feat(ai): show provider response telemetry in the chat

AUTONOMOUS_STATE.json

@@ -7,7 +7,7 @@
     "asset-curator",
     "qa-playtester"
   ],
-  "last_cycle": 192,
+  "last_cycle": 193,
   "completed_targets": [
     {
       "cycle": 75,
@@ -1126,6 +1126,12 @@
       "target": "Make provider telemetry inspectable per request (Tier 2 audit data layer)",
       "metric_delta": "The x-omniroute-* telemetry was console-only; the objective wants full request/response inspectability. Added parseProviderTelemetry (numbers, drops non-numeric, null when nothing), refactored formatProviderTelemetry onto it, a per-request registry localAiRequestProviderTelemetry with customGetAiRequestProviderTelemetry(id), an onProviderTelemetry callback threaded through sendChatCompletion AND streamChatCompletion, and storage on all three turn paths (sub-agent attributes to the parent). Cleared by the test reset. +1 test (parse shape, non-numeric drop, live storage via a create with telemetry headers) red pre-fix, mutation-checked (removing the callback yields null). 1630/1630 across 81 suites, eslint+prettier clean, divergence clean.",
       "commit": "feat(ai): expose per-request provider telemetry"
+    },
+    {
+      "cycle": 193,
+      "target": "Surface the per-request provider telemetry in the chat (Tier 2 consumer)",
+      "metric_delta": "The cycle-192 telemetry registry had no consumer. Added formatProviderTelemetryNote (pure, in AiRequestChat/Utils.js) and a small 'Provider response: <model> · <ms> · ~in/out tokens · cache' line in ChatMessages' status block (fork-modified), reading customGetAiRequestProviderTelemetry. +3 tests (full format, missing token half, null cases) mutation-checked (dropping the cache part fails the full-format test). Note is only shown when the endpoint reports x-omniroute-* headers, so it stays silent for Ollama (which sends none) - verified live that Ollama emits none; the format itself is pinned by tests. 1633/1633 across 81 suites, eslint+prettier clean, production build exit 0, dev watcher compiled, divergence clean.",
+      "commit": "feat(ai): show provider response telemetry in the chat"
     }
   ],
   "backlog": [

newIDE/app/src/AiGeneration/AiRequestChat/CanPayForAiRequest.spec.js

@@ -14,6 +14,7 @@ import {
   shouldShowLocalColdStartHint,
   useLocalColdStartHint,
   hasModelResponded,
+  formatProviderTelemetryNote,
   COLD_START_HINT_DELAY_MS,
 } from './Utils';
 import {
@@ -528,6 +529,37 @@ describe('useLocalColdStartHint', () => {
   });
 });
 
+describe('formatProviderTelemetryNote', () => {
+  it('formats model, latency, tokens and cache', () => {
+    expect(
+      formatProviderTelemetryNote({
+        model: 'deepseek-v4.1-flash',
+        latencyMs: 3361.4,
+        tokensIn: 8856,
+        tokensOut: 90,
+        cache: 'MISS',
+      })
+    ).toBe(
+      'deepseek-v4.1-flash · 3361 ms · ~8856 in / 90 out tokens · cache MISS'
+    );
+  });
+
+  it('shows what is present and a ? for a missing half of the token pair', () => {
+    expect(formatProviderTelemetryNote({ tokensOut: 12 })).toBe(
+      '~? in / 12 out tokens'
+    );
+    expect(formatProviderTelemetryNote({ latencyMs: 20 })).toBe('20 ms');
+  });
+
+  it('is null when there is nothing to show', () => {
+    expect(formatProviderTelemetryNote(null)).toBe(null);
+    expect(formatProviderTelemetryNote(undefined)).toBe(null);
+    expect(formatProviderTelemetryNote({})).toBe(null);
+    expect(formatProviderTelemetryNote({ latencyMs: NaN })).toBe(null);
+    expect(formatProviderTelemetryNote({ model: '' })).toBe(null);
+  });
+});
+
 describe('hasModelResponded', () => {
   it('is false before the model produces anything', () => {
     expect(hasModelResponded(null)).toBe(false);

newIDE/app/src/AiGeneration/AiRequestChat/ChatMessages.js

@@ -53,6 +53,7 @@ import {
   getAvailableCredits,
   useLocalColdStartHint,
   hasModelResponded,
+  formatProviderTelemetryNote,
 } from './Utils';
 import Text from '../../UI/Text';
 import { ColumnStackLayout, LineStackLayout } from '../../UI/Layout';
@@ -69,6 +70,7 @@ import {
   customGetAiRequestSystemCompacted,
   customGetAiRequestTokenTotal,
   customGetAiRequestPartialContent,
+  customGetAiRequestProviderTelemetry,
 } from '../../AI/CustomAIClient';
 
 /**
@@ -1387,6 +1389,25 @@ export const ChatMessages: React.ComponentType<Props> = React.memo<Props>(
           </Line>
         ) : null}
 
+        {!isWorking &&
+        formatProviderTelemetryNote(
+          customGetAiRequestProviderTelemetry(aiRequest.id)
+        ) ? (
+          <Line justifyContent="flex-start">
+            <Text
+              noMargin
+              displayInlineAsSpan
+              size="body-small"
+              color="secondary"
+            >
+              <Trans>Provider response:</Trans>{' '}
+              {formatProviderTelemetryNote(
+                customGetAiRequestProviderTelemetry(aiRequest.id)
+              )}
+            </Text>
+          </Line>
+        ) : null}
+
         {aiRequest.status === 'error' ? (
           <Line justifyContent="flex-start">
             <AiRequestErrorRow

newIDE/app/src/AiGeneration/AiRequestChat/Utils.js

@@ -211,6 +211,39 @@ export const shouldShowSendAgainLabel = ({
  * turn — later slow turns are context-size slowness, not a model loading into
  * memory, and narrating every one as a cold start is noise.
  */
+/**
+ * A short, human-readable note for the provider telemetry a routing proxy
+ * reported for the last turn (`x-omniroute-*`), or null when there is nothing
+ * to show. Kept pure so the chat's rendering stays a one-liner and the format
+ * is unit-testable.
+ */
+export const formatProviderTelemetryNote = (telemetry: ?Object): ?string => {
+  if (!telemetry || typeof telemetry !== 'object') return null;
+  const parts = [];
+  if (typeof telemetry.model === 'string' && telemetry.model) {
+    parts.push(telemetry.model);
+  }
+  if (
+    typeof telemetry.latencyMs === 'number' &&
+    Number.isFinite(telemetry.latencyMs)
+  ) {
+    parts.push(`${Math.round(telemetry.latencyMs)} ms`);
+  }
+  const hasIn = typeof telemetry.tokensIn === 'number';
+  const hasOut = typeof telemetry.tokensOut === 'number';
+  if (hasIn || hasOut) {
+    parts.push(
+      `~${hasIn ? telemetry.tokensIn : '?'} in / ${
+        hasOut ? telemetry.tokensOut : '?'
+      } out tokens`
+    );
+  }
+  if (typeof telemetry.cache === 'string' && telemetry.cache) {
+    parts.push(`cache ${telemetry.cache}`);
+  }
+  return parts.length > 0 ? parts.join(' · ') : null;
+};
+
 export const hasModelResponded = (output: ?Array<Object>): boolean =>
   !!(output || []).some(
     message =>