diff --git a/src/llm/llama-server-client.test.ts b/src/llm/llama-server-client.test.ts
index 0bfcc8d8..ff0e46e9 100644
--- a/src/llm/llama-server-client.test.ts
+++ b/src/llm/llama-server-client.test.ts
@@ -50,7 +50,10 @@ describe("LlamaServerClient.complete", () => {
 
     expect(result.content).toBe('{"tool":"finish","args":{}}');
     expect(result.reasoningContent).toBe("");
-    expect(result.timing.promptTokens).toBe(40);
+    // 40 evaluated this request + 30 reused from the KV cache: the
+    // prompt the model saw was 70 tokens, and that is what occupancy
+    // consumers (the TUI context chip among them) need reported.
+    expect(result.timing.promptTokens).toBe(70);
     expect(result.cacheHitTokens).toBe(30);
     expect(result.slotId).toBe(2);
     expect(result.modelId).toBe("qwen-test");
@@ -65,6 +68,29 @@ describe("LlamaServerClient.complete", () => {
     expect(snapshot.body.repeat_last_n).toBe(256);
   });
 
+  it("reports the bare evaluated count when nothing was cached", async () => {
+    const client = new LlamaServerClient({
+      baseUrl: "http://127.0.0.1:9999",
+      fetchImpl: createMockFetch(async () =>
+        new Response(
+          JSON.stringify({
+            content: "ok",
+            stop: true,
+            truncated: false,
+            timings: { prompt_ms: 10, predicted_ms: 20, prompt_n: 40, predicted_n: 8 },
+            slot_id: 0,
+          }),
+          { status: 200, headers: { "content-type": "application/json" } },
+        ),
+      ),
+    });
+
+    const result = await client.complete({ prompt: "hello", maxTokens: 16 });
+
+    expect(result.timing.promptTokens).toBe(40);
+    expect(result.cacheHitTokens).toBe(0);
+  });
+
   it("forwards explicit repeatPenalty / repeatLastN overrides", async () => {
     let captured: Record<string, unknown> | null = null;
     const client = new LlamaServerClient({
diff --git a/src/llm/llama-server-client.ts b/src/llm/llama-server-client.ts
index 3abe229a..f2d8c2e4 100644
--- a/src/llm/llama-server-client.ts
+++ b/src/llm/llama-server-client.ts
@@ -551,6 +551,17 @@ function normaliseCompletionResponse(
   payload: Record<string, unknown>,
 ): CompletionResult {
   const timings = (payload.timings ?? {}) as Record<string, unknown>;
+  // `prompt_n` / `tokens_evaluated` count only the tokens llama-server
+  // actually evaluated this request — the prefix reused from the KV
+  // cache (`tokens_cached`) is excluded. Every consumer of
+  // `timing.promptTokens` treats it as "how big was the prompt" (their
+  // fallback is `prompt.tokens.total`, and the TUI shows it as occupied
+  // context), so report the whole prompt: evaluated + cached. On a warm
+  // cache the raw `prompt_n` is a small fraction of the prompt and the
+  // context readout collapsed to it, then leapt back to the estimator's
+  // full figure the moment anything reprojected it.
+  const evaluatedTokens = toNumber(timings.prompt_n ?? payload.tokens_evaluated);
+  const cachedTokens = toNumber(payload.tokens_cached);
   return {
     content: typeof payload.content === "string" ? payload.content : "",
     reasoningContent:
@@ -562,10 +573,10 @@ function normaliseCompletionResponse(
     timing: {
       promptMs: toNumber(timings.prompt_ms),
       predictedMs: toNumber(timings.predicted_ms),
-      promptTokens: toNumber(timings.prompt_n ?? payload.tokens_evaluated),
+      promptTokens: evaluatedTokens + cachedTokens,
       predictedTokens: toNumber(timings.predicted_n ?? payload.tokens_predicted),
     },
-    cacheHitTokens: toNumber(payload.tokens_cached),
+    cacheHitTokens: cachedTokens,
     slotId: toNumber(payload.slot_id ?? payload.id_slot, -1),
     modelId:
       typeof payload.model === "string" ? payload.model : null,
diff --git a/src/tui/agent-event-reducer.ts b/src/tui/agent-event-reducer.ts
index 2f8baaf6..57d2f01f 100644
--- a/src/tui/agent-event-reducer.ts
+++ b/src/tui/agent-event-reducer.ts
@@ -716,9 +716,11 @@ function reduceStepEvent(
     case "llm_completed": {
       // `prompt_built` carried an estimate (`estimateTokens` over-counts
       // by design); the provider just reported what its own tokenizer
-      // actually counted — llama.cpp from `tokens_evaluated`, an
-      // OpenAI-compatible cloud from `usage.prompt_tokens`. Prefer it,
-      // and leave the estimate standing when nothing was reported.
+      // actually counted — llama.cpp from `prompt_n + tokens_cached`
+      // (the whole prompt, not just the slice evaluated past the KV
+      // cache), an OpenAI-compatible cloud from `usage.prompt_tokens`.
+      // Prefer it, and leave the estimate standing when nothing was
+      // reported.
       const counted = event.completion.timing?.promptTokens ?? 0;
       if (counted <= 0) return state;
       return {