Accept independent Grok reasoning counters in prompt receipts

Co-Authored-By: Paperclip <noreply@paperclip.ing>

packages/paperclip-runner/src/drivers/acpx/grok-usage.test.ts

@@ -21,12 +21,29 @@ describe("Grok terminal prompt usage", () => {
 
   it.each([
     { usageIsIncomplete: true }, { usageIsIncomplete: "false" }, { cachedReadTokens: 505796 },
-    { inputTokens: -1 }, { outputTokens: 1.5 }, { reasoningTokens: 5764 }, { totalTokens: 1 },
+    { inputTokens: -1 }, { outputTokens: 1.5 }, { reasoningTokens: -1 }, { reasoningTokens: 1.5 }, { totalTokens: 1 },
     { modelCalls: undefined }, { inputTokens: Number.MAX_SAFE_INTEGER + 1 },
   ])("refuses incomplete or invalid token authority: %j", patch => {
     expect(grokPromptUsage({ ...meta, usage: { ...usage, ...patch } }, "session-1")).toBeNull();
   });
 
+  it("settles a real Grok prompt whose reasoning counter exceeds its output counter", () => {
+    const collector = createGrokPromptUsageCollector();
+    collector.begin("runner-1", "session-1");
+    collector.observe("outbound", request);
+    collector.observe("inbound", { ...response, result: { _meta: { ...meta, usage: {
+      inputTokens: 517350, outputTokens: 7969, totalTokens: 525319, cachedReadTokens: 458624,
+      cacheCreationTokens: 0, reasoningTokens: 9859, modelCalls: 15, costUsdTicks: 4537320000,
+    } } } });
+    const state = { lastRequestId: "runner-1", terminalPromptUsage: collector.snapshot() };
+    expect(persistedAcpxTurnUsage({}, state, "runner-1")).toEqual({
+      type: "status", tag: "usage_update", text: "terminal prompt usage",
+      cost: { amount: 0.453732, currency: "USD", basis: "turn" },
+      breakdown: { inputTokens: 58726, outputTokens: 7969, totalTokens: 525319,
+        cachedReadTokens: 458624, cachedWriteTokens: 0, thoughtTokens: 0 },
+    });
+  });
+
   it.each([{ costUsdTicks: undefined }, { costUsdTicks: -1 }, { costIsPartial: true }, { costIsPartial: "false" }])(
     "preserves complete tokens without inventing cost: %j", patch => {
       const result = grokPromptUsage({ ...meta, usage: { ...usage, ...patch } }, "session-1");

packages/paperclip-runner/src/drivers/acpx/grok-usage.ts

@@ -16,8 +16,7 @@ export function grokPromptUsage(metaValue: unknown, sessionId: string): Omit<Ter
   const { inputTokens, outputTokens, totalTokens, cachedReadTokens, cacheCreationTokens, reasoningTokens, modelCalls } = usage;
   if (![inputTokens, outputTokens, totalTokens, cachedReadTokens, cacheCreationTokens, reasoningTokens, modelCalls].every(integer)
     || Number(inputTokens) + Number(outputTokens) !== totalTokens
-    || Number(cachedReadTokens) + Number(cacheCreationTokens) > Number(inputTokens)
-    || Number(reasoningTokens) > Number(outputTokens)) return null;
+    || Number(cachedReadTokens) + Number(cacheCreationTokens) > Number(inputTokens)) return null;
   const costKnown = integer(usage.costUsdTicks) && (usage.costIsPartial === undefined || usage.costIsPartial === false);
   return {
     promptId: meta.promptId,

tests/runner-e2e/README.md

@@ -2188,6 +2188,12 @@ an API-key pass must not be reported as subscription evidence. See
 [`doc/grok-native-runner.md`](../../doc/grok-native-runner.md) for installation,
 credential boundaries and repetition requirements.
 
+Grok accounting uses the complete prompt ledger bound to the matching ACP
+response. Its reasoning counter can exceed its output counter; it is not a
+validation bound or an extra term in the ledger's input-plus-output total.
+Reported cost remains authoritative. See the upstream
+[`UsageTotals` ledger](https://github.com/xai-org/grok-build/blob/2bdd1d6a6369de0e8c68132ea4539e9abd9e14a8/crates/codegen/xai-chat-state/src/usage.rs).
+
 ### Kimi Code qualification
 
 `kimi-qualification` is a separate explicit-only candidate suite for maintained