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

package/dist/profiles.js

@@ -1445,6 +1445,156 @@ 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/profiles.mjs

@@ -8,7 +8,7 @@ import {
   latencyTierOf,
   profilesByProvider,
   tryGetProfile
-} from "./chunk-QKXTMVCT.mjs";
+} from "./chunk-4UO4CCSP.mjs";
 export {
   ALIASES,
   LATENCY_TIER_MS,

package/dist/{types-axAX1Bg0.d.mts → types-8pWRDbUg.d.mts}

@@ -1,4 +1,4 @@
-import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-B2dRyJXO.mjs';
+import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-BTvyl8-w.mjs';
 
 /**
  * Glass-Box observability types (alpha.17).

package/dist/{types-C-Wp7HpI.d.ts → types-BdYlJz0Z.d.ts}

@@ -1,4 +1,4 @@
-import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-ClU56aBc.js';
+import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-Qnw6L265.js';
 
 /**
  * Glass-Box observability types (alpha.17).

package/dist/{types-CSDweZsl.d.mts → types-Cya0jPAF.d.mts}

@@ -1,4 +1,4 @@
-import { i as Adapter, x as SectionKind } from './ir-B2dRyJXO.mjs';
+import { i as Adapter, x as SectionKind } from './ir-BTvyl8-w.mjs';
 
 /**
  * Internal config + hook types for createGlassboxRoutes().

package/dist/{types-ByN3r0_7.d.ts → types-DxYXuFxj.d.ts}

@@ -1,4 +1,4 @@
-import { i as Adapter, x as SectionKind } from './ir-ClU56aBc.js';
+import { i as Adapter, x as SectionKind } from './ir-Qnw6L265.js';
 
 /**
  * Internal config + hook types for createGlassboxRoutes().

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@warmdrift/kgauto-compiler",
-  "version": "2.0.0-alpha.64",
+  "version": "2.0.0-alpha.66",
   "description": "Prompt compiler with executable provider knowledge for multi-model AI apps: normalized multi-provider transport with fallback chains, compile-time cliff guards, a curated model registry, and a telemetry flight recorder. Swap models without rewriting prompts.",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",