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

package/dist/index.js

@@ -1718,6 +1718,156 @@ 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 = {
@@ -2972,6 +3122,10 @@ function lower(ir, profile, hints = {}) {
       return lowerOpenAI(ir, profile, hints);
     case "deepseek":
       return lowerDeepSeek(ir, profile);
+    case "zai":
+      return lowerZai(ir, profile, hints);
+    case "moonshot":
+      return lowerMoonshot(ir, profile);
     default:
       throw new Error(`No lowering implementation for provider: ${profile.provider}`);
   }
@@ -3260,6 +3414,68 @@ function lowerDeepSeek(ir, profile) {
     }
   };
 }
+function buildOpenAICompatibleParts(ir) {
+  const ordered = sortSections(ir.sections);
+  const systemText = ordered.map((s) => s.text).join("\n\n");
+  const messages = systemText ? [{ role: "system", content: systemText }] : [];
+  for (const m of ir.history ?? []) {
+    if (m.role === "system") continue;
+    messages.push({ role: m.role, content: m.parts ?? m.content });
+  }
+  if (ir.currentTurn && ir.currentTurn.role !== "system") {
+    messages.push({
+      role: ir.currentTurn.role,
+      content: ir.currentTurn.parts ?? ir.currentTurn.content
+    });
+  }
+  const history = (ir.history ?? []).filter((m) => m.role !== "system");
+  const histMarkIndex = resolveHistoryMarkIndex(history.length, ir.historyCachePolicy);
+  const historyCacheableTokens = histMarkIndex >= 0 ? sumHistoryTokens(history, histMarkIndex) : 0;
+  return {
+    messages,
+    tools: ir.tools && ir.tools.length > 0 ? toOpenAITools(ir.tools) : void 0,
+    response_format: ir.constraints?.structuredOutput ? { type: "json_object" } : void 0,
+    historyCacheableTokens
+  };
+}
+function lowerZai(ir, profile, hints) {
+  const parts = buildOpenAICompatibleParts(ir);
+  return {
+    request: {
+      provider: "zai",
+      model: profile.id,
+      messages: parts.messages,
+      tools: parts.tools,
+      response_format: parts.response_format,
+      // Z.ai thinking defaults to enabled server-side; emit an explicit
+      // disable only when a cliff forced thinking off (the GLM analogue of
+      // Gemini's thinkingBudget=0).
+      thinking: hints.forceThinkingZero ? { type: "disabled" } : void 0
+    },
+    diagnostics: {
+      cacheableTokens: 0,
+      historyCacheableTokens: parts.historyCacheableTokens,
+      estimatedCacheSavingsUsd: 0
+    }
+  };
+}
+function lowerMoonshot(ir, profile) {
+  const parts = buildOpenAICompatibleParts(ir);
+  return {
+    request: {
+      provider: "moonshot",
+      model: profile.id,
+      messages: parts.messages,
+      tools: parts.tools,
+      response_format: parts.response_format
+    },
+    diagnostics: {
+      cacheableTokens: 0,
+      historyCacheableTokens: parts.historyCacheableTokens,
+      estimatedCacheSavingsUsd: 0
+    }
+  };
+}
 function sortSections(sections) {
   return [...sections].sort((a, b) => {
     const wa = a.weight ?? 100;
@@ -5035,6 +5251,9 @@ function buildPayload(input, reg) {
     system_prompt_chars: input.systemPromptChars ?? reg?.systemPromptChars,
     fell_over_from: fellOverFrom,
     fallback_reason: fallbackReason,
+    // alpha.66 — omitted-undefined keeps the key out of the JSON body
+    // entirely for non-retry rows (pre-migration-039 brain safety).
+    retried_same_model: input.retriedSameModel,
     // alpha.29 — translator activity (migration 019). Send NULL when no
     // rewrites fired so the brain's "did the translator do anything?"
     // queries can use `IS NOT NULL` cleanly.
@@ -5316,7 +5535,9 @@ var SUPPORTED_PROVIDERS = Object.freeze([
   "anthropic",
   "google",
   "openai",
-  "deepseek"
+  "deepseek",
+  "zai",
+  "moonshot"
 ]);
 function isSupportedProvider(p) {
   return SUPPORTED_PROVIDERS.includes(p);
@@ -5329,7 +5550,15 @@ 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 : {};
@@ -5381,6 +5610,10 @@ 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/streaming.ts
 var ANTHROPIC_URL = "https://api.anthropic.com/v1/messages";
@@ -5673,6 +5906,8 @@ function retryableError(status, code, message, raw) {
 var ANTHROPIC_URL2 = "https://api.anthropic.com/v1/messages";
 var OPENAI_URL = "https://api.openai.com/v1/chat/completions";
 var DEEPSEEK_URL = "https://api.deepseek.com/chat/completions";
+var ZAI_URL = "https://api.z.ai/api/paas/v4/chat/completions";
+var MOONSHOT_URL = "https://api.moonshot.ai/v1/chat/completions";
 async function execute(request, opts = {}) {
   const merged = applyOverrides(request, opts.providerOverrides);
   switch (merged.provider) {
@@ -5684,6 +5919,18 @@ async function execute(request, opts = {}) {
       return executeOpenAI(merged, opts);
     case "deepseek":
       return executeDeepSeek(merged, opts);
+    case "zai":
+      return executeOpenAICompatible(
+        merged,
+        opts,
+        { provider: "zai", url: ZAI_URL, missingKeyMessage: "ZAI_API_KEY missing" }
+      );
+    case "moonshot":
+      return executeOpenAICompatible(
+        merged,
+        opts,
+        { provider: "moonshot", url: MOONSHOT_URL, missingKeyMessage: "MOONSHOT_API_KEY missing" }
+      );
     default: {
       const _exhaustive = merged;
       throw new Error(`execute(): no executor for provider: ${JSON.stringify(_exhaustive)}`);
@@ -5833,6 +6080,34 @@ async function executeDeepSeek(request, opts) {
   if (!res.ok) return classifyHttpError2(res.status, json);
   return { ok: true, status: res.status, response: normalizeOpenAILike(json) };
 }
+async function executeOpenAICompatible(request, opts, spec) {
+  const apiKey = resolveProviderKey(spec.provider, { apiKeys: opts.apiKeys });
+  if (!apiKey) {
+    return terminalError(401, "auth", spec.missingKeyMessage);
+  }
+  if (opts.onChunk) {
+    return streamOpenAILike(spec.url, request, apiKey, spec.provider, {
+      onChunk: opts.onChunk,
+      fetchImpl: opts.fetchImpl
+    });
+  }
+  const { provider: _provider, ...body } = request;
+  const fetchFn = opts.fetchImpl ?? fetch;
+  let res;
+  let json;
+  try {
+    res = await fetchFn(spec.url, {
+      method: "POST",
+      headers: { authorization: `Bearer ${apiKey}`, "content-type": "application/json" },
+      body: JSON.stringify(body)
+    });
+    json = await res.json().catch(() => ({}));
+  } catch (err) {
+    return retryableError2(0, "network_error", String(err), null);
+  }
+  if (!res.ok) return classifyHttpError2(res.status, json);
+  return { ok: true, status: res.status, response: normalizeOpenAILike(json) };
+}
 function normalizeOpenAILike(raw) {
   const r = raw;
   const choice = r.choices?.[0];
@@ -6690,6 +6965,8 @@ async function call(ir, opts = {}) {
   let activeCompile = initial;
   let lastErr;
   const failedProviders = /* @__PURE__ */ new Set();
+  const sameModelRetryEnabled = opts.sameModelRetry ?? isSameModelRetryEnabledFromEnv();
+  let retriedSameModel = false;
   for (let i = 0; i < targetsToTry.length; i++) {
     const targetModel = targetsToTry[i];
     const targetProfile = tryGetProfile(targetModel);
@@ -6727,15 +7004,37 @@ async function call(ir, opts = {}) {
     );
     const targetSupportsStreaming = targetProfile?.streaming === true;
     const streamingOnChunk = opts.onChunk && !opts.noStream && targetSupportsStreaming ? opts.onChunk : void 0;
-    const exec = await execute(activeCompile.request, {
+    const execOpts = {
       apiKeys: opts.apiKeys,
       fetchImpl: opts.fetchImpl,
       providerOverrides: opts.providerOverrides,
       onChunk: streamingOnChunk
-    });
-    const validated = exec.ok ? validateStructuredContract(exec, ir) : exec;
+    };
+    const exec = await execute(activeCompile.request, execOpts);
+    let validated = exec.ok ? validateStructuredContract(exec, ir) : exec;
+    let servedByRetry = false;
+    if (!validated.ok && isStructuredContractViolation(validated.errorCode) && sameModelRetryEnabled && !retriedSameModel) {
+      retriedSameModel = true;
+      attempts.push({
+        model: targetModel,
+        status: validated.errorType,
+        errorCode: validated.errorCode,
+        message: validated.message
+      });
+      const retryRequest = validated.errorCode === "max_tokens_on_structured_output" ? raiseOutputBudget(activeCompile.request, targetProfile) : activeCompile.request;
+      safeEmit(
+        () => emitExecuteAttempt(traceId, ir.appId, { model: targetModel, attemptIndex: i })
+      );
+      const retryExec = await execute(retryRequest, execOpts);
+      validated = retryExec.ok ? validateStructuredContract(retryExec, ir) : retryExec;
+      servedByRetry = true;
+    }
     if (validated.ok) {
-      attempts.push({ model: targetModel, status: "success" });
+      attempts.push({
+        model: targetModel,
+        status: "success",
+        ...servedByRetry ? { sameModelRetry: true } : {}
+      });
       const latencyMs2 = Date.now() - start;
       safeEmit(
         () => emitExecuteSuccess(traceId, ir.appId, {
@@ -6772,7 +7071,10 @@ async function call(ir, opts = {}) {
         finishReason: validated.response.finishReason,
         totalMs: latencyMs2,
         fellOverFrom: fellOver ? initial.target : void 0,
-        fallbackReason
+        fallbackReason,
+        // alpha.66 — migration 039. Omitted (not false) when no retry fired
+        // so pre-039 brains never see the key.
+        retriedSameModel: retriedSameModel || void 0
       });
       if (fellOver) {
         const firstFailed = attempts.find((a) => a.status !== "success");
@@ -6830,6 +7132,7 @@ async function call(ir, opts = {}) {
         servedBy: targetModel,
         fellOverFrom: fellOver ? initial.target : void 0,
         fallbackReason,
+        retriedSameModel: retriedSameModel || void 0,
         unreachableFiltered,
         policyBlockedFiltered,
         traceId,
@@ -6844,7 +7147,8 @@ async function call(ir, opts = {}) {
       model: targetModel,
       status: validated.errorType,
       errorCode: validated.errorCode,
-      message: validated.message
+      message: validated.message,
+      ...servedByRetry ? { sameModelRetry: true } : {}
     });
     lastErr = validated;
     if (validated.errorType === "terminal" || opts.noFallback) {
@@ -6855,15 +7159,20 @@ async function call(ir, opts = {}) {
       break;
     }
   }
+  const lastAttempted = [...attempts].reverse().find(
+    (a) => a.status !== "success" && a.errorCode !== "auth_inferred" && a.errorCode !== "compile_error"
+  );
   const latencyMs = Date.now() - start;
   await record({
     handle: initial.handle,
-    tokensIn: 0,
-    tokensOut: 0,
+    tokensIn: lastErr?.tokens?.input ?? 0,
+    tokensOut: lastErr?.tokens?.output ?? 0,
     latencyMs,
     success: false,
     errorType: lastErr?.errorCode,
-    promptPreview: extractPromptPreview(ir)
+    promptPreview: extractPromptPreview(ir),
+    actualModel: lastAttempted?.model,
+    retriedSameModel: retriedSameModel || void 0
   });
   const filteredNote = unreachableFiltered && unreachableFiltered.length > 0 ? ` (also auto-filtered: [${unreachableFiltered.join(", ")}] \u2014 no API key)` : "";
   const blockedNote = policyBlockedFiltered && policyBlockedFiltered.length > 0 ? ` (also policy-blocked: [${policyBlockedFiltered.join(", ")}])` : "";
@@ -7108,7 +7417,10 @@ function validateStructuredContract(exec, ir) {
       errorType: "retryable",
       errorCode: "max_tokens_on_structured_output",
       message: `Provider returned finishReason="${exec.response.finishReason}" on a structured-output call \u2014 output truncated mid-token, JSON cannot be valid`,
-      raw: exec.response.raw
+      raw: exec.response.raw,
+      // alpha.66 — the provider DID serve this call; carry the real usage so
+      // exhausted-chain failure rows aren't stamped tokens_in=0 (PB 05-18 ask 4).
+      tokens: { input: exec.response.tokens.input, output: exec.response.tokens.output }
     };
   }
   if (!exec.response.text) {
@@ -7124,10 +7436,35 @@ function validateStructuredContract(exec, ir) {
       errorType: "retryable",
       errorCode: "structured_output_parse_failed",
       message: err instanceof Error ? err.message : String(err),
-      raw: exec.response.raw
+      raw: exec.response.raw,
+      tokens: { input: exec.response.tokens.input, output: exec.response.tokens.output }
     };
   }
 }
+function isStructuredContractViolation(code) {
+  return code === "max_tokens_on_structured_output" || code === "structured_output_parse_failed";
+}
+function raiseOutputBudget(request, profile) {
+  if (!profile) return request;
+  if (request.provider === "anthropic") {
+    if (typeof request.max_tokens === "number" && request.max_tokens < profile.maxOutputTokens) {
+      return { ...request, max_tokens: profile.maxOutputTokens };
+    }
+    return request;
+  }
+  if (request.provider === "google") {
+    const gc = request.generationConfig;
+    const cap = gc?.maxOutputTokens;
+    if (typeof cap === "number" && cap < profile.maxOutputTokens) {
+      return {
+        ...request,
+        generationConfig: { ...gc, maxOutputTokens: profile.maxOutputTokens }
+      };
+    }
+    return request;
+  }
+  return request;
+}
 function normalizeFallbackReason(attempts) {
   const first = attempts.find((a) => a.status !== "success");
   if (!first) return void 0;
@@ -7160,6 +7497,8 @@ function safeEmit(fn) {
 var _internal = {
   compileAndRegister,
   validateStructuredContract,
+  isStructuredContractViolation,
+  raiseOutputBudget,
   extractPromptPreview,
   normalizeFallbackReason,
   generateTraceId,
@@ -7948,7 +8287,7 @@ function createBrainForwardRoutes(config) {
 }
 
 // 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_HEADERS2 = { "Content-Type": "application/json" };
@@ -7997,6 +8336,14 @@ 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",
@@ -8028,6 +8375,11 @@ 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);
     }