Merge pull request #31 from IfkumRfnl/fix/cache-accounting

fix(stream): correct cached input accounting

CHANGELOG.md

@@ -1,5 +1,9 @@
 # Changelog
 
+## 0.4.4 - 2026-08-03
+
+- Fix cached input tokens being counted twice.
+
 ## 0.4.3 - 2026-08-02
 
 - Allow pi to start when model discovery is unavailable. The provider now caches the last successfully fetched model catalog so previously discovered Command Code models remain selectable offline; a first offline start without a cache keeps Command Code unavailable until `/reload` succeeds.

src/core.ts

@@ -388,10 +388,14 @@ export function createStreamCommandCode(deps: CoreDependencies) {
             const usage = commandCodeUsage(event)
             if (usage) {
               const details = commandCodeInputTokenDetails(usage)
-              output.usage.input = numberValue(usage.inputTokens) ?? 0
+              const totalInput = numberValue(usage.inputTokens) ?? 0
+              const input = numberValue(details?.noCacheTokens)
+              const cacheRead = numberValue(details?.cacheReadTokens) ?? 0
+              const cacheWrite = numberValue(details?.cacheWriteTokens) ?? 0
+              output.usage.input = input ?? Math.max(0, totalInput - cacheRead - cacheWrite)
               output.usage.output = numberValue(usage.outputTokens) ?? 0
-              output.usage.cacheRead = numberValue(details?.cacheReadTokens) ?? 0
+              output.usage.cacheRead = cacheRead
-              output.usage.cacheWrite = numberValue(details?.cacheWriteTokens) ?? 0
+              output.usage.cacheWrite = cacheWrite
               output.usage.totalTokens =
                 output.usage.input +
                 output.usage.output +

tests/test-stream.ts

@@ -102,9 +102,9 @@ describe("streamCommandCode — successful streams", () => {
           type: "finish",
           finishReason: "stop",
           totalUsage: {
-            inputTokens: 5,
+            inputTokens: 3124,
-            outputTokens: 2,
+            outputTokens: 15,
-            inputTokenDetails: { cacheReadTokens: 3, cacheWriteTokens: 1 },
+            inputTokenDetails: { noCacheTokens: 52, cacheReadTokens: 3072 },
           },
         }),
       ],
@@ -134,10 +134,77 @@ describe("streamCommandCode — successful streams", () => {
       done.message.content[0]?.type === "text" ? done.message.content[0].text : "",
       "Hello",
     )
-    assert.equal(done.message.usage.totalTokens, 11)
+    assert.equal(done.message.usage.input, 52)
+    assert.equal(done.message.usage.cacheRead, 3072)
+    assert.equal(done.message.usage.cacheWrite, 0)
+    assert.equal(done.message.usage.totalTokens, 3139)
     assert.equal(calculatedUsages.length, 1)
   })
 
+  it("derives uncached input when noCacheTokens is missing", async () => {
+    server.mockResponse({
+      type: "success",
+      events: [
+        JSON.stringify({
+          type: "finish",
+          finishReason: "stop",
+          totalUsage: {
+            inputTokens: 100,
+            outputTokens: 10,
+            inputTokenDetails: { cacheReadTokens: 75 },
+          },
+        }),
+      ],
+    })
+    const { streamCommandCode } = createTestDeps({ apiBase: server.baseUrl() })
+
+    const events = await collectEvents(
+      streamCommandCode(makeModel(), makeContext(), { apiKey: "mock-key" }),
+    )
+
+    const done = events.at(-1)
+    assert.equal(done?.type, "done")
+    if (done?.type !== "done") throw new Error("expected done")
+    assert.equal(done.message.usage.input, 25)
+    assert.equal(done.message.usage.cacheRead, 75)
+    assert.equal(done.message.usage.cacheWrite, 0)
+    assert.equal(done.message.usage.totalTokens, 110)
+  })
+
+  it("accounts for cache writes separately from uncached input", async () => {
+    server.mockResponse({
+      type: "success",
+      events: [
+        JSON.stringify({
+          type: "finish",
+          finishReason: "stop",
+          totalUsage: {
+            inputTokens: 100,
+            outputTokens: 10,
+            inputTokenDetails: {
+              noCacheTokens: 20,
+              cacheReadTokens: 70,
+              cacheWriteTokens: 10,
+            },
+          },
+        }),
+      ],
+    })
+    const { streamCommandCode } = createTestDeps({ apiBase: server.baseUrl() })
+
+    const events = await collectEvents(
+      streamCommandCode(makeModel(), makeContext(), { apiKey: "mock-key" }),
+    )
+
+    const done = events.at(-1)
+    assert.equal(done?.type, "done")
+    if (done?.type !== "done") throw new Error("expected done")
+    assert.equal(done.message.usage.input, 20)
+    assert.equal(done.message.usage.cacheRead, 70)
+    assert.equal(done.message.usage.cacheWrite, 10)
+    assert.equal(done.message.usage.totalTokens, 110)
+  })
+
   it("ends on finish without waiting for an open upstream connection", async () => {
     server.mockResponse({
       type: "success",