@warmdrift/kgauto-compiler 2.0.0-alpha.3 → 2.0.0-alpha.4

package/README.md

@@ -1,4 +1,4 @@
-# @warmdrift/kgauto-compiler — v2.0.0-alpha.3
+# @warmdrift/kgauto-compiler — v2.0.0-alpha.4
 
 > Prompt compiler + central learning brain for multi-model AI apps.
 > **Swap models without rewriting prompts.**
@@ -18,7 +18,7 @@ mutations.
 - **Package:** alpha — coexists with v1 (`@warmdrift/kgauto@1.2.0`) under
   the temporary name `@warmdrift/kgauto-compiler`. Renames to v2 final once
   v1 is fully retired from production.
-- **Tests:** 132/132 passing
+- **Tests:** 147/147 passing
 - **Build:** clean (43KB ESM, 60KB CJS)
 - **Brain:** schema ready (see `brain/migrations/001_initial_schema.sql`);
   awaiting dedicated Supabase provisioning.

package/dist/index.d.mts

@@ -1,5 +1,5 @@
-import { M as ModelProfile, C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, R as RecordInput, O as OracleScore, e as CompileResult } from './profiles-BiyrF36f.mjs';
-export { f as ALIASES, g as CacheStrategy, h as CallAttempt, i as CallError, j as CliffRule, k as Constraints, I as IntentDeclaration, L as LoweringSpec, l as Message, m as MutationApplied, n as NormalizedTokens, o as PromptSection, p as Provider, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, T as ToolCall, s as ToolDefinition, t as allProfiles, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-BiyrF36f.mjs';
+import { M as ModelProfile, C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, R as RecordInput, O as OracleScore, e as CompileResult } from './profiles-CH_nKPjp.mjs';
+export { f as ALIASES, g as CacheStrategy, h as CallAttempt, i as CallError, j as CliffRule, k as Constraints, I as IntentDeclaration, L as LoweringSpec, l as Message, m as MutationApplied, n as NormalizedTokens, o as PromptSection, p as Provider, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, T as ToolCall, s as ToolDefinition, t as allProfiles, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-CH_nKPjp.mjs';
 export { ALL_ARCHETYPES, ContextBucket, DIALECT_VERSION, HistoryDepth, INTENT_ARCHETYPES, IntentArchetypeName, OutputMode, ShapeSignature, ToolCountBucket, bucketContext, bucketHistory, bucketToolCount, hashShape, isArchetype, learningKey } from './dialect.mjs';
 
 /**

package/dist/index.d.ts

@@ -1,5 +1,5 @@
-import { M as ModelProfile, C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, R as RecordInput, O as OracleScore, e as CompileResult } from './profiles-C5lVqF8_.js';
-export { f as ALIASES, g as CacheStrategy, h as CallAttempt, i as CallError, j as CliffRule, k as Constraints, I as IntentDeclaration, L as LoweringSpec, l as Message, m as MutationApplied, n as NormalizedTokens, o as PromptSection, p as Provider, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, T as ToolCall, s as ToolDefinition, t as allProfiles, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-C5lVqF8_.js';
+import { M as ModelProfile, C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, R as RecordInput, O as OracleScore, e as CompileResult } from './profiles-CDttLtaD.js';
+export { f as ALIASES, g as CacheStrategy, h as CallAttempt, i as CallError, j as CliffRule, k as Constraints, I as IntentDeclaration, L as LoweringSpec, l as Message, m as MutationApplied, n as NormalizedTokens, o as PromptSection, p as Provider, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, T as ToolCall, s as ToolDefinition, t as allProfiles, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-CDttLtaD.js';
 export { ALL_ARCHETYPES, ContextBucket, DIALECT_VERSION, HistoryDepth, INTENT_ARCHETYPES, IntentArchetypeName, OutputMode, ShapeSignature, ToolCountBucket, bucketContext, bucketHistory, bucketToolCount, hashShape, isArchetype, learningKey } from './dialect.js';
 
 /**

package/dist/index.js

@@ -1309,6 +1309,9 @@ function buildPayload(input, reg) {
   const compileTarget = reg?.model;
   const actual = input.actualModel ?? compileTarget;
   const requested = input.actualModel && compileTarget && input.actualModel !== compileTarget ? compileTarget : void 0;
+  const mutationsApplied = input.mutationsApplied ?? reg?.mutationsApplied ?? [];
+  const costModel = actual;
+  const costUsdActual = costModel ? computeCostUsd(costModel, input.tokensIn, input.tokensOut) : void 0;
   return {
     handle: input.handle,
     app_id: reg?.appId,
@@ -1318,7 +1321,7 @@ function buildPayload(input, reg) {
     provider: reg?.provider,
     shape_key: reg?.shapeKey,
     learning_key: reg?.learningKey,
-    mutations_applied: reg?.mutationsApplied ?? [],
+    mutations_applied: mutationsApplied,
     tokens_in: input.tokensIn,
     tokens_out: input.tokensOut,
     estimated_tokens_in: reg?.estimatedTokensIn,
@@ -1332,9 +1335,21 @@ function buildPayload(input, reg) {
     oracle_rationale: input.oracleScore?.rationale,
     prompt_preview: input.promptPreview,
     response_preview: input.responsePreview,
-    dialect_version: "v1"
+    dialect_version: "v1",
+    cache_read_input_tokens: input.cacheReadInputTokens,
+    cache_creation_input_tokens: input.cacheCreationInputTokens,
+    cost_usd_actual: costUsdActual,
+    ttft_ms: input.ttftMs
   };
 }
+function computeCostUsd(modelId, tokensIn, tokensOut) {
+  if (tokensIn === 0 && tokensOut === 0) return void 0;
+  const profile = tryGetProfile(modelId);
+  if (!profile) return void 0;
+  const inUsd = tokensIn / 1e6 * profile.costInputPer1m;
+  const outUsd = tokensOut / 1e6 * profile.costOutputPer1m;
+  return Math.round((inUsd + outUsd) * 1e6) / 1e6;
+}
 
 // src/ir.ts
 var CallError = class extends Error {
@@ -1607,7 +1622,7 @@ async function call(ir, opts = {}) {
       attempts.push({ model: targetModel, status: "success" });
       const latencyMs2 = Date.now() - start;
       const responseWithStructured = withStructuredOutput(exec.response, ir);
-      void record({
+      await record({
         handle: initial.handle,
         tokensIn: responseWithStructured.tokens.input,
         tokensOut: responseWithStructured.tokens.output,
@@ -1616,7 +1631,11 @@ async function call(ir, opts = {}) {
         emptyResponse: responseWithStructured.tokens.output === 0,
         toolsCalled: responseWithStructured.toolCalls.map((tc) => tc.name),
         actualModel: targetModel !== initial.target ? targetModel : void 0,
-        responsePreview: responseWithStructured.text.slice(0, 200)
+        mutationsApplied: targetModel !== initial.target ? activeCompile.mutationsApplied.map((m) => m.id) : void 0,
+        promptPreview: extractPromptPreview(ir),
+        responsePreview: responseWithStructured.text.slice(0, 200),
+        cacheReadInputTokens: responseWithStructured.tokens.cached,
+        cacheCreationInputTokens: responseWithStructured.tokens.cacheCreated
       });
       return {
         handle: initial.handle,
@@ -1641,13 +1660,14 @@ async function call(ir, opts = {}) {
     }
   }
   const latencyMs = Date.now() - start;
-  void record({
+  await record({
     handle: initial.handle,
     tokensIn: 0,
     tokensOut: 0,
     latencyMs,
     success: false,
-    errorType: lastErr?.errorCode
+    errorType: lastErr?.errorCode,
+    promptPreview: extractPromptPreview(ir)
   });
   throw new CallError(
     `call(): all attempts failed${lastErr ? ` \u2014 ${lastErr.errorCode}: ${lastErr.message}` : ""}`,
@@ -1665,6 +1685,13 @@ function compileAndRegister(ir, opts) {
   registerCompile(ir.appId, ir.intent.archetype, ir, result);
   return result;
 }
+function extractPromptPreview(ir) {
+  const turn = ir.currentTurn?.content;
+  if (turn) return turn.slice(0, 200);
+  const lastHist = ir.history?.[ir.history.length - 1]?.content;
+  if (lastHist) return lastHist.slice(0, 200);
+  return void 0;
+}
 function withStructuredOutput(response, ir) {
   if (!ir.constraints?.structuredOutput) return response;
   if (!response.text) return response;

package/dist/index.mjs

@@ -792,6 +792,9 @@ function buildPayload(input, reg) {
   const compileTarget = reg?.model;
   const actual = input.actualModel ?? compileTarget;
   const requested = input.actualModel && compileTarget && input.actualModel !== compileTarget ? compileTarget : void 0;
+  const mutationsApplied = input.mutationsApplied ?? reg?.mutationsApplied ?? [];
+  const costModel = actual;
+  const costUsdActual = costModel ? computeCostUsd(costModel, input.tokensIn, input.tokensOut) : void 0;
   return {
     handle: input.handle,
     app_id: reg?.appId,
@@ -801,7 +804,7 @@ function buildPayload(input, reg) {
     provider: reg?.provider,
     shape_key: reg?.shapeKey,
     learning_key: reg?.learningKey,
-    mutations_applied: reg?.mutationsApplied ?? [],
+    mutations_applied: mutationsApplied,
     tokens_in: input.tokensIn,
     tokens_out: input.tokensOut,
     estimated_tokens_in: reg?.estimatedTokensIn,
@@ -815,9 +818,21 @@ function buildPayload(input, reg) {
     oracle_rationale: input.oracleScore?.rationale,
     prompt_preview: input.promptPreview,
     response_preview: input.responsePreview,
-    dialect_version: "v1"
+    dialect_version: "v1",
+    cache_read_input_tokens: input.cacheReadInputTokens,
+    cache_creation_input_tokens: input.cacheCreationInputTokens,
+    cost_usd_actual: costUsdActual,
+    ttft_ms: input.ttftMs
   };
 }
+function computeCostUsd(modelId, tokensIn, tokensOut) {
+  if (tokensIn === 0 && tokensOut === 0) return void 0;
+  const profile = tryGetProfile(modelId);
+  if (!profile) return void 0;
+  const inUsd = tokensIn / 1e6 * profile.costInputPer1m;
+  const outUsd = tokensOut / 1e6 * profile.costOutputPer1m;
+  return Math.round((inUsd + outUsd) * 1e6) / 1e6;
+}
 
 // src/ir.ts
 var CallError = class extends Error {
@@ -1090,7 +1105,7 @@ async function call(ir, opts = {}) {
       attempts.push({ model: targetModel, status: "success" });
       const latencyMs2 = Date.now() - start;
       const responseWithStructured = withStructuredOutput(exec.response, ir);
-      void record({
+      await record({
         handle: initial.handle,
         tokensIn: responseWithStructured.tokens.input,
         tokensOut: responseWithStructured.tokens.output,
@@ -1099,7 +1114,11 @@ async function call(ir, opts = {}) {
         emptyResponse: responseWithStructured.tokens.output === 0,
         toolsCalled: responseWithStructured.toolCalls.map((tc) => tc.name),
         actualModel: targetModel !== initial.target ? targetModel : void 0,
-        responsePreview: responseWithStructured.text.slice(0, 200)
+        mutationsApplied: targetModel !== initial.target ? activeCompile.mutationsApplied.map((m) => m.id) : void 0,
+        promptPreview: extractPromptPreview(ir),
+        responsePreview: responseWithStructured.text.slice(0, 200),
+        cacheReadInputTokens: responseWithStructured.tokens.cached,
+        cacheCreationInputTokens: responseWithStructured.tokens.cacheCreated
       });
       return {
         handle: initial.handle,
@@ -1124,13 +1143,14 @@ async function call(ir, opts = {}) {
     }
   }
   const latencyMs = Date.now() - start;
-  void record({
+  await record({
     handle: initial.handle,
     tokensIn: 0,
     tokensOut: 0,
     latencyMs,
     success: false,
-    errorType: lastErr?.errorCode
+    errorType: lastErr?.errorCode,
+    promptPreview: extractPromptPreview(ir)
   });
   throw new CallError(
     `call(): all attempts failed${lastErr ? ` \u2014 ${lastErr.errorCode}: ${lastErr.message}` : ""}`,
@@ -1148,6 +1168,13 @@ function compileAndRegister(ir, opts) {
   registerCompile(ir.appId, ir.intent.archetype, ir, result);
   return result;
 }
+function extractPromptPreview(ir) {
+  const turn = ir.currentTurn?.content;
+  if (turn) return turn.slice(0, 200);
+  const lastHist = ir.history?.[ir.history.length - 1]?.content;
+  if (lastHist) return lastHist.slice(0, 200);
+  return void 0;
+}
 function withStructuredOutput(response, ir) {
   if (!ir.constraints?.structuredOutput) return response;
   if (!response.text) return response;

package/dist/{profiles-C5lVqF8_.d.ts → profiles-CDttLtaD.d.ts}

@@ -386,6 +386,38 @@ interface RecordInput {
      * the originally-requested model.
      */
     actualModel?: string;
+    /**
+     * Override `mutations_applied` for this outcome. Set by `call()` when
+     * fallback fires — the served compile's mutations (which actually shaped
+     * the request that went on the wire) replace the initial compile's
+     * mutations (registered against the handle). Without this override, fallback
+     * traffic is attributed to the initial compile's mutations and the brain's
+     * mutation effectiveness stats become misleading.
+     *
+     * alpha.4: extends s11 truth-in-logging to mutations.
+     */
+    mutationsApplied?: string[];
+    /**
+     * Cache read input tokens, when supported by the provider.
+     * - Anthropic: `usage.cache_read_input_tokens`
+     * - Google (implicit caching): `usageMetadata.cachedContentTokenCount`
+     * - OpenAI: `usage.prompt_tokens_details.cached_tokens`
+     *
+     * Powers the cost-and-efficiency-watcher (interfaces/kgauto.md, alpha.4):
+     * `tokens_in - cache_read_input_tokens` is the un-cached new context per call.
+     */
+    cacheReadInputTokens?: number;
+    /**
+     * Cache creation input tokens (Anthropic-specific).
+     * `usage.cache_creation_input_tokens`. The first call that pays the 25%
+     * upcharge to write a cache marker; subsequent calls hit `cacheRead`.
+     */
+    cacheCreationInputTokens?: number;
+    /**
+     * Time to first token (ms). Optional; populated when the provider/SDK
+     * surfaces it. Distinct from `latencyMs` (end-to-end wall clock).
+     */
+    ttftMs?: number;
 }
 
 /**

package/dist/{profiles-BiyrF36f.d.mts → profiles-CH_nKPjp.d.mts}

@@ -386,6 +386,38 @@ interface RecordInput {
      * the originally-requested model.
      */
     actualModel?: string;
+    /**
+     * Override `mutations_applied` for this outcome. Set by `call()` when
+     * fallback fires — the served compile's mutations (which actually shaped
+     * the request that went on the wire) replace the initial compile's
+     * mutations (registered against the handle). Without this override, fallback
+     * traffic is attributed to the initial compile's mutations and the brain's
+     * mutation effectiveness stats become misleading.
+     *
+     * alpha.4: extends s11 truth-in-logging to mutations.
+     */
+    mutationsApplied?: string[];
+    /**
+     * Cache read input tokens, when supported by the provider.
+     * - Anthropic: `usage.cache_read_input_tokens`
+     * - Google (implicit caching): `usageMetadata.cachedContentTokenCount`
+     * - OpenAI: `usage.prompt_tokens_details.cached_tokens`
+     *
+     * Powers the cost-and-efficiency-watcher (interfaces/kgauto.md, alpha.4):
+     * `tokens_in - cache_read_input_tokens` is the un-cached new context per call.
+     */
+    cacheReadInputTokens?: number;
+    /**
+     * Cache creation input tokens (Anthropic-specific).
+     * `usage.cache_creation_input_tokens`. The first call that pays the 25%
+     * upcharge to write a cache marker; subsequent calls hit `cacheRead`.
+     */
+    cacheCreationInputTokens?: number;
+    /**
+     * Time to first token (ms). Optional; populated when the provider/SDK
+     * surfaces it. Distinct from `latencyMs` (end-to-end wall clock).
+     */
+    ttftMs?: number;
 }
 
 /**

package/dist/profiles.d.mts

@@ -1,2 +1,2 @@
-export { f as ALIASES, g as CacheStrategy, j as CliffRule, L as LoweringSpec, M as ModelProfile, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, t as allProfiles, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-BiyrF36f.mjs';
+export { f as ALIASES, g as CacheStrategy, j as CliffRule, L as LoweringSpec, M as ModelProfile, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, t as allProfiles, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-CH_nKPjp.mjs';
 import './dialect.mjs';

package/dist/profiles.d.ts

@@ -1,2 +1,2 @@
-export { f as ALIASES, g as CacheStrategy, j as CliffRule, L as LoweringSpec, M as ModelProfile, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, t as allProfiles, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-C5lVqF8_.js';
+export { f as ALIASES, g as CacheStrategy, j as CliffRule, L as LoweringSpec, M as ModelProfile, q as RecoveryRule, S as StructuredOutputCapability, r as SystemPromptMode, t as allProfiles, u as getProfile, v as profilesByProvider, w as tryGetProfile } from './profiles-CDttLtaD.js';
 import './dialect.js';

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@warmdrift/kgauto-compiler",
-  "version": "2.0.0-alpha.3",
+  "version": "2.0.0-alpha.4",
   "description": "Prompt compiler + central learning brain for multi-model AI apps. Swap models without rewriting prompts.",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",