@warmdrift/kgauto-compiler 2.0.0-alpha.45 → 2.0.0-alpha.46

package/dist/index.d.mts

@@ -363,6 +363,15 @@ interface ShadowProbeRecordInput {
     tokensCurrentOut?: number;
     tokensCandidateIn?: number;
     tokensCandidateOut?: number;
+    /**
+     * Wall-clock latency of the served call, ms (alpha.46). The third swap-decision
+     * axis alongside quality + cost — a verdict needs to know the candidate is
+     * slower, not just better/cheaper. Served latency mirrors
+     * `compile_outcomes.latency_ms`; carried here so a probe row is self-contained.
+     */
+    latencyCurrentMs?: number;
+    /** Wall-clock latency of the candidate replay call, ms (alpha.46). */
+    latencyCandidateMs?: number;
 }
 /**
  * Pure builder for the `probe_outcomes` row written by the inline shadow-probe.
@@ -390,6 +399,8 @@ declare function buildShadowProbeRow(input: ShadowProbeRecordInput): {
     tokens_current_out: number | null;
     tokens_candidate_in: number | null;
     tokens_candidate_out: number | null;
+    latency_current_ms: number | null;
+    latency_candidate_ms: number | null;
     prompt_fidelity: number;
     replay_source: 'inline-full-ir';
 };

package/dist/index.d.ts

@@ -363,6 +363,15 @@ interface ShadowProbeRecordInput {
     tokensCurrentOut?: number;
     tokensCandidateIn?: number;
     tokensCandidateOut?: number;
+    /**
+     * Wall-clock latency of the served call, ms (alpha.46). The third swap-decision
+     * axis alongside quality + cost — a verdict needs to know the candidate is
+     * slower, not just better/cheaper. Served latency mirrors
+     * `compile_outcomes.latency_ms`; carried here so a probe row is self-contained.
+     */
+    latencyCurrentMs?: number;
+    /** Wall-clock latency of the candidate replay call, ms (alpha.46). */
+    latencyCandidateMs?: number;
 }
 /**
  * Pure builder for the `probe_outcomes` row written by the inline shadow-probe.
@@ -390,6 +399,8 @@ declare function buildShadowProbeRow(input: ShadowProbeRecordInput): {
     tokens_current_out: number | null;
     tokens_candidate_in: number | null;
     tokens_candidate_out: number | null;
+    latency_current_ms: number | null;
+    latency_candidate_ms: number | null;
     prompt_fidelity: number;
     replay_source: 'inline-full-ir';
 };

package/dist/index.js

@@ -4494,6 +4494,8 @@ function buildShadowProbeRow(input) {
     tokens_current_out: input.tokensCurrentOut ?? null,
     tokens_candidate_in: input.tokensCandidateIn ?? null,
     tokens_candidate_out: input.tokensCandidateOut ?? null,
+    latency_current_ms: input.latencyCurrentMs ?? null,
+    latency_candidate_ms: input.latencyCandidateMs ?? null,
     // Full IR was replayed (not a truncated preview), so fidelity is 1.0 — the
     // prompt-fidelity guard never fires on these rows.
     prompt_fidelity: 1,
@@ -6014,7 +6016,8 @@ async function call(ir, opts = {}) {
           ir,
           opts,
           servedModel: targetModel,
-          servedResponse: validated.response
+          servedResponse: validated.response,
+          servedLatencyMs: latencyMs2
         });
         if (isBrainSync()) {
           await probe;
@@ -6127,11 +6130,13 @@ async function runShadowProbe(args) {
           },
           args.opts
         );
+        const candStart = Date.now();
         const exec = await execute(candCompile.request, {
           apiKeys: args.opts.apiKeys,
           fetchImpl: args.opts.fetchImpl,
           providerOverrides: args.opts.providerOverrides
         });
+        const candidateLatencyMs = Date.now() - candStart;
         if (!exec.ok) continue;
         await recordShadowProbe({
           appId: args.ir.appId,
@@ -6145,7 +6150,12 @@ async function runShadowProbe(args) {
           tokensCurrentIn: args.servedResponse.tokens.input,
           tokensCurrentOut: args.servedResponse.tokens.output,
           tokensCandidateIn: exec.response.tokens.input,
-          tokensCandidateOut: exec.response.tokens.output
+          tokensCandidateOut: exec.response.tokens.output,
+          // alpha.46 — speed is the third swap-decision axis. Served latency
+          // mirrors the user's actual wait; candidate latency is measured only
+          // here and stored nowhere else (irrecoverable if not captured now).
+          latencyCurrentMs: args.servedLatencyMs,
+          latencyCandidateMs: candidateLatencyMs
         });
       } catch {
       }

package/dist/index.mjs

@@ -2844,6 +2844,8 @@ function buildShadowProbeRow(input) {
     tokens_current_out: input.tokensCurrentOut ?? null,
     tokens_candidate_in: input.tokensCandidateIn ?? null,
     tokens_candidate_out: input.tokensCandidateOut ?? null,
+    latency_current_ms: input.latencyCurrentMs ?? null,
+    latency_candidate_ms: input.latencyCandidateMs ?? null,
     // Full IR was replayed (not a truncated preview), so fidelity is 1.0 — the
     // prompt-fidelity guard never fires on these rows.
     prompt_fidelity: 1,
@@ -3649,7 +3651,8 @@ async function call(ir, opts = {}) {
           ir,
           opts,
           servedModel: targetModel,
-          servedResponse: validated.response
+          servedResponse: validated.response,
+          servedLatencyMs: latencyMs2
         });
         if (isBrainSync()) {
           await probe;
@@ -3762,11 +3765,13 @@ async function runShadowProbe(args) {
           },
           args.opts
         );
+        const candStart = Date.now();
         const exec = await execute(candCompile.request, {
           apiKeys: args.opts.apiKeys,
           fetchImpl: args.opts.fetchImpl,
           providerOverrides: args.opts.providerOverrides
         });
+        const candidateLatencyMs = Date.now() - candStart;
         if (!exec.ok) continue;
         await recordShadowProbe({
           appId: args.ir.appId,
@@ -3780,7 +3785,12 @@ async function runShadowProbe(args) {
           tokensCurrentIn: args.servedResponse.tokens.input,
           tokensCurrentOut: args.servedResponse.tokens.output,
           tokensCandidateIn: exec.response.tokens.input,
-          tokensCandidateOut: exec.response.tokens.output
+          tokensCandidateOut: exec.response.tokens.output,
+          // alpha.46 — speed is the third swap-decision axis. Served latency
+          // mirrors the user's actual wait; candidate latency is measured only
+          // here and stored nowhere else (irrecoverable if not captured now).
+          latencyCurrentMs: args.servedLatencyMs,
+          latencyCandidateMs: candidateLatencyMs
         });
       } catch {
       }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@warmdrift/kgauto-compiler",
-  "version": "2.0.0-alpha.45",
+  "version": "2.0.0-alpha.46",
   "description": "Prompt compiler + central learning brain for multi-model AI apps. Swap models without rewriting prompts.",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",