diff --git a/src/app/api/insights/chat/__tests__/__snapshots__/golden-transcripts.test.ts.snap b/src/app/api/insights/chat/__tests__/__snapshots__/golden-transcripts.test.ts.snap
index b9570c8ea..7581c58cd 100644
--- a/src/app/api/insights/chat/__tests__/__snapshots__/golden-transcripts.test.ts.snap
+++ b/src/app/api/insights/chat/__tests__/__snapshots__/golden-transcripts.test.ts.snap
@@ -29,7 +29,7 @@ exports[`coach chat golden transcripts > 01a tool mode, two rounds, present data
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -78,7 +78,7 @@ exports[`coach chat golden transcripts > 01a tool mode, two rounds, present data
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "appendMessage",
@@ -278,7 +278,7 @@ exports[`coach chat golden transcripts > 01b tool mode, present data, one figure
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -327,7 +327,7 @@ exports[`coach chat golden transcripts > 01b tool mode, present data, one figure
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "annotate",
@@ -539,7 +539,7 @@ exports[`coach chat golden transcripts > 02 tool mode, every tool missed, fabric
       ],
     },
     {
-      "amount": 12889,
+      "amount": 13339,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -588,7 +588,7 @@ exports[`coach chat golden transcripts > 02 tool mode, every tool missed, fabric
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "annotate",
@@ -1415,7 +1415,7 @@ exports[`coach chat golden transcripts > 07a all providers failed (plain) 1`] =
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -1453,7 +1453,7 @@ exports[`coach chat golden transcripts > 07a all providers failed (plain) 1`] =
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 12891,
+      "reserved": 13341,
     },
     {
       "call": "annotate",
@@ -1507,7 +1507,7 @@ exports[`coach chat golden transcripts > 07b all providers failed (every hop rat
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -1545,7 +1545,7 @@ exports[`coach chat golden transcripts > 07b all providers failed (every hop rat
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 12891,
+      "reserved": 13341,
     },
     {
       "call": "annotate",
@@ -1599,7 +1599,7 @@ exports[`coach chat golden transcripts > 07c all providers failed (primary crede
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -1637,7 +1637,7 @@ exports[`coach chat golden transcripts > 07c all providers failed (primary crede
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 12891,
+      "reserved": 13341,
     },
     {
       "call": "annotate",
@@ -1691,7 +1691,7 @@ exports[`coach chat golden transcripts > 08 an unwrapped tagged provider error 1
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -1729,7 +1729,7 @@ exports[`coach chat golden transcripts > 08 an unwrapped tagged provider error 1
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 12891,
+      "reserved": 13341,
     },
     {
       "call": "annotate",
@@ -1784,7 +1784,7 @@ exports[`coach chat golden transcripts > 09 an aborted request leaves a cancelle
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -1822,7 +1822,7 @@ exports[`coach chat golden transcripts > 09 an aborted request leaves a cancelle
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 12891,
+      "reserved": 13341,
     },
     {
       "call": "appendMessage",
@@ -1883,7 +1883,7 @@ exports[`coach chat golden transcripts > 10 outbound block replaces the reply an
       ],
     },
     {
-      "amount": 12892,
+      "amount": 13342,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -1932,7 +1932,7 @@ exports[`coach chat golden transcripts > 10 outbound block replaces the reply an
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "annotate",
@@ -2262,7 +2262,7 @@ exports[`coach chat golden transcripts > 11a KEYVALUES valid 1`] = `
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -2311,7 +2311,7 @@ exports[`coach chat golden transcripts > 11a KEYVALUES valid 1`] = `
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "appendMessage",
@@ -2515,7 +2515,7 @@ exports[`coach chat golden transcripts > 11b KEYVALUES partial 1`] = `
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -2564,7 +2564,7 @@ exports[`coach chat golden transcripts > 11b KEYVALUES partial 1`] = `
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "annotate",
@@ -2766,7 +2766,7 @@ exports[`coach chat golden transcripts > 11c KEYVALUES malformed (no closing mar
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -2815,7 +2815,7 @@ exports[`coach chat golden transcripts > 11c KEYVALUES malformed (no closing mar
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "annotate",
@@ -3001,7 +3001,7 @@ exports[`coach chat golden transcripts > 11d sentinel-only reply is an empty rep
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -3050,7 +3050,7 @@ exports[`coach chat golden transcripts > 11d sentinel-only reply is an empty rep
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "annotate",
@@ -3104,7 +3104,7 @@ exports[`coach chat golden transcripts > 11e empty provider reply 1`] = `
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -3153,7 +3153,7 @@ exports[`coach chat golden transcripts > 11e empty provider reply 1`] = `
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
   ],
   "contentType": "text/event-stream; charset=utf-8",
@@ -3198,7 +3198,7 @@ exports[`coach chat golden transcripts > 12a SUGGEST-REMINDER surfaced 1`] = `
       ],
     },
     {
-      "amount": 12896,
+      "amount": 13346,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -3247,7 +3247,7 @@ exports[`coach chat golden transcripts > 12a SUGGEST-REMINDER surfaced 1`] = `
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1837,
+      "reserved": 2287,
     },
     {
       "call": "prisma.user.update",
@@ -3471,7 +3471,7 @@ exports[`coach chat golden transcripts > 12b SUGGEST-REMINDER suppressed 1`] = `
       ],
     },
     {
-      "amount": 12896,
+      "amount": 13346,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -3520,7 +3520,7 @@ exports[`coach chat golden transcripts > 12b SUGGEST-REMINDER suppressed 1`] = `
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1837,
+      "reserved": 2287,
     },
     {
       "call": "annotate",
@@ -3708,7 +3708,7 @@ exports[`coach chat golden transcripts > 13 SUGGEST-ACTION 1`] = `
       ],
     },
     {
-      "amount": 12892,
+      "amount": 13342,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -3757,7 +3757,7 @@ exports[`coach chat golden transcripts > 13 SUGGEST-ACTION 1`] = `
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "annotate",
@@ -3985,7 +3985,7 @@ exports[`coach chat golden transcripts > 14a REMEMBER valid 1`] = `
       ],
     },
     {
-      "amount": 12896,
+      "amount": 13346,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -4034,7 +4034,7 @@ exports[`coach chat golden transcripts > 14a REMEMBER valid 1`] = `
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1837,
+      "reserved": 2287,
     },
     {
       "call": "captureReminderFromSentinel",
@@ -4242,7 +4242,7 @@ exports[`coach chat golden transcripts > 14b REMEMBER malformed 1`] = `
       ],
     },
     {
-      "amount": 12896,
+      "amount": 13346,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -4291,7 +4291,7 @@ exports[`coach chat golden transcripts > 14b REMEMBER malformed 1`] = `
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1837,
+      "reserved": 2287,
     },
     {
       "call": "annotate",
@@ -4494,7 +4494,7 @@ exports[`coach chat golden transcripts > 15 an unknown learn link is scrubbed, a
       ],
     },
     {
-      "amount": 12895,
+      "amount": 13345,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -4543,7 +4543,7 @@ exports[`coach chat golden transcripts > 15 an unknown learn link is scrubbed, a
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1837,
+      "reserved": 2287,
     },
     {
       "call": "annotate",
@@ -5207,7 +5207,7 @@ exports[`coach chat golden transcripts > 17a a workout-scoped first turn 1`] = `
       ],
     },
     {
-      "amount": 12890,
+      "amount": 13340,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -5256,7 +5256,7 @@ exports[`coach chat golden transcripts > 17a a workout-scoped first turn 1`] = `
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "appendMessage",
@@ -5465,7 +5465,7 @@ exports[`coach chat golden transcripts > 17b a workout-scoped first turn with th
       ],
     },
     {
-      "amount": 12890,
+      "amount": 13340,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -5514,7 +5514,7 @@ exports[`coach chat golden transcripts > 17b a workout-scoped first turn with th
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "appendMessage",
@@ -5702,7 +5702,7 @@ exports[`coach chat golden transcripts > 18 a persistence throw after the provid
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -5751,7 +5751,7 @@ exports[`coach chat golden transcripts > 18 a persistence throw after the provid
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "event.setError",
@@ -5817,7 +5817,7 @@ exports[`coach chat golden transcripts > 19 an existing conversation with prior
       ],
     },
     {
-      "amount": 13053,
+      "amount": 13503,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -5866,7 +5866,7 @@ exports[`coach chat golden transcripts > 19 an existing conversation with prior
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1851,
+      "reserved": 2301,
     },
     {
       "call": "annotate",
@@ -6078,7 +6078,7 @@ exports[`coach chat golden transcripts > 20 a guided-question turn with a scope
       ],
     },
     {
-      "amount": 12892,
+      "amount": 13342,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -6127,7 +6127,7 @@ exports[`coach chat golden transcripts > 20 a guided-question turn with a scope
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "appendMessage",
@@ -6295,7 +6295,7 @@ exports[`coach chat golden transcripts > 21 the German locale with a stored self
       ],
     },
     {
-      "amount": 12917,
+      "amount": 13367,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -6344,7 +6344,7 @@ exports[`coach chat golden transcripts > 21 the German locale with a stored self
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1839,
+      "reserved": 2289,
     },
     {
       "call": "appendMessage",
@@ -6548,7 +6548,7 @@ exports[`coach chat golden transcripts > 22 tool mode with live steps: found, em
       ],
     },
     {
-      "amount": 12891,
+      "amount": 13341,
       "call": "reserveBudget",
       "cap": 2000000,
       "dateKey": "2026-09-26",
@@ -6597,7 +6597,7 @@ exports[`coach chat golden transcripts > 22 tool mode with live steps: found, em
       "cached": 0,
       "call": "reconcileSpend",
       "dateKey": "2026-09-26",
-      "reserved": 1836,
+      "reserved": 2286,
     },
     {
       "call": "appendMessage",
diff --git a/src/app/api/insights/chat/__tests__/dialog-harness.ts b/src/app/api/insights/chat/__tests__/dialog-harness.ts
index 14f787db9..c9b2dc079 100644
--- a/src/app/api/insights/chat/__tests__/dialog-harness.ts
+++ b/src/app/api/insights/chat/__tests__/dialog-harness.ts
@@ -63,6 +63,7 @@ export interface ProviderCall {
   system: string;
   messages: Array<{ role: string; content: string }>;
   toolChoice: string | undefined;
+  cacheKey: string | undefined;
   toolCalls: Array<{ name: string; args: Record<string, unknown> }>;
 }
 export const providerCalls: ProviderCall[] = [];
@@ -278,6 +279,7 @@ async function runRawCompletionWithFallback(args: {
     system: string;
     messages: Array<{ role: string; content: string }>;
     toolChoice?: string;
+    cacheKey?: string;
   };
 }) {
   const round = world.script.shift() ?? { text: "" };
@@ -289,6 +291,7 @@ async function runRawCompletionWithFallback(args: {
       content: msg.content,
     })),
     toolChoice: args.params.toolChoice,
+    cacheKey: args.params.cacheKey,
     toolCalls: calls,
   });
   const offered = args.params.toolChoice !== "none";
diff --git a/src/app/api/insights/chat/__tests__/route-activity.test.ts b/src/app/api/insights/chat/__tests__/route-activity.test.ts
index 92cb32fe5..04c689803 100644
--- a/src/app/api/insights/chat/__tests__/route-activity.test.ts
+++ b/src/app/api/insights/chat/__tests__/route-activity.test.ts
@@ -279,7 +279,9 @@ describe("concurrent turns", () => {
     expect(m.checkRateLimit).toHaveBeenCalledWith(
       "coach-turn-active:u1",
       2,
-      180_000,
+      // The longest turn's wall time (200 s since v1.41.2) plus 30 s for
+      // its reply to stream.
+      230_000,
     );
   });
 });
diff --git a/src/app/api/insights/chat/route.ts b/src/app/api/insights/chat/route.ts
index 971f391f8..b55b23d4f 100644
--- a/src/app/api/insights/chat/route.ts
+++ b/src/app/api/insights/chat/route.ts
@@ -40,6 +40,7 @@ import { runCoachTurn } from "@/lib/ai/coach/turn/pipeline";
 import type { ReasoningLevel } from "@/lib/ai/reasoning/levels";
 import { resolveCoachTurnReasoningLevel } from "@/lib/ai/reasoning/controls";
 import { streamProviderError, streamRefusal } from "@/lib/ai/coach/turn/sse";
+import { TURN_LIMITS } from "@/lib/ai/coach/tools/turn-budget";
 
 /**
  * The `coach` capability gate for a turn. Every unavailable reason throws the
@@ -72,11 +73,12 @@ async function coachCapabilityRefusal(): Promise<Response | null> {
  */
 const COACH_CONCURRENT_TURNS = 2;
 /**
- * How long a turn's slot is held at most: a turn's wall time (150 s) plus
- * the streaming of its reply. A slot whose release never arrived (a crashed
- * process) frees itself when the window ends.
+ * How long a turn's slot is held at most: the longest turn's wall time
+ * (`TURN_LIMITS.user.wallMs`, 200 s) plus 30 s for the streaming of its
+ * reply. A slot whose release never arrived (a crashed process) frees itself
+ * when the window ends.
  */
-const COACH_TURN_SLOT_MS = 180_000;
+const COACH_TURN_SLOT_MS = TURN_LIMITS.user.wallMs + 30_000;
 
 async function handleChatRequest(request: NextRequest): Promise<Response> {
   const auth = await requireAuth();
diff --git a/src/lib/ai/__tests__/cached-input-usage.test.ts b/src/lib/ai/__tests__/cached-input-usage.test.ts
new file mode 100644
index 000000000..2a555aad5
--- /dev/null
+++ b/src/lib/ai/__tests__/cached-input-usage.test.ts
@@ -0,0 +1,368 @@
+/**
+ * Cached input, per provider, end to end into the Coach turn budget.
+ *
+ * The turn budget counts cached input at a tenth (`CACHED_INPUT_WEIGHT`).
+ * That only works when two things hold for every client:
+ *
+ *   1. `tokensUsed` is the gross count, cached input included, and
+ *      `cachedInputTokens` is the cached share of it. The budget and the
+ *      daily ledger both subtract the cached share from the gross.
+ *   2. The cached count is read from where the provider puts it: Codex and
+ *      the Responses wire under `usage.input_tokens_details.cached_tokens`,
+ *      Chat Completions (OpenAI, gateways, local servers that report it)
+ *      under `usage.prompt_tokens_details.cached_tokens`, Anthropic as
+ *      `usage.cache_read_input_tokens` beside an `input_tokens` that does
+ *      NOT include it.
+ *
+ * Each fixture below is a usage payload in the provider's own shape for the
+ * same round: 20,000 tokens of input of which 18,000 came from the cache,
+ * and 500 of output. Every client must turn it into gross 20,500 / cached
+ * 18,000, which the budget weighs at 2,000 + 500 + 1,800 = 4,300.
+ *
+ * And the cache has to be hit at all: the Codex backend routes a request to
+ * a cache by `prompt_cache_key` (the official client sets it to its thread
+ * id), and `api.openai.com` documents the same field. A turn's rounds share
+ * one key, so round two finds round one's prefix.
+ */
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+vi.mock("undici", async (importOriginal) => {
+  const actual = await importOriginal<typeof import("undici")>();
+  return {
+    ...actual,
+    fetch: (input: unknown, init?: unknown) =>
+      (globalThis.fetch as unknown as (i: unknown, n?: unknown) => unknown)(
+        input,
+        init,
+      ),
+  };
+});
+vi.mock("../local-host-allowlist", () => ({ aiEgressPolicyFor: () => ({}) }));
+
+import { AnthropicClient } from "../anthropic-client";
+import { CodexClient } from "../codex-client";
+import { OpenAIClient } from "../openai-client";
+import { LocalOpenAICompatibleClient } from "../local-client";
+import { singleUserTurn, type CompletionResult } from "../types";
+import { weightedRoundTokens } from "@/lib/ai/coach/tools/turn-budget";
+
+const INPUT = 20_000;
+const CACHED = 18_000;
+const OUTPUT = 500;
+const GROSS = INPUT + OUTPUT;
+const WEIGHTED = INPUT - CACHED + OUTPUT + CACHED / 10;
+
+function expectDiscounted(result: CompletionResult) {
+  expect(result.tokensUsed).toBe(GROSS);
+  expect(result.cachedInputTokens).toBe(CACHED);
+  expect(
+    weightedRoundTokens({
+      tokens: result.tokensUsed,
+      cachedTokens: result.cachedInputTokens,
+    }),
+  ).toBe(WEIGHTED);
+}
+
+function sse(events: string[]): Response {
+  const encoder = new TextEncoder();
+  return new Response(
+    new ReadableStream<Uint8Array>({
+      start(controller) {
+        for (const ev of events) controller.enqueue(encoder.encode(ev));
+        controller.close();
+      },
+    }),
+    { status: 200, headers: { "content-type": "text/event-stream" } },
+  );
+}
+
+function codexStream(): Response {
+  return sse([
+    `event: response.output_item.done\ndata: ${JSON.stringify({
+      type: "response.output_item.done",
+      item: {
+        type: "message",
+        role: "assistant",
+        content: [{ type: "output_text", text: "ok" }],
+      },
+    })}\n\n`,
+    `event: response.completed\ndata: ${JSON.stringify({
+      type: "response.completed",
+      response: {
+        id: "resp1",
+        usage: {
+          input_tokens: INPUT,
+          input_tokens_details: { cached_tokens: CACHED },
+          output_tokens: OUTPUT,
+          output_tokens_details: { reasoning_tokens: 0 },
+          total_tokens: GROSS,
+        },
+      },
+    })}\n\n`,
+  ]);
+}
+
+/** The Chat Completions usage block, as OpenAI and most gateways send it. */
+const chatUsage = {
+  prompt_tokens: INPUT,
+  completion_tokens: OUTPUT,
+  total_tokens: GROSS,
+  prompt_tokens_details: { cached_tokens: CACHED },
+};
+
+function jsonResponse(body: unknown) {
+  return {
+    ok: true,
+    status: 200,
+    headers: { get: () => "application/json" },
+    json: () => Promise.resolve(body),
+    text: () => Promise.resolve(JSON.stringify(body)),
+  };
+}
+
+function codexClient() {
+  return new CodexClient({
+    accessToken: "t",
+    accountId: "a",
+    onTokenRefresh: vi
+      .fn()
+      .mockResolvedValue({ accessToken: "t", accountId: "a" }),
+    slugChain: ["gpt-5-codex"],
+  });
+}
+
+beforeEach(() => {
+  vi.restoreAllMocks();
+});
+
+describe("cached input reaches the turn budget, per provider", () => {
+  it("codex: input_tokens_details.cached_tokens", async () => {
+    vi.stubGlobal("fetch", vi.fn().mockResolvedValue(codexStream()));
+    const result = await codexClient().generateCompletion(
+      singleUserTurn({ system: "s", user: "u" }),
+    );
+    expectDiscounted(result);
+  });
+
+  it("openai (Chat Completions): prompt_tokens_details.cached_tokens", async () => {
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue(
+        jsonResponse({
+          choices: [{ message: { content: "ok" }, finish_reason: "stop" }],
+          usage: chatUsage,
+        }),
+      ),
+    );
+    const client = new OpenAIClient({
+      apiKey: "sk-test",
+      model: "gpt-4o-mini",
+      baseUrl: "https://api.openai.com/v1",
+    });
+    expectDiscounted(
+      await client.generateCompletion(
+        singleUserTurn({ system: "s", user: "u" }),
+      ),
+    );
+  });
+
+  it("gateway (OpenRouter and the like): the same Chat Completions block", async () => {
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue(
+        jsonResponse({
+          choices: [{ message: { content: "ok" }, finish_reason: "stop" }],
+          usage: chatUsage,
+        }),
+      ),
+    );
+    const client = new OpenAIClient({
+      apiKey: "sk-or-test",
+      model: "openai/gpt-4o-mini",
+      baseUrl: "https://openrouter.ai/api/v1",
+      providerType: "openai-compatible",
+    });
+    expectDiscounted(
+      await client.generateCompletion(
+        singleUserTurn({ system: "s", user: "u" }),
+      ),
+    );
+  });
+
+  it("anthropic: cache reads ride beside input_tokens, not inside it", async () => {
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue(
+        jsonResponse({
+          content: [{ type: "text", text: "ok" }],
+          stop_reason: "end_turn",
+          usage: {
+            // Anthropic's input_tokens counts only what came after the last
+            // cache breakpoint; reads and writes are reported beside it.
+            input_tokens: INPUT - CACHED,
+            cache_read_input_tokens: CACHED,
+            cache_creation_input_tokens: 0,
+            output_tokens: OUTPUT,
+          },
+        }),
+      ),
+    );
+    const client = new AnthropicClient({
+      apiKey: "sk-ant-test",
+      model: "claude-sonnet-4-5",
+    });
+    expectDiscounted(
+      await client.generateCompletion(
+        singleUserTurn({ system: "s", user: "u" }),
+      ),
+    );
+  });
+
+  it("anthropic: a cache write counts in full", async () => {
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue(
+        jsonResponse({
+          content: [{ type: "text", text: "ok" }],
+          stop_reason: "end_turn",
+          usage: {
+            input_tokens: 2_000,
+            cache_read_input_tokens: 0,
+            cache_creation_input_tokens: 18_000,
+            output_tokens: OUTPUT,
+          },
+        }),
+      ),
+    );
+    const client = new AnthropicClient({
+      apiKey: "sk-ant-test",
+      model: "claude-sonnet-4-5",
+    });
+    const result = await client.generateCompletion(
+      singleUserTurn({ system: "s", user: "u" }),
+    );
+    expect(result.tokensUsed).toBe(GROSS);
+    expect(result.cachedInputTokens ?? 0).toBe(0);
+  });
+
+  it("local: prompt_tokens_details.cached_tokens where the server reports it", async () => {
+    vi.stubGlobal(
+      "fetch",
+      vi.fn().mockResolvedValue(
+        jsonResponse({
+          choices: [{ message: { content: "ok" }, finish_reason: "stop" }],
+          usage: chatUsage,
+        }),
+      ),
+    );
+    const client = new LocalOpenAICompatibleClient({
+      apiKey: null,
+      model: "llama3:8b",
+      baseUrl: "http://localhost:11434/v1",
+    });
+    expectDiscounted(
+      await client.generateCompletion(
+        singleUserTurn({ system: "s", user: "u" }),
+      ),
+    );
+  });
+
+  it("local, streamed: the usage frame's cached count", async () => {
+    vi.stubGlobal(
+      "fetch",
+      vi
+        .fn()
+        .mockResolvedValue(
+          sse([
+            'data: {"choices":[{"delta":{"content":"ok"}}]}\n\n',
+            `data: ${JSON.stringify({ choices: [{ delta: {}, finish_reason: "stop" }], usage: chatUsage })}\n\n`,
+            "data: [DONE]\n\n",
+          ]),
+        ),
+    );
+    const client = new LocalOpenAICompatibleClient({
+      apiKey: null,
+      model: "llama3:8b",
+      baseUrl: "http://localhost:11434/v1",
+    });
+    expectDiscounted(
+      await client.generateCompletionStream(
+        singleUserTurn({ system: "s", user: "u" }),
+        () => {},
+      ),
+    );
+  });
+});
+
+describe("a turn's rounds share one prompt cache key", () => {
+  it("codex sends prompt_cache_key and one session id for every round with the same key", async () => {
+    const fetchMock = vi
+      .fn()
+      .mockImplementation(() => Promise.resolve(codexStream()));
+    vi.stubGlobal("fetch", fetchMock);
+    const client = codexClient();
+    const params = {
+      ...singleUserTurn({ system: "s", user: "u" }),
+      cacheKey: "coach-turn-key",
+    };
+    await client.generateCompletion(params);
+    await client.generateCompletion(params);
+
+    const [first, second] = fetchMock.mock.calls.map(
+      (call) => call[1] as { body: string; headers: Record<string, string> },
+    );
+    expect(JSON.parse(first.body).prompt_cache_key).toBe("coach-turn-key");
+    expect(JSON.parse(second.body).prompt_cache_key).toBe("coach-turn-key");
+    expect(first.headers.session_id).toBe(second.headers.session_id);
+    expect(first.headers.session_id).toMatch(
+      /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/,
+    );
+  });
+
+  it("codex without a key sends none and keeps a fresh session per call", async () => {
+    const fetchMock = vi
+      .fn()
+      .mockImplementation(() => Promise.resolve(codexStream()));
+    vi.stubGlobal("fetch", fetchMock);
+    const client = codexClient();
+    await client.generateCompletion(singleUserTurn({ system: "s", user: "u" }));
+    await client.generateCompletion(singleUserTurn({ system: "s", user: "u" }));
+    const [first, second] = fetchMock.mock.calls.map(
+      (call) => call[1] as { body: string; headers: Record<string, string> },
+    );
+    expect("prompt_cache_key" in JSON.parse(first.body)).toBe(false);
+    expect(first.headers.session_id).not.toBe(second.headers.session_id);
+  });
+
+  it("api.openai.com gets prompt_cache_key; a gateway never does", async () => {
+    const reply = () =>
+      Promise.resolve(
+        jsonResponse({
+          choices: [{ message: { content: "ok" }, finish_reason: "stop" }],
+          usage: chatUsage,
+        }),
+      );
+    const fetchMock = vi.fn().mockImplementation(reply);
+    vi.stubGlobal("fetch", fetchMock);
+    const params = {
+      ...singleUserTurn({ system: "s", user: "u" }),
+      cacheKey: "coach-turn-key",
+    };
+    await new OpenAIClient({
+      apiKey: "sk-test",
+      model: "gpt-4o-mini",
+      baseUrl: "https://api.openai.com/v1",
+    }).generateCompletion(params);
+    await new OpenAIClient({
+      apiKey: "sk-or-test",
+      model: "openai/gpt-4o-mini",
+      baseUrl: "https://openrouter.ai/api/v1",
+      providerType: "openai-compatible",
+    }).generateCompletion(params);
+    const [direct, gateway] = fetchMock.mock.calls.map((call) =>
+      JSON.parse((call[1] as { body: string }).body),
+    );
+    expect(direct.prompt_cache_key).toBe("coach-turn-key");
+    expect("prompt_cache_key" in gateway).toBe(false);
+  });
+});
diff --git a/src/lib/ai/anthropic-client.ts b/src/lib/ai/anthropic-client.ts
index 654de706b..79a681234 100644
--- a/src/lib/ai/anthropic-client.ts
+++ b/src/lib/ai/anthropic-client.ts
@@ -412,6 +412,7 @@ export class AnthropicClient implements AIProvider {
         input_tokens?: number;
         output_tokens?: number;
         cache_read_input_tokens?: number;
+        cache_creation_input_tokens?: number;
         output_tokens_details?: { thinking_tokens?: number };
       };
     };
@@ -487,9 +488,19 @@ export class AnthropicClient implements AIProvider {
     const content =
       wantsJson && rawText ? extractJsonObject(rawText) : (rawText ?? "");
 
+    // v1.41.2 — the gross count includes the cached prefix. Anthropic reports
+    // cache reads and cache writes BESIDE `input_tokens`, which holds only
+    // what came after the last breakpoint; every other client reports a total
+    // that already contains the cached share. The turn budget and the daily
+    // ledger subtract `cachedInputTokens` from `tokensUsed`, so leaving the
+    // reads out of the gross made a mostly cached round look nearly free and
+    // a cache write (billed above the base rate) went uncounted.
     const inputTokens = json.usage?.input_tokens ?? 0;
+    const cacheReadTokens = json.usage?.cache_read_input_tokens ?? 0;
+    const cacheWriteTokens = json.usage?.cache_creation_input_tokens ?? 0;
     const outputTokens = json.usage?.output_tokens ?? 0;
-    const tokensUsed = inputTokens + outputTokens || null;
+    const tokensUsed =
+      inputTokens + cacheReadTokens + cacheWriteTokens + outputTokens || null;
     const cachedInputTokens = json.usage?.cache_read_input_tokens ?? null;
 
     // The presence of tool_use blocks is authoritative for the tool-loop gate:
diff --git a/src/lib/ai/coach/eval/__tests__/scenarios-live.test.ts b/src/lib/ai/coach/eval/__tests__/scenarios-live.test.ts
index ecf3383e1..61891818d 100644
--- a/src/lib/ai/coach/eval/__tests__/scenarios-live.test.ts
+++ b/src/lib/ai/coach/eval/__tests__/scenarios-live.test.ts
@@ -5,6 +5,7 @@
  * chat turn's own context, the reply through the turn's parsers) and the
  * grade, so a live run measures the model and nothing else.
  */
+import { TURN_LIMITS } from "@/lib/ai/coach/tools/turn-budget";
 import { afterEach, describe, expect, it, vi } from "vitest";
 
 import type { CompletionParams, CompletionResult } from "@/lib/ai/types";
@@ -119,23 +120,27 @@ describe("runScenarioLive", () => {
 
   it("forces an answer at the round cap of the person's own plan", async () => {
     const call = toolRound("get_metric_table", { metric: "pulse" });
+    // v1.41.2 — the cap is sixteen (was twelve); read from the limits so
+    // the scenario run and the turn cannot drift apart again.
+    const cap = TURN_LIMITS.user.maxRounds;
+    expect(cap).toBe(16);
     const provider = scripted([
-      ...Array.from({ length: 11 }, () => call),
+      ...Array.from({ length: cap - 1 }, () => call),
       { content: "So far." },
     ]);
     const observation = await runScenarioLive({
       scenario: scenario("forced-final.en"),
       provider,
     });
-    expect(observation.providerCalls).toBe(12);
+    expect(observation.providerCalls).toBe(cap);
     expect(provider.sent.map((p) => p.toolChoice)).toEqual([
-      ...Array.from({ length: 11 }, () => "auto"),
+      ...Array.from({ length: cap - 1 }, () => "auto"),
       "none",
     ]);
     expect(observation.prose).toBe("So far.");
-    expect(observation.toolRounds).toBe(11);
+    expect(observation.toolRounds).toBe(cap - 1);
     // The same call every round: graded as repeats.
-    expect(observation.repeatedCalls).toBe(10);
+    expect(observation.repeatedCalls).toBe(cap - 2);
   });
 
   it("ends the run on a clarifying question asked through the tool", async () => {
diff --git a/src/lib/ai/coach/eval/__tests__/scenarios.test.ts b/src/lib/ai/coach/eval/__tests__/scenarios.test.ts
index 9b077b2fc..bfdbc0325 100644
--- a/src/lib/ai/coach/eval/__tests__/scenarios.test.ts
+++ b/src/lib/ai/coach/eval/__tests__/scenarios.test.ts
@@ -63,6 +63,7 @@ import type {
   CoachStep,
 } from "@/lib/ai/coach/types";
 import { COACH_FOLLOW_UP_KEYS } from "@/lib/ai/coach/dialog-keys";
+import { TURN_LIMITS } from "@/lib/ai/coach/tools/turn-budget";
 import { POST } from "@/app/api/insights/chat/route";
 
 import {
@@ -130,7 +131,7 @@ const call = (name: string, args: Record<string, unknown> = {}) => ({
 });
 
 /** Ten distinct reads: with `get_sleep` first, eleven fetching rounds. */
-const FORCED_READS: ReadonlyArray<[string, string]> = [
+const FORCED_READS: ReadonlyArray<[string, string, string?]> = [
   ["pulse", "last7days"],
   ["pulse", "last30days"],
   ["pulse", "last90days"],
@@ -141,6 +142,11 @@ const FORCED_READS: ReadonlyArray<[string, string]> = [
   ["bp", "lastYear"],
   ["sleep", "last90days"],
   ["sleep", "lastYear"],
+  // v1.41.2 — four more for the sixteen-round cap.
+  ["sleep", "last7days"],
+  ["pulse", "last30days", "previous"],
+  ["sleep", "last30days"],
+  ["bp", "last30days", "previous"],
 ];
 
 function script(name: string, lang: Lang): Round[] {
@@ -204,12 +210,18 @@ function script(name: string, lang: Lang): Round[] {
       return [];
     case "forced-final":
       // v1.41 — a new read every round, never a repeat, until the round cap
-      // of the person's own plan (twelve, the forced answer included) makes
-      // the twelfth round the answer.
+      // of the person's own plan (sixteen since v1.41.2, the forced answer
+      // included) makes the last round the answer.
       return [
         { calls: [call("get_sleep")] },
-        ...FORCED_READS.map(([metric, window]) => ({
-          calls: [call("get_metric_table", { metric, window })],
+        ...FORCED_READS.map(([metric, window, period]) => ({
+          calls: [
+            call("get_metric_table", {
+              metric,
+              window,
+              ...(period ? { period } : {}),
+            }),
+          ],
         })),
         { text: PROSE.forced[lang] },
       ];
@@ -766,9 +778,14 @@ describe.each(["de", "en"] as const)("in %s", (lang) => {
     const scenario = byName("forced-final", lang);
     const { observation, frames } = await run(scenario);
     expect(providerCalls.map((c) => c.toolChoice)).toEqual([
-      ...Array.from({ length: 11 }, () => "auto"),
+      ...Array.from({ length: TURN_LIMITS.user.maxRounds - 1 }, () => "auto"),
       "none",
     ]);
+    // v1.41.2 — every round of the turn carries the same prompt-cache key,
+    // so a provider that routes by key finds the prefix the last round left.
+    const keys = new Set(providerCalls.map((c) => c.cacheKey));
+    expect(keys.size).toBe(1);
+    expect([...keys][0]).toMatch(/^[0-9a-f]{32}$/);
     const provenance = framesOf<{ metricSource: { forcedFinal?: boolean } }>(
       frames,
       "provenance",
diff --git a/src/lib/ai/coach/eval/run-case.ts b/src/lib/ai/coach/eval/run-case.ts
index f6d2856b0..9eecf3986 100644
--- a/src/lib/ai/coach/eval/run-case.ts
+++ b/src/lib/ai/coach/eval/run-case.ts
@@ -26,6 +26,7 @@
  */
 import type { Locale } from "@/lib/i18n/config";
 import { UNBOUNDED_REACH } from "@/lib/ai/coach/history-reach";
+import { TURN_LIMITS } from "@/lib/ai/coach/tools/turn-budget";
 import type {
   AiMessage,
   AiToolCall,
@@ -145,7 +146,7 @@ export interface ScenarioProvider {
  * How many rounds a live scenario may take: the round cap of the person's
  * own plan (v1.41), the forced answer included.
  */
-const SCENARIO_MAX_ROUNDS = 12;
+const SCENARIO_MAX_ROUNDS = TURN_LIMITS.user.maxRounds;
 
 function parseArgs(raw: string): Record<string, unknown> {
   try {
@@ -236,7 +237,7 @@ function priorResultTurns(scenario: CoachScenario) {
  * the chat turn's own context (system prompt, tool-mode and dialog
  * addenda, DATA INVENTORY, EARLIER TABLES, transcript), put to a real
  * model over the real tool catalogue, comparisons and the dialog tools
- * for up to twelve rounds (v1.41), a question ending the run. Every call is
+ * for as many rounds as a turn on the person's own plan (v1.41), a question ending the run. Every call is
  * captured with its arguments and answered with a stand-in result; the
  * reply goes through the clarification and chip parsers the turn uses.
  *
diff --git a/src/lib/ai/coach/prompt-cache-key.ts b/src/lib/ai/coach/prompt-cache-key.ts
new file mode 100644
index 000000000..c70a12e96
--- /dev/null
+++ b/src/lib/ai/coach/prompt-cache-key.ts
@@ -0,0 +1,20 @@
+/**
+ * v1.41.2 — the prompt-cache key of a Coach conversation.
+ *
+ * Every round of a turn resends the same system prompt and history, and the
+ * next turn of the conversation resends it again with one more exchange. A
+ * provider that routes requests to its prompt cache by key (Codex,
+ * `api.openai.com`) only finds the prefix an earlier call left there when the
+ * calls carry the same key. One key per conversation does that; a turn with
+ * no conversation yet gets a key of its own, still shared by its rounds.
+ *
+ * Hashed so the provider never sees the conversation id itself.
+ */
+import { createHash, randomUUID } from "node:crypto";
+
+export function coachPromptCacheKey(conversationId?: string | null): string {
+  return createHash("sha256")
+    .update(`coach-prompt:${conversationId || randomUUID()}`)
+    .digest("hex")
+    .slice(0, 32);
+}
diff --git a/src/lib/ai/coach/stream-events.ts b/src/lib/ai/coach/stream-events.ts
index 5c54094bb..c0757bc43 100644
--- a/src/lib/ai/coach/stream-events.ts
+++ b/src/lib/ai/coach/stream-events.ts
@@ -76,7 +76,7 @@ export const coachStepSchema = z
     id: z
       .string()
       .regex(/^s[1-9]\d{0,2}$/)
-      .describe("`s1`..`s48`, unique within a turn."),
+      .describe("`s1`..`s64`, unique within a turn."),
     tool: z.enum([...COACH_TOOL_NAMES, "show_result", "snapshot"]),
     labelKey: z
       .string()
diff --git a/src/lib/ai/coach/tools/__tests__/turn-budget.test.ts b/src/lib/ai/coach/tools/__tests__/turn-budget.test.ts
index 2a251d442..b43779869 100644
--- a/src/lib/ai/coach/tools/__tests__/turn-budget.test.ts
+++ b/src/lib/ai/coach/tools/__tests__/turn-budget.test.ts
@@ -25,18 +25,40 @@ function clock(start = 0) {
 }
 
 describe("TURN_LIMITS", () => {
+  // v1.41.2 — raised by about a third from 120k / 150 s / 12 and
+  // 40k / 90 s / 6: a turn on a person's own plan stopped for tokens after
+  // five rounds while it was still reading.
   it("gives the person's own plan more than the operator's", () => {
     expect(TURN_LIMITS.user).toEqual({
-      tokens: 120_000,
-      wallMs: 150_000,
-      maxRounds: 12,
+      tokens: 165_000,
+      wallMs: 200_000,
+      maxRounds: 16,
     });
     expect(TURN_LIMITS.operator).toEqual({
-      tokens: 40_000,
-      wallMs: 90_000,
-      maxRounds: 6,
+      tokens: 55_000,
+      wallMs: 120_000,
+      maxRounds: 8,
     });
   });
+
+  it("keeps the final answer at the same share of each payer's tokens", () => {
+    expect(FINAL_ANSWER_TOKENS / TURN_LIMITS.user.tokens).toBeCloseTo(0.01);
+    expect(FINAL_ANSWER_TOKENS / TURN_LIMITS.operator.tokens).toBeCloseTo(0.03);
+  });
+
+  it("holds the turn's slot for at least the longest turn", async () => {
+    // The chat route's concurrency slot frees itself after its window; a
+    // window shorter than the wall time would let a third turn start while
+    // a long one is still running.
+    const { readFileSync } = await import("node:fs");
+    const route = readFileSync(
+      `${process.cwd()}/src/app/api/insights/chat/route.ts`,
+      "utf8",
+    );
+    expect(route).toMatch(
+      /COACH_TURN_SLOT_MS = TURN_LIMITS\.user\.wallMs \+ \d/,
+    );
+  });
 });
 
 describe("token accounting", () => {
@@ -73,9 +95,9 @@ describe("check()", () => {
       effort: "medium",
       initialInputTokens: 10_000,
     });
-    // 34k spent; the next round with medium thinking and the final answer
-    // need about 8k more, past the operator's 40k.
-    budget.endRound({ tokens: 34_000, cachedTokens: 0, durationMs: 1_000 });
+    // 49k spent; the next round with medium thinking and the final answer
+    // need about 8k more, past the operator's 55k.
+    budget.endRound({ tokens: 49_000, cachedTokens: 0, durationMs: 1_000 });
     expect(budget.check()).toBe("budget");
   });
 
@@ -103,9 +125,9 @@ describe("check()", () => {
       initialInputTokens: 1_000,
       now: c.now,
     });
-    c.advance(30_000);
+    c.advance(60_000);
     budget.endRound({ tokens: 1_000, cachedTokens: 0, durationMs: 30_000 });
-    // 30 s elapsed + two typical 30 s rounds = 90 s: not over yet.
+    // 60 s elapsed + two typical 30 s rounds = 120 s: not over yet.
     expect(budget.check()).toBeNull();
     c.advance(1);
     budget.endRound({ tokens: 1_000, cachedTokens: 900, durationMs: 30_000 });
diff --git a/src/lib/ai/coach/tools/loop.ts b/src/lib/ai/coach/tools/loop.ts
index f784a9096..ea45a44ad 100644
--- a/src/lib/ai/coach/tools/loop.ts
+++ b/src/lib/ai/coach/tools/loop.ts
@@ -30,6 +30,7 @@
  * A clarifying question ends the turn without a prose round: the question
  * is the reply.
  */
+import { coachPromptCacheKey } from "@/lib/ai/coach/prompt-cache-key";
 import type { CoachHistoryReach } from "@/lib/ai/coach/history-reach";
 import { annotate } from "@/lib/logging/context";
 import type { Locale } from "@/lib/i18n/config";
@@ -277,6 +278,9 @@ export async function runCoachToolLoop(args: {
       ),
     });
   const progress = createProgressTracker();
+  // v1.41.2 — one key for every round, so round two finds round one's prefix
+  // in the provider's prompt cache.
+  const cacheKey = coachPromptCacheKey(turn?.conversationId);
 
   const messages: AiMessage[] = [...args.messages];
   let system = args.system;
@@ -351,6 +355,7 @@ export async function runCoachToolLoop(args: {
           : {}),
         tools,
         toolChoice: isFinal ? ("none" as const) : ("auto" as const),
+        cacheKey,
       },
     });
     const result = fallback.result;
diff --git a/src/lib/ai/coach/tools/turn-budget.ts b/src/lib/ai/coach/tools/turn-budget.ts
index 3d021441c..16f8ba986 100644
--- a/src/lib/ai/coach/tools/turn-budget.ts
+++ b/src/lib/ai/coach/tools/turn-budget.ts
@@ -46,17 +46,26 @@ export interface TurnLimits {
  * Fixed per payer, not configurable: the person's own plan, key or local
  * model gets room for a long "why" chain; a turn the operator pays for gets
  * a third of the tokens and half the rounds.
+ *
+ * v1.41.2 — about a third more on both sides (was 120k / 150 s / 12 rounds
+ * and 40k / 90 s / 6 rounds). A real turn on a person's own plan stopped for
+ * tokens after five rounds while it was still reading, and the operator
+ * side moves with it so the split stays the same.
  */
 export const TURN_LIMITS: Readonly<Record<TurnPayer, TurnLimits>> = {
-  user: { tokens: 120_000, wallMs: 150_000, maxRounds: 12 },
-  operator: { tokens: 40_000, wallMs: 90_000, maxRounds: 6 },
+  user: { tokens: 165_000, wallMs: 200_000, maxRounds: 16 },
+  operator: { tokens: 55_000, wallMs: 120_000, maxRounds: 8 },
 };
 
 /** The answer budget of one round, before any thinking budget. */
 export const ROUND_ANSWER_TOKENS = 600;
 
-/** What the final answer is given on top of its input. */
-export const FINAL_ANSWER_TOKENS = 1_200;
+/**
+ * What the final answer is given on top of its input. v1.41.2 — raised with
+ * the limits (was 1,200) so it keeps the same share of each payer's tokens:
+ * 1 % of the person's own budget, 3 % of the operator's.
+ */
+export const FINAL_ANSWER_TOKENS = 1_650;
 
 /** Cached input counts at this share of its size. */
 export const CACHED_INPUT_WEIGHT = 0.1;
diff --git a/src/lib/ai/coach/turn/__tests__/steps.test.ts b/src/lib/ai/coach/turn/__tests__/steps.test.ts
index af70675a1..ccd097b50 100644
--- a/src/lib/ai/coach/turn/__tests__/steps.test.ts
+++ b/src/lib/ai/coach/turn/__tests__/steps.test.ts
@@ -462,7 +462,7 @@ describe("toStep — mapping", () => {
         parsedArgs: parsed,
         locale: "en",
       })?.id,
-    ).toBe("s48");
+    ).toBe("s64");
     expect(
       toStep({
         call: { id: "u", name: "get_everything", arguments: "{}" },
diff --git a/src/lib/ai/coach/turn/model.ts b/src/lib/ai/coach/turn/model.ts
index dc2a44810..f8c7996f7 100644
--- a/src/lib/ai/coach/turn/model.ts
+++ b/src/lib/ai/coach/turn/model.ts
@@ -4,6 +4,7 @@
  * completion over the full snapshot. A provider failure is classified into
  * the structured `coach.*` code the stream answers with.
  */
+import { coachPromptCacheKey } from "@/lib/ai/coach/prompt-cache-key";
 import { annotate } from "@/lib/logging/context";
 import type { Locale } from "@/lib/i18n/config";
 import { localeLanguageNames as LANGUAGE_NAMES } from "@/lib/i18n/config";
@@ -652,6 +653,7 @@ export async function runTurnModel(args: {
           // v1.22 (#89) — per-idle-gap timeout for the streaming local call.
           timeoutMs: ctx.aiResponseTimeoutMs,
         }),
+        cacheKey: coachPromptCacheKey(conversationId),
         ...(args.reasoning
           ? {
               reasoning: args.reasoning,
diff --git a/src/lib/ai/coach/turn/steps.ts b/src/lib/ai/coach/turn/steps.ts
index 2543185db..d09ce5d3f 100644
--- a/src/lib/ai/coach/turn/steps.ts
+++ b/src/lib/ai/coach/turn/steps.ts
@@ -40,9 +40,10 @@ import { COMPARE_SERIES_TOOL_NAME } from "@/lib/ai/coach/tools/compare-series";
 
 /**
  * At most this many steps per turn; later calls run but show no step.
- * v1.41 — 48: a budgeted turn runs up to twelve rounds.
+ * v1.41 — four per round of the longest turn: 48 for twelve rounds, 64 since
+ * v1.41.2 raised the cap to sixteen.
  */
-export const MAX_TURN_STEPS = 48;
+export const MAX_TURN_STEPS = 64;
 
 const SCOPE_SOURCES: ReadonlySet<string> = new Set(
   coachScopeSourceSchema.options,
diff --git a/src/lib/ai/coach/types.ts b/src/lib/ai/coach/types.ts
index 2db401956..686e6232f 100644
--- a/src/lib/ai/coach/types.ts
+++ b/src/lib/ai/coach/types.ts
@@ -498,7 +498,7 @@ export type CoachStepReason =
   | "invalid_arguments";
 
 export interface CoachStep {
-  /** `s1`..`s48`, unique within a turn; frames upsert by id. */
+  /** `s1`..`s64`, unique within a turn; frames upsert by id. */
   id: string;
   tool: CoachToolName | "show_result" | "snapshot";
   /** Closed catalog key, e.g. `coach.step.read`. */
diff --git a/src/lib/ai/codex-client.ts b/src/lib/ai/codex-client.ts
index 8a4fe0fdf..2eb7f2eaf 100644
--- a/src/lib/ai/codex-client.ts
+++ b/src/lib/ai/codex-client.ts
@@ -1,4 +1,4 @@
-import { randomUUID } from "node:crypto";
+import { createHash, randomUUID } from "node:crypto";
 import { callTimeoutMs } from "./effective-timeout";
 import { safeFetch } from "@/lib/safe-fetch";
 import type {
@@ -431,6 +431,21 @@ export interface CodexAttemptDiagnostics {
   workingSlug: string | null;
 }
 
+/**
+ * A UUID-shaped id derived from a cache key, for the session headers: the
+ * same key gives the same id, and the key itself never rides a header.
+ */
+function uuidFromKey(key: string): string {
+  const hex = createHash("sha256").update(`codex-session:${key}`).digest("hex");
+  return [
+    hex.slice(0, 8),
+    hex.slice(8, 12),
+    `4${hex.slice(13, 16)}`,
+    `${((parseInt(hex[16], 16) & 0x3) | 0x8).toString(16)}${hex.slice(17, 20)}`,
+    hex.slice(20, 32),
+  ].join("-");
+}
+
 export class CodexClient implements AIProvider {
   readonly type = "codex" as const;
   /** See `AIProvider.responseTimeoutSeconds`; stamped by the resolver. */
@@ -650,8 +665,15 @@ export class CodexClient implements AIProvider {
     slug: string,
     wire: CodexReasoningWire,
   ): Promise<Response> {
-    const sessionId = randomUUID();
-    const threadId = randomUUID();
+    // v1.41.2 — calls that share a cache key share one session and thread
+    // id and send the key as `prompt_cache_key`, the way the official client
+    // sends its thread id (docs/codex-protocol-spec.md §3b/§3c). Before, every
+    // round of a Coach turn arrived as a new session with no key, so the
+    // backend had no reason to route it to the cache the previous round
+    // filled, and each round was charged its whole prefix again.
+    const cacheKey = params.cacheKey?.trim() || null;
+    const sessionId = cacheKey ? uuidFromKey(cacheKey) : randomUUID();
+    const threadId = cacheKey ? sessionId : randomUUID();
 
     // v1.18.11 — fold vision inputs (Lab-OCR) into the user turn when present.
     // The Codex/ChatGPT-OAuth backend accepts the `input_image` content block
@@ -713,6 +735,7 @@ export class CodexClient implements AIProvider {
             : null,
           store: false,
           stream: true,
+          ...(cacheKey ? { prompt_cache_key: cacheKey } : {}),
           // Required field; empty array when not asking for reasoning. With
           // reasoning on, the encrypted items come back so the next tool
           // round can hand them over statelessly (`store` must stay false).
diff --git a/src/lib/ai/local-client.ts b/src/lib/ai/local-client.ts
index 3c3a5f184..d4ea0fbfb 100644
--- a/src/lib/ai/local-client.ts
+++ b/src/lib/ai/local-client.ts
@@ -11,6 +11,7 @@ import type {
 import {
   buildOpenAIMessages,
   mapFinishReason,
+  parseCachedTokens,
   raiseEmbeddedError,
   readEmbeddedError,
   sanitiseBodyExcerpt,
@@ -411,6 +412,7 @@ export class LocalOpenAICompatibleClient implements AIProvider {
       }>;
       usage?: {
         total_tokens?: number;
+        prompt_tokens_details?: { cached_tokens?: number };
         completion_tokens_details?: { reasoning_tokens?: number };
       };
     };
@@ -459,6 +461,11 @@ export class LocalOpenAICompatibleClient implements AIProvider {
     return {
       content,
       tokensUsed: json.usage?.total_tokens ?? null,
+      // v1.41.2 — llama.cpp, vLLM and the hosted gateways this client also
+      // serves report a prompt-cache hit the Chat Completions way; without it
+      // the Coach turn budget counted a cached prefix in full. Ollama sends
+      // none, which leaves it null.
+      cachedInputTokens: parseCachedTokens(json),
       model: this.config.model,
       providerType: "local",
       // v1.28.28 (#470) — surface why the model stopped so the JSON callers
@@ -543,7 +550,10 @@ export class LocalOpenAICompatibleClient implements AIProvider {
           message?: { content?: string };
           finish_reason?: string;
         }>;
-        usage?: { total_tokens?: number };
+        usage?: {
+          total_tokens?: number;
+          prompt_tokens_details?: { cached_tokens?: number };
+        };
       } | null;
       // The server ignored `stream` and answered buffered — the embedded
       // failure can ride this body exactly as it rides the non-streaming one.
@@ -577,6 +587,7 @@ export class LocalOpenAICompatibleClient implements AIProvider {
       return {
         content: buffered,
         tokensUsed: json?.usage?.total_tokens ?? null,
+        cachedInputTokens: json ? parseCachedTokens(json) : null,
         model: this.config.model,
         providerType: "local",
         finishReason: mapFinishReason(json?.choices?.[0]?.finish_reason),
@@ -589,6 +600,7 @@ export class LocalOpenAICompatibleClient implements AIProvider {
     let sseBuffer = "";
     let content = "";
     let tokensUsed: number | null = null;
+    let cachedInputTokens: number | null = null;
     let reasoningTokens: number | null = null;
     let finishReason: string | undefined;
 
@@ -679,6 +691,7 @@ export class LocalOpenAICompatibleClient implements AIProvider {
             }>;
             usage?: {
               total_tokens?: number;
+              prompt_tokens_details?: { cached_tokens?: number };
               completion_tokens_details?: { reasoning_tokens?: number };
             };
           };
@@ -702,6 +715,8 @@ export class LocalOpenAICompatibleClient implements AIProvider {
           if (typeof chunk.usage?.total_tokens === "number") {
             tokensUsed = chunk.usage.total_tokens;
           }
+          const cached = chunk.usage?.prompt_tokens_details?.cached_tokens;
+          if (typeof cached === "number") cachedInputTokens = cached;
           const rt = chunk.usage?.completion_tokens_details?.reasoning_tokens;
           if (typeof rt === "number") reasoningTokens = rt;
           const chunkFinish = chunk.choices?.[0]?.finish_reason;
@@ -761,6 +776,7 @@ export class LocalOpenAICompatibleClient implements AIProvider {
     return {
       content,
       tokensUsed,
+      cachedInputTokens,
       model: this.config.model,
       providerType: "local",
       finishReason: mapFinishReason(finishReason),
diff --git a/src/lib/ai/openai-capabilities.ts b/src/lib/ai/openai-capabilities.ts
index 75b723a8e..1766ec05d 100644
--- a/src/lib/ai/openai-capabilities.ts
+++ b/src/lib/ai/openai-capabilities.ts
@@ -34,7 +34,7 @@ function isModernCanonicalModel(model: string): boolean {
   );
 }
 
-function isCanonicalOpenAIEndpoint(baseUrl: string): boolean {
+export function isCanonicalOpenAIEndpoint(baseUrl: string): boolean {
   try {
     const url = new URL(baseUrl);
     return (
diff --git a/src/lib/ai/openai-client.ts b/src/lib/ai/openai-client.ts
index 227613d67..46c084158 100644
--- a/src/lib/ai/openai-client.ts
+++ b/src/lib/ai/openai-client.ts
@@ -3,6 +3,7 @@ import { callTimeoutMs } from "./effective-timeout";
 import { annotate } from "@/lib/logging/context";
 import type { AIProvider, CompletionParams, CompletionResult } from "./types";
 import {
+  isCanonicalOpenAIEndpoint,
   selectOpenAIChatCompletionsCapabilities,
   type OpenAIChatCompletionsCapabilities,
 } from "./openai-capabilities";
@@ -208,6 +209,13 @@ export class OpenAIClient implements AIProvider {
             : {}),
           ...(tools ? { tools } : {}),
           ...(params.toolChoice ? { tool_choice: params.toolChoice } : {}),
+          // v1.41.2 — the prompt-cache routing key `api.openai.com` documents
+          // (https://platform.openai.com/docs/guides/prompt-caching). Only
+          // there: a gateway or proxy behind another base URL may refuse a
+          // field it does not know.
+          ...(params.cacheKey && isCanonicalOpenAIEndpoint(this.config.baseUrl)
+            ? { prompt_cache_key: params.cacheKey }
+            : {}),
           // #1126 — the gateway entry's reasoning setting; Default sends
           // nothing. `api.openai.com` and Codex never get it. v1.41 — a call
           // that carries its own resolved level (`params.reasoning`) wins.
diff --git a/src/lib/ai/openai-wire.ts b/src/lib/ai/openai-wire.ts
index 673c128a7..f7ebb3ff1 100644
--- a/src/lib/ai/openai-wire.ts
+++ b/src/lib/ai/openai-wire.ts
@@ -150,7 +150,9 @@ export function parseOpenAIToolCalls(
   }));
 }
 
-export function parseCachedTokens(json: OpenAIResponseJson): number | null {
+export function parseCachedTokens(json: {
+  usage?: { prompt_tokens_details?: { cached_tokens?: number } };
+}): number | null {
   return json.usage?.prompt_tokens_details?.cached_tokens ?? null;
 }
 
diff --git a/src/lib/ai/types.ts b/src/lib/ai/types.ts
index c33949131..be66d3c42 100644
--- a/src/lib/ai/types.ts
+++ b/src/lib/ai/types.ts
@@ -374,6 +374,16 @@ export interface CompletionParams {
    * client sends exactly what it sent before the field existed.
    */
   reasoning?: { effort: ReasoningLevel; summaries: boolean };
+  /**
+   * v1.41.2 — a stable key for calls that resend the same prompt prefix (the
+   * rounds of one Coach turn, the turns of one conversation). Providers that
+   * route a request to a prompt cache by key read it: Codex as
+   * `prompt_cache_key` and its session headers, `api.openai.com` as
+   * `prompt_cache_key`. Without it a Codex call is routed afresh every time
+   * and a later round pays for its whole prefix again. An opaque value: the
+   * caller hashes whatever it derives the key from. Absent → nothing is sent.
+   */
+  cacheKey?: string;
   /**
    * v1.41 — live reasoning events, called as they arrive on clients that read
    * the provider's stream, or once after the reply on clients that do not.
@@ -390,6 +400,12 @@ export interface CompletionResult {
    * from the provider's cache, where the provider reports it (Anthropic
    * `cache_read_input_tokens`, OpenAI/Codex `cached_tokens`). Null / absent
    * when the provider does not report it or the prefix did not hit.
+   *
+   * v1.41.2 — always a share of `tokensUsed`, never beside it: the turn
+   * budget and the daily ledger charge `tokensUsed - cachedInputTokens` at
+   * full weight and the cached share at a tenth. A client whose provider
+   * reports cached input outside its input count (Anthropic) adds it to
+   * `tokensUsed`.
    */
   cachedInputTokens?: number | null;
   model: string;