 +1658,15 @@ var PROVIDER_ENV_KEYS = Object.freeze({
     "GOOGLE_GENERATIVE_AI_API_KEY"
   ]),
   openai: Object.freeze(["OPENAI_API_KEY"]),
-  deepseek: Object.freeze(["DEEPSEEK_API_KEY"])
+  deepseek: Object.freeze(["DEEPSEEK_API_KEY"]),
+  // alpha.65 — Z.ai (GLM family). ZAI_API_KEY is the canonical name; the
+  // Z_AI_API_KEY variant covers the underscore convention some tooling uses
+  // for the brand's "Z.ai" spelling.
+  zai: Object.freeze(["ZAI_API_KEY", "Z_AI_API_KEY"]),
+  // alpha.65 — Moonshot AI (Kimi family). MOONSHOT_API_KEY is canonical
+  // (api.moonshot.ai); KIMI_API_KEY covers the platform.kimi.ai rebrand
+  // surface so consumers keyed under either name resolve.
+  moonshot: Object.freeze(["MOONSHOT_API_KEY", "KIMI_API_KEY"])
 });
 function defaultEnv() {
   return typeof process !== "undefined" && process.env ? process.env : {};

package/dist/glassbox-routes/index.mjs

@@ -1,10 +1,10 @@
 import {
   ARCHETYPE_FLOOR_DEFAULT,
   getDefaultFallbackChain
-} from "../chunk-YZRPNSSQ.mjs";
+} from "../chunk-SBFSYCQG.mjs";
 import {
   tryGetProfile
-} from "../chunk-QKXTMVCT.mjs";
+} from "../chunk-4UO4CCSP.mjs";
 import {
   subscribe,
   subscribeApp

package/dist/glassbox-routes/react/index.d.mts

@@ -1,6 +1,6 @@
 import * as react_jsx_runtime from 'react/jsx-runtime';
-import { a as TraceDetail } from '../../types-CSDweZsl.mjs';
-import '../../ir-B2dRyJXO.mjs';
+import { a as TraceDetail } from '../../types-Cya0jPAF.mjs';
+import '../../ir-BTvyl8-w.mjs';
 import '../../dialect.mjs';
 
 /**

package/dist/glassbox-routes/react/index.d.ts

@@ -1,6 +1,6 @@
 import * as react_jsx_runtime from 'react/jsx-runtime';
-import { a as TraceDetail } from '../../types-ByN3r0_7.js';
-import '../../ir-ClU56aBc.js';
+import { a as TraceDetail } from '../../types-DxYXuFxj.js';
+import '../../ir-Qnw6L265.js';
 import '../../dialect.js';
 
 /**

package/dist/index.d.mts

@@ -1,5 +1,5 @@
-import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-B2dRyJXO.mjs';
-export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, E as EffortLevel, F as FallbackReason, r as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, s as MutationApplied, t as NormalizedTokens, u as OutcomeKind, v as PerAxisMetricsByModel, w as PromptSection, x as SectionKind, y as ShadowProbeConfig, T as ToolCall, z as ToolDefinition, D as captureGoldenIr, J as parseGoldenCaptureRate, K as resolveGoldenCaptureRate, L as shouldCaptureGolden } from './ir-B2dRyJXO.mjs';
+import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-BTvyl8-w.mjs';
+export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, E as EffortLevel, F as FallbackReason, r as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, s as MutationApplied, t as NormalizedTokens, u as OutcomeKind, v as PerAxisMetricsByModel, w as PromptSection, x as SectionKind, y as ShadowProbeConfig, T as ToolCall, z as ToolDefinition, D as captureGoldenIr, J as parseGoldenCaptureRate, K as resolveGoldenCaptureRate, L as shouldCaptureGolden } from './ir-BTvyl8-w.mjs';
 import { ModelProfile, ArchetypeConvention } from './profiles.mjs';
 export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, allProfiles, getProfile, latencyTierOf, profilesByProvider, tryGetProfile } from './profiles.mjs';
 export { BrainForwardConfig, BrainForwardRoutes, createBrainForwardRoutes } from './brain-proxy.mjs';
@@ -78,6 +78,19 @@ interface ExecuteErr {
     errorCode: string;
     message: string;
     raw: unknown;
+    /**
+     * alpha.66 — real token usage when the failure is a SYNTHESIZED
+     * structured-output contract violation (the provider returned 200 with
+     * usage; the response just violated the contract). Absent on transport-
+     * level failures (4xx/5xx/network) where no usage exists. Lets the
+     * brain's failure rows carry true `tokens_in` for the cliff class —
+     * PB's 2026-05-18 `chain-walk-on-structured-output-parse-failed` ask 4
+     * (tokens_in=0 on a 72s parse-fail made diagnosis ambiguous).
+     */
+    tokens?: {
+        input: number;
+        output: number;
+    };
 }
 type ExecuteResult = ExecuteOk | ExecuteErr;
 declare function execute(request: CompiledRequest, opts?: ExecuteOptions): Promise<ExecuteResult>;
@@ -597,6 +610,12 @@ interface OutcomePayload {
     system_prompt_chars?: number;
     fell_over_from?: string;
     fallback_reason?: 'rate_limit' | 'provider_auth_failed' | 'provider_error' | 'cliff' | 'cost_cap' | 'contract_violation';
+    /**
+     * True when the same-model retry (CallOptions.sameModelRetry) fired
+     * during this call. Key OMITTED (not false) when no retry fired — safe
+     * against brains that haven't applied migration 039 yet.
+     */
+    retried_same_model?: boolean;
     /**
      * Per-call SectionRewrite[] captured at compile time. Omitted (sent as
      * undefined → stored NULL) when no rewrites fired. Powers cross-app
@@ -923,7 +942,7 @@ declare function runGoldenEval(opts: GoldenEvalOptions): Promise<GoldenEvalRunRe
  * guard in `tests/version.test.ts` fails the suite (and therefore
  * `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
  */
-declare const LIBRARY_VERSION = "2.0.0-alpha.64";
+declare const LIBRARY_VERSION = "2.0.0-alpha.66";
 
 /**
  * Oracle contract — how an app tells the brain whether a response was good.
@@ -1808,7 +1827,7 @@ declare function getPerAxisMetrics(opts: GetPerAxisMetricsOpts): Promise<PerAxis
  * no execute() handler, no env-var convention). Narrowing here keeps the
  * reachability check structurally honest.
  */
-type SupportedProvider = 'anthropic' | 'google' | 'openai' | 'deepseek';
+type SupportedProvider = 'anthropic' | 'google' | 'openai' | 'deepseek' | 'zai' | 'moonshot';
 /**
  * Per-provider env var names kgauto recognizes. Order doesn't matter —
  * first-present wins. Multiple names per provider supported because Google

package/dist/index.d.ts

@@ -1,5 +1,5 @@
-import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-ClU56aBc.js';
-export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, E as EffortLevel, F as FallbackReason, r as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, s as MutationApplied, t as NormalizedTokens, u as OutcomeKind, v as PerAxisMetricsByModel, w as PromptSection, x as SectionKind, y as ShadowProbeConfig, T as ToolCall, z as ToolDefinition, D as captureGoldenIr, J as parseGoldenCaptureRate, K as resolveGoldenCaptureRate, L as shouldCaptureGolden } from './ir-ClU56aBc.js';
+import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-Qnw6L265.js';
+export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, E as EffortLevel, F as FallbackReason, r as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, s as MutationApplied, t as NormalizedTokens, u as OutcomeKind, v as PerAxisMetricsByModel, w as PromptSection, x as SectionKind, y as ShadowProbeConfig, T as ToolCall, z as ToolDefinition, D as captureGoldenIr, J as parseGoldenCaptureRate, K as resolveGoldenCaptureRate, L as shouldCaptureGolden } from './ir-Qnw6L265.js';
 import { ModelProfile, ArchetypeConvention } from './profiles.js';
 export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, allProfiles, getProfile, latencyTierOf, profilesByProvider, tryGetProfile } from './profiles.js';
 export { BrainForwardConfig, BrainForwardRoutes, createBrainForwardRoutes } from './brain-proxy.js';
@@ -78,6 +78,19 @@ interface ExecuteErr {
     errorCode: string;
     message: string;
     raw: unknown;
+    /**
+     * alpha.66 — real token usage when the failure is a SYNTHESIZED
+     * structured-output contract violation (the provider returned 200 with
+     * usage; the response just violated the contract). Absent on transport-
+     * level failures (4xx/5xx/network) where no usage exists. Lets the
+     * brain's failure rows carry true `tokens_in` for the cliff class —
+     * PB's 2026-05-18 `chain-walk-on-structured-output-parse-failed` ask 4
+     * (tokens_in=0 on a 72s parse-fail made diagnosis ambiguous).
+     */
+    tokens?: {
+        input: number;
+        output: number;
+    };
 }
 type ExecuteResult = ExecuteOk | ExecuteErr;
 declare function execute(request: CompiledRequest, opts?: ExecuteOptions): Promise<ExecuteResult>;
@@ -597,6 +610,12 @@ interface OutcomePayload {
     system_prompt_chars?: number;
     fell_over_from?: string;
     fallback_reason?: 'rate_limit' | 'provider_auth_failed' | 'provider_error' | 'cliff' | 'cost_cap' | 'contract_violation';
+    /**
+     * True when the same-model retry (CallOptions.sameModelRetry) fired
+     * during this call. Key OMITTED (not false) when no retry fired — safe
+     * against brains that haven't applied migration 039 yet.
+     */
+    retried_same_model?: boolean;
     /**
      * Per-call SectionRewrite[] captured at compile time. Omitted (sent as
      * undefined → stored NULL) when no rewrites fired. Powers cross-app
@@ -923,7 +942,7 @@ declare function runGoldenEval(opts: GoldenEvalOptions): Promise<GoldenEvalRunRe
  * guard in `tests/version.test.ts` fails the suite (and therefore
  * `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
  */
-declare const LIBRARY_VERSION = "2.0.0-alpha.64";
+declare const LIBRARY_VERSION = "2.0.0-alpha.66";
 
 /**
  * Oracle contract — how an app tells the brain whether a response was good.
@@ -1808,7 +1827,7 @@ declare function getPerAxisMetrics(opts: GetPerAxisMetricsOpts): Promise<PerAxis
  * no execute() handler, no env-var convention). Narrowing here keeps the
  * reachability check structurally honest.
  */
-type SupportedProvider = 'anthropic' | 'google' | 'openai' | 'deepseek';
+type SupportedProvider = 'anthropic' | 'google' | 'openai' | 'deepseek' | 'zai' | 'moonshot';
 /**
  * Per-provider env var names kgauto recognizes. Order doesn't matter —
  * first-present wins. Multiple names per provider supported because Google