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

package/dist/index.mjs

@@ -15,7 +15,7 @@ import {
 import {
   LIBRARY_VERSION,
   createKeyHealthRoute
-} from "./chunk-DGPYLFDG.mjs";
+} from "./chunk-CMB6PZWS.mjs";
 import {
   ABSOLUTE_FLOOR,
   ARCHETYPE_FLOOR_DEFAULT,
@@ -37,10 +37,11 @@ import {
   isBrainQueryActiveFor,
   isModelReachable,
   isProviderReachable,
+  isSameModelRetryEnabledFromEnv,
   loadChainsFromBrain,
   readBrainReadEnv,
   resolveProviderKey
-} from "./chunk-YZRPNSSQ.mjs";
+} from "./chunk-SBFSYCQG.mjs";
 import {
   ALIASES,
   LATENCY_TIER_MS,
@@ -51,7 +52,7 @@ import {
   latencyTierOf,
   profilesByProvider,
   tryGetProfile
-} from "./chunk-QKXTMVCT.mjs";
+} from "./chunk-4UO4CCSP.mjs";
 import {
   emitAdvisoryFired,
   emitCompileDone,
@@ -1102,6 +1103,10 @@ 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}`);
   }
@@ -1390,6 +1395,68 @@ 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;
@@ -3165,6 +3232,9 @@ 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.
@@ -3732,6 +3802,8 @@ 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) {
@@ -3743,6 +3815,18 @@ 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)}`);
@@ -3892,6 +3976,34 @@ 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];
@@ -4095,6 +4207,8 @@ 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);
@@ -4132,15 +4246,37 @@ 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, {
@@ -4177,7 +4313,10 @@ 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");
@@ -4235,6 +4374,7 @@ async function call(ir, opts = {}) {
         servedBy: targetModel,
         fellOverFrom: fellOver ? initial.target : void 0,
         fallbackReason,
+        retriedSameModel: retriedSameModel || void 0,
         unreachableFiltered,
         policyBlockedFiltered,
         traceId,
@@ -4249,7 +4389,8 @@ 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) {
@@ -4260,15 +4401,20 @@ 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: [${policyBlockedFiltered.join(", ")}])` : "";
@@ -4513,7 +4659,10 @@ function validateStructuredContract(exec, ir) {
       errorType: "retryable",
       errorCode: "max_tokens_on_structured_output",
       message: `Provider returned finishReason="${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) {
@@ -4529,10 +4678,35 @@ 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;
@@ -4565,6 +4739,8 @@ function safeEmit(fn) {
 var _internal = {
   compileAndRegister,
   validateStructuredContract,
+  isStructuredContractViolation,
+  raiseOutputBudget,
   extractPromptPreview,
   normalizeFallbackReason,
   generateTraceId,

package/dist/{ir-B2dRyJXO.d.mts → ir-BTvyl8-w.d.mts}

@@ -373,7 +373,7 @@ interface PromptIR {
      */
     historyCachePolicy?: HistoryCachePolicy;
 }
-type Provider = 'anthropic' | 'google' | 'openai' | 'deepseek' | 'mistral' | 'xai';
+type Provider = 'anthropic' | 'google' | 'openai' | 'deepseek' | 'zai' | 'moonshot' | 'mistral' | 'xai';
 /**
  * Mutation IDs that fired during compile. Empty in v1 (no mutation engine
  * yet). Populated when the brain is online and pushing mutations.
@@ -456,6 +456,33 @@ type CompiledRequest = {
         content: unknown;
     }>;
     tools?: unknown[];
+} | {
+    provider: 'zai';
+    model: string;
+    messages: Array<{
+        role: string;
+        content: unknown;
+    }>;
+    tools?: unknown[];
+    response_format?: unknown;
+    /**
+     * alpha.65 — Z.ai thinking knob (docs.z.ai chat-completion reference):
+     * `thinking.type` is 'enabled' (default) or 'disabled'. Emitted only
+     * when a cliff forces thinking off (`force_thinking_budget_zero`);
+     * omitted otherwise so the provider default applies.
+     */
+    thinking?: {
+        type: 'enabled' | 'disabled';
+    };
+} | {
+    provider: 'moonshot';
+    model: string;
+    messages: Array<{
+        role: string;
+        content: unknown;
+    }>;
+    tools?: unknown[];
+    response_format?: unknown;
 };
 /**
  * Best-practice advisory emitted by the compiler at compile time. Non-fatal —
@@ -866,6 +893,8 @@ interface ApiKeys {
     google?: string;
     openai?: string;
     deepseek?: string;
+    zai?: string;
+    moonshot?: string;
 }
 /**
  * Per-provider override fields shallow-merged into the lowered request before
@@ -877,6 +906,8 @@ interface ProviderOverrides {
     google?: Record<string, unknown>;
     openai?: Record<string, unknown>;
     deepseek?: Record<string, unknown>;
+    zai?: Record<string, unknown>;
+    moonshot?: Record<string, unknown>;
 }
 /**
  * Full-IR inline shadow-probe config (Shape B, Phase 1 — 2026-05-29 s51).
@@ -1034,12 +1065,40 @@ interface CallOptions {
      * capture never fires.
      */
     goldenCapture?: GoldenCaptureOptions;
+    /**
+     * alpha.66 — one same-model retry on a structured-output contract violation
+     * (`max_tokens_on_structured_output` / `structured_output_parse_failed`)
+     * BEFORE chain-walking to a lower-ranked tier. Filed by playbacksam
+     * 2026-07-18 (`structured-output-truncation-retry-same-model`): on PB's
+     * summarize surface the walk destination is measured quality-inferior
+     * (golden eval run #3), so one more attempt at the leader beats walking.
+     *
+     * Consent posture (matches KGAUTO_AUTO_PROMOTE / KGAUTO_GOLDEN_CAPTURE):
+     * default OFF; `true` wins, explicit `false` beats the env, undefined
+     * falls through to the `KGAUTO_SAME_MODEL_RETRY` env var ('1'/'true').
+     * Default-off because the retry adds one full model round-trip of latency
+     * in the failure class — the consumer decides whether their fallback tier
+     * is bad enough to pay that.
+     *
+     * Bounds: at most ONE retry per call(), on whichever target hits the
+     * contract violation first. Truncation-class retries additionally raise
+     * any explicit wire output cap sitting below the profile's
+     * `maxOutputTokens` (e.g. a terse-clamp or providerOverrides cap);
+     * requests already at the profile ceiling re-roll unchanged.
+     */
+    sameModelRetry?: boolean;
 }
 interface CallAttempt {
     model: string;
     status: 'success' | 'retryable' | 'terminal';
     errorCode?: string;
     message?: string;
+    /**
+     * alpha.66 — true on the attempt row that was the one-shot same-model
+     * retry (CallOptions.sameModelRetry). The preceding row for the same model
+     * carries the contract violation that triggered it.
+     */
+    sameModelRetry?: boolean;
 }
 /**
  * Why fallback fired. Normalized for `CallResult.fallbackReason` (alpha.9).
@@ -1095,6 +1154,14 @@ interface CallResult {
     fellOverFrom?: string;
     /** Set only when fallback fired. Normalized cause. */
     fallbackReason?: FallbackReason;
+    /**
+     * alpha.66 — true when the one-shot same-model retry
+     * (CallOptions.sameModelRetry) fired during this call, regardless of
+     * whether the retry rescued the call or the chain walked afterwards.
+     * Cross-check `attempts[]` (the `sameModelRetry: true` row) and
+     * `fellOverFrom` to distinguish rescued-by-retry from retried-then-walked.
+     */
+    retriedSameModel?: boolean;
     /**
      * alpha.10. Models that auto-filter dropped from the fallback walk because
      * their provider had no reachable API key. Empty when nothing was filtered
@@ -1292,6 +1359,13 @@ interface RecordInput {
      * keep in sync with the wire-contract enum (TraceDetail.fallbackReason).
      */
     fallbackReason?: 'rate_limit' | 'provider_auth_failed' | 'provider_error' | 'cliff' | 'cost_cap' | 'contract_violation';
+    /**
+     * alpha.66 — true when the one-shot same-model retry fired during this
+     * call (migration 039 `retried_same_model`). Powers the offline rollup:
+     * retry rate on the cliff class, and — joined with `fell_over_from` —
+     * how often the retry rescued the call vs the chain walking anyway.
+     */
+    retriedSameModel?: boolean;
 }
 /**
  * alpha.20 Entry 4: kinds of consumer-declared outcomes feeding the quality

package/dist/{ir-ClU56aBc.d.ts → ir-Qnw6L265.d.ts}

@@ -373,7 +373,7 @@ interface PromptIR {
      */
     historyCachePolicy?: HistoryCachePolicy;
 }
-type Provider = 'anthropic' | 'google' | 'openai' | 'deepseek' | 'mistral' | 'xai';
+type Provider = 'anthropic' | 'google' | 'openai' | 'deepseek' | 'zai' | 'moonshot' | 'mistral' | 'xai';
 /**
  * Mutation IDs that fired during compile. Empty in v1 (no mutation engine
  * yet). Populated when the brain is online and pushing mutations.
@@ -456,6 +456,33 @@ type CompiledRequest = {
         content: unknown;
     }>;
     tools?: unknown[];
+} | {
+    provider: 'zai';
+    model: string;
+    messages: Array<{
+        role: string;
+        content: unknown;
+    }>;
+    tools?: unknown[];
+    response_format?: unknown;
+    /**
+     * alpha.65 — Z.ai thinking knob (docs.z.ai chat-completion reference):
+     * `thinking.type` is 'enabled' (default) or 'disabled'. Emitted only
+     * when a cliff forces thinking off (`force_thinking_budget_zero`);
+     * omitted otherwise so the provider default applies.
+     */
+    thinking?: {
+        type: 'enabled' | 'disabled';
+    };
+} | {
+    provider: 'moonshot';
+    model: string;
+    messages: Array<{
+        role: string;
+        content: unknown;
+    }>;
+    tools?: unknown[];
+    response_format?: unknown;
 };
 /**
  * Best-practice advisory emitted by the compiler at compile time. Non-fatal —
@@ -866,6 +893,8 @@ interface ApiKeys {
     google?: string;
     openai?: string;
     deepseek?: string;
+    zai?: string;
+    moonshot?: string;
 }
 /**
  * Per-provider override fields shallow-merged into the lowered request before
@@ -877,6 +906,8 @@ interface ProviderOverrides {
     google?: Record<string, unknown>;
     openai?: Record<string, unknown>;
     deepseek?: Record<string, unknown>;
+    zai?: Record<string, unknown>;
+    moonshot?: Record<string, unknown>;
 }
 /**
  * Full-IR inline shadow-probe config (Shape B, Phase 1 — 2026-05-29 s51).
@@ -1034,12 +1065,40 @@ interface CallOptions {
      * capture never fires.
      */
     goldenCapture?: GoldenCaptureOptions;
+    /**
+     * alpha.66 — one same-model retry on a structured-output contract violation
+     * (`max_tokens_on_structured_output` / `structured_output_parse_failed`)
+     * BEFORE chain-walking to a lower-ranked tier. Filed by playbacksam
+     * 2026-07-18 (`structured-output-truncation-retry-same-model`): on PB's
+     * summarize surface the walk destination is measured quality-inferior
+     * (golden eval run #3), so one more attempt at the leader beats walking.
+     *
+     * Consent posture (matches KGAUTO_AUTO_PROMOTE / KGAUTO_GOLDEN_CAPTURE):
+     * default OFF; `true` wins, explicit `false` beats the env, undefined
+     * falls through to the `KGAUTO_SAME_MODEL_RETRY` env var ('1'/'true').
+     * Default-off because the retry adds one full model round-trip of latency
+     * in the failure class — the consumer decides whether their fallback tier
+     * is bad enough to pay that.
+     *
+     * Bounds: at most ONE retry per call(), on whichever target hits the
+     * contract violation first. Truncation-class retries additionally raise
+     * any explicit wire output cap sitting below the profile's
+     * `maxOutputTokens` (e.g. a terse-clamp or providerOverrides cap);
+     * requests already at the profile ceiling re-roll unchanged.
+     */
+    sameModelRetry?: boolean;
 }
 interface CallAttempt {
     model: string;
     status: 'success' | 'retryable' | 'terminal';
     errorCode?: string;
     message?: string;
+    /**
+     * alpha.66 — true on the attempt row that was the one-shot same-model
+     * retry (CallOptions.sameModelRetry). The preceding row for the same model
+     * carries the contract violation that triggered it.
+     */
+    sameModelRetry?: boolean;
 }
 /**
  * Why fallback fired. Normalized for `CallResult.fallbackReason` (alpha.9).
@@ -1095,6 +1154,14 @@ interface CallResult {
     fellOverFrom?: string;
     /** Set only when fallback fired. Normalized cause. */
     fallbackReason?: FallbackReason;
+    /**
+     * alpha.66 — true when the one-shot same-model retry
+     * (CallOptions.sameModelRetry) fired during this call, regardless of
+     * whether the retry rescued the call or the chain walked afterwards.
+     * Cross-check `attempts[]` (the `sameModelRetry: true` row) and
+     * `fellOverFrom` to distinguish rescued-by-retry from retried-then-walked.
+     */
+    retriedSameModel?: boolean;
     /**
      * alpha.10. Models that auto-filter dropped from the fallback walk because
      * their provider had no reachable API key. Empty when nothing was filtered
@@ -1292,6 +1359,13 @@ interface RecordInput {
      * keep in sync with the wire-contract enum (TraceDetail.fallbackReason).
      */
     fallbackReason?: 'rate_limit' | 'provider_auth_failed' | 'provider_error' | 'cliff' | 'cost_cap' | 'contract_violation';
+    /**
+     * alpha.66 — true when the one-shot same-model retry fired during this
+     * call (migration 039 `retried_same_model`). Powers the offline rollup:
+     * retry rate on the cliff class, and — joined with `fell_over_from` —
+     * how often the retry rescued the call vs the chain walking anyway.
+     */
+    retriedSameModel?: boolean;
 }
 /**
  * alpha.20 Entry 4: kinds of consumer-declared outcomes feeding the quality

package/dist/key-health.d.mts

@@ -73,7 +73,7 @@
  *     ]
  *   }
  */
-type KeyHealthProvider = 'anthropic' | 'deepseek' | 'google' | 'openai';
+type KeyHealthProvider = 'anthropic' | 'deepseek' | 'google' | 'moonshot' | 'openai';
 interface KeyHealthConfig {
     /** Consumer app id echoed as `app_id` in the response (e.g. 'playbacksam'). */
     appId: string;

package/dist/key-health.d.ts

@@ -73,7 +73,7 @@
  *     ]
  *   }
  */
-type KeyHealthProvider = 'anthropic' | 'deepseek' | 'google' | 'openai';
+type KeyHealthProvider = 'anthropic' | 'deepseek' | 'google' | 'moonshot' | 'openai';
 interface KeyHealthConfig {
     /** Consumer app id echoed as `app_id` in the response (e.g. 'playbacksam'). */
     appId: string;

package/dist/key-health.js

@@ -25,7 +25,7 @@ __export(key_health_exports, {
 module.exports = __toCommonJS(key_health_exports);
 
 // 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" };
@@ -74,6 +74,14 @@ 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",
@@ -105,6 +113,11 @@ 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/key-health.mjs

@@ -1,6 +1,6 @@
 import {
   createKeyHealthRoute
-} from "./chunk-DGPYLFDG.mjs";
+} from "./chunk-CMB6PZWS.mjs";
 export {
   createKeyHealthRoute
 };

package/dist/profiles.d.mts

@@ -1,4 +1,4 @@
-import { k as Provider } from './ir-B2dRyJXO.mjs';
+import { k as Provider } from './ir-BTvyl8-w.mjs';
 import { IntentArchetypeName } from './dialect.mjs';
 
 /**

package/dist/profiles.d.ts

@@ -1,4 +1,4 @@
-import { k as Provider } from './ir-ClU56aBc.js';
+import { k as Provider } from './ir-Qnw6L265.js';
 import { IntentArchetypeName } from './dialect.js';
 
 /**