feat(harness): stop sharing agent $ cost with the client

miot-harness/src/miot_harness/observability/callbacks.py

@@ -216,8 +216,9 @@ def on_llm_end(
                 "cache_read_input_tokens": usage.cache_read_input_tokens,
                 "cache_creation_input_tokens": usage.cache_creation_input_tokens,
             }
-            if cost_usd is not None:
-                data["cost_usd"] = cost_usd
+            # NOTE: cost_usd is deliberately NOT emitted to the client. The dollar
+            # cost stays server-side (span attribute above / eval + provenance logs)
+            # so we can track our own spend without exposing it to the tenant.
             # Fail open: the progress sink is an observability concern,
             # so a buggy or saturated event bus must NOT bubble up and
             # tear down the LLM call. Catch Exception (not BaseException)

miot-harness/tests/observability/test_usage_recorded_event.py

@@ -56,6 +56,8 @@ def test_callback_emits_usage_recorded_when_progress_wired(
     usage_events = [e for e in events if e.type == "usage.recorded"]
     assert len(usage_events) == 1
     payload = usage_events[0].data
+    # The dollar cost is NEVER shared with the client, even for a priced
+    # model — it stays server-side (span attribute + eval/provenance logs).
     # Total input (200) minus cache_read (5) minus cache_creation (3) = 192
     assert payload == {
         "agent": "filter_expert",
@@ -64,9 +66,8 @@ def test_callback_emits_usage_recorded_when_progress_wired(
         "output_tokens": 80,
         "cache_read_input_tokens": 5,
         "cache_creation_input_tokens": 3,
-        "cost_usd": payload["cost_usd"],
     }
-    assert payload["cost_usd"] > 0
+    assert "cost_usd" not in payload
     assert usage_events[0].run_id == "run_abc"
 
 

turbo-repo/packages/miot-chat/src/tui/__tests__/components/FooterLine.test.tsx

@@ -9,9 +9,7 @@ function usage(partial: Partial<UsageTotals> = {}): UsageTotals {
     outputTokens: 0,
     cacheReadTokens: 0,
     cacheCreationTokens: 0,
-    costUsd: 0,
     lastAgent: null,
-    lastCostUsd: null,
     ...partial,
   };
 }
@@ -44,23 +42,22 @@ describe("<FooterLine />", () => {
     expect(lastFrame() ?? "").toContain("profile staging");
   });
 
-  it("shows a usage segment when usageTotals has any tokens", () => {
+  it("shows a token usage segment when usageTotals has any tokens", () => {
     const { lastFrame } = render(
       <FooterLine
         {...props({
           usageTotals: usage({
             inputTokens: 1234,
             outputTokens: 56,
-            costUsd: 0.0123,
             lastAgent: "synthesizer",
-            lastCostUsd: 0.0123,
           }),
         })}
       />,
     );
     const frame = lastFrame() ?? "";
     expect(frame).toContain("1234→56");
-    expect(frame).toContain("$0.0123");
+    // The dollar cost is never shown to the client.
+    expect(frame).not.toContain("$");
   });
 
   it("hides the usage segment when there are no tokens yet", () => {

turbo-repo/packages/miot-chat/src/tui/__tests__/transcript.project.test.ts

@@ -30,9 +30,7 @@ function emptySlice(): TranscriptSlice {
       outputTokens: 0,
       cacheReadTokens: 0,
       cacheCreationTokens: 0,
-      costUsd: 0,
       lastAgent: null,
-      lastCostUsd: null,
     },
   };
 }
@@ -434,7 +432,7 @@ describe("transcript projector — thinking + usage (plan: SSE rich events)", ()
     expect(s2.transcript[0]).toMatchObject({ kind: "thinking", status: "complete" });
   });
 
-  it("usage.recorded accumulates totals across calls", () => {
+  it("usage.recorded accumulates token totals across calls", () => {
     const ctx = mkCtx();
     const s1 = applyHarnessEvent(
       emptySlice(),
@@ -446,7 +444,6 @@ describe("transcript projector — thinking + usage (plan: SSE rich events)", ()
           output_tokens: 100,
           cache_read_input_tokens: 50,
           cache_creation_input_tokens: 25,
-          cost_usd: 0.0042,
         },
       }),
       "r1",
@@ -457,9 +454,7 @@ describe("transcript projector — thinking + usage (plan: SSE rich events)", ()
       outputTokens: 100,
       cacheReadTokens: 50,
       cacheCreationTokens: 25,
-      costUsd: 0.0042,
       lastAgent: "filter_expert",
-      lastCostUsd: 0.0042,
     });
 
     const s2 = applyHarnessEvent(
@@ -480,8 +475,9 @@ describe("transcript projector — thinking + usage (plan: SSE rich events)", ()
     expect(s2.usageTotals.inputTokens).toBe(3000);
     expect(s2.usageTotals.outputTokens).toBe(300);
     expect(s2.usageTotals.lastAgent).toBe("synthesizer");
-    expect(s2.usageTotals.lastCostUsd).toBeNull();
-    expect(s2.usageTotals.costUsd).toBeCloseTo(0.0042);
+    // Dollar cost is never carried on the client-side usage totals.
+    expect(s2.usageTotals).not.toHaveProperty("costUsd");
+    expect(s2.usageTotals).not.toHaveProperty("lastCostUsd");
   });
 
   it("agent.completed is intentionally dropped from the transcript", () => {

turbo-repo/packages/miot-chat/src/tui/chrome/FooterLine.tsx

@@ -22,8 +22,7 @@ export function FooterLine(props: FooterLineProps): React.ReactElement {
   segments.push(`ctx≈${props.approxTokens}tok (${props.contextPercent}%)`);
   const u = props.usageTotals;
   if (u && (u.inputTokens > 0 || u.outputTokens > 0)) {
-    const cost = u.costUsd > 0 ? ` $${u.costUsd.toFixed(4)}` : "";
-    segments.push(`${u.inputTokens}→${u.outputTokens}${cost}`);
+    segments.push(`${u.inputTokens}→${u.outputTokens}`);
   }
   segments.push(props.baseUrl);
   if (props.profileName) {

turbo-repo/packages/miot-chat/src/tui/session/types.ts

@@ -43,27 +43,23 @@ export type TranscriptItem =
 
 /**
  * Running totals across the conversation. Reset by CLEAR /
- * RESET_CONVERSATION; accumulated turn-over-turn. `lastCostUsd` is
- * the cost of the most recent LLM call (chip footer uses it).
+ * RESET_CONVERSATION; accumulated turn-over-turn. Dollar cost is
+ * intentionally not tracked here — it is not shared with the client.
  */
 export interface UsageTotals {
   inputTokens: number;
   outputTokens: number;
   cacheReadTokens: number;
   cacheCreationTokens: number;
-  costUsd: number;
   lastAgent: string | null;
-  lastCostUsd: number | null;
 }
 
 export const ZERO_USAGE: UsageTotals = {
   inputTokens: 0,
   outputTokens: 0,
   cacheReadTokens: 0,
   cacheCreationTokens: 0,
-  costUsd: 0,
   lastAgent: null,
-  lastCostUsd: null,
 };
 
 export interface PendingApproval {

turbo-repo/packages/miot-chat/src/tui/transcript/project.ts

@@ -193,8 +193,6 @@ export function applyHarnessEvent(
         typeof event.data.cache_creation_input_tokens === "number"
           ? event.data.cache_creation_input_tokens
           : 0;
-      const cost =
-        typeof event.data.cost_usd === "number" ? event.data.cost_usd : 0;
       const agent =
         typeof event.data.agent === "string" ? event.data.agent : null;
       return {
@@ -204,10 +202,7 @@ export function applyHarnessEvent(
           outputTokens: slice.usageTotals.outputTokens + outT,
           cacheReadTokens: slice.usageTotals.cacheReadTokens + cacheR,
           cacheCreationTokens: slice.usageTotals.cacheCreationTokens + cacheC,
-          costUsd: slice.usageTotals.costUsd + cost,
           lastAgent: agent,
-          lastCostUsd:
-            typeof event.data.cost_usd === "number" ? event.data.cost_usd : null,
         },
       };
     }

turbo-repo/packages/miot-harness-client/src/types.ts

@@ -119,7 +119,6 @@ export interface UsageRecordedData {
   output_tokens: number;
   cache_read_input_tokens: number;
   cache_creation_input_tokens: number;
-  cost_usd?: number;
 }
 
 export const TERMINAL_EVENT_TYPES = new Set<HarnessEventType>([