fix(ollama): report prompt_eval_cached_count as cache-read input

scripts/test-layout/layout.json

@@ -1330,7 +1330,7 @@
     "ocx-launcher-desktop-handoff.test.ts": "cli", "cli-path-diagnostics.test.ts": "cli",
     "ocx-launcher-source.test.ts": "cli",
     "ocx-run.test.ts": "cli",
-    "ollama-native-parser.test.ts": "providers/ollama", "ollama-native-reasoning-wire.test.ts": "providers/ollama", "ollama-native-structured-output.test.ts": "providers/ollama", "ollama-native-tool-continuations.test.ts": "providers/ollama",
+    "ollama-native-parser.test.ts": "providers/ollama", "ollama-native-cached-usage.test.ts": "providers/ollama", "ollama-native-reasoning-wire.test.ts": "providers/ollama", "ollama-native-structured-output.test.ts": "providers/ollama", "ollama-native-tool-continuations.test.ts": "providers/ollama",
     "ollama-native-v4.test.ts": "providers/ollama", "ollama-native-scaling.test.ts": "providers/ollama", "ollama-native.test.ts": "providers/ollama", "ollama-show-enrichment-v7.test.ts": "providers/ollama", "ollama-show-enrichment.test.ts": "providers/ollama",
     "ollama-show-ignore-abort.test.ts": "providers/ollama",
     "opaque-blob-wrapped-rejection.test.ts": "responses", "openai-api-virtual-models.test.ts": "adapters/openai", "openai-chat-bounded-tool-names.test.ts": "adapters/openai", "openai-chat-dangling-toolcalls.test.ts": "adapters/openai",

src/adapters/ollama-native.ts

@@ -641,7 +641,17 @@ function usageFromNative(value: JsonRecord | undefined): OcxUsage | undefined {
   const input = isFiniteNonNegativeInteger(value.prompt_eval_count) ? value.prompt_eval_count : undefined;
   const output = isFiniteNonNegativeInteger(value.eval_count) ? value.eval_count : undefined;
   if (input === undefined && output === undefined) return undefined;
-  return { inputTokens: input ?? 0, outputTokens: output ?? 0 };
+  // Ollama reports prompt_eval_cached_count as the cached subset of prompt_eval_count (api/types.go;
+  // its CLI prints prompt_eval_count - cached as the uncached share), which is OCX's cache-read
+  // convention. Omitted means unknown, so no cache field is invented; a count that exceeds the
+  // prompt is not a subset and is ignored rather than clamped (#8101).
+  const cached = input !== undefined && isFiniteNonNegativeInteger(value.prompt_eval_cached_count)
+    && value.prompt_eval_cached_count <= input ? value.prompt_eval_cached_count : undefined;
+  return {
+    inputTokens: input ?? 0,
+    outputTokens: output ?? 0,
+    ...(cached !== undefined ? { cachedInputTokens: cached, cacheReadInputTokens: cached } : {}),
+  };
 }
 
 function stopReasonFromNative(value: unknown): string | undefined {

tests/fixtures/test-layout-expected.json

@@ -1092,7 +1092,7 @@
   "oauth-tos-warning.test.ts": "gui", "oauth-upsert-preserves-api-key.test.ts": "oauth",
   "ocx-launcher-runtime.test.ts": "cli", "ocx-launcher-fixture-trust.test.ts": "cli", "ocx-launcher-source.test.ts": "cli",
   "ocx-launcher-desktop-handoff.test.ts": "cli", "cli-path-diagnostics.test.ts": "cli",
-  "ocx-run.test.ts": "cli", "ollama-native-parser.test.ts": "providers/ollama",
+  "ocx-run.test.ts": "cli", "ollama-native-parser.test.ts": "providers/ollama", "ollama-native-cached-usage.test.ts": "providers/ollama",
   "ollama-native-reasoning-wire.test.ts": "providers/ollama",
   "ollama-native-scaling.test.ts": "providers/ollama",
   "ollama-native-structured-output.test.ts": "providers/ollama", "ollama-native-v4.test.ts": "providers/ollama",

tests/providers/ollama/ollama-native-cached-usage.test.ts

@@ -0,0 +1,53 @@
+import { describe, expect, test } from "bun:test";
+import { createOllamaNativeAdapter } from "../../../src/adapters/ollama-native";
+import type { AdapterEvent } from "../../../src/types";
+import { createTestTranslatorBudget } from "../../helpers/translator-budget";
+
+// #8101: the native usage parser dropped prompt_eval_cached_count, so cached input always read 0.
+const adapter = () => createOllamaNativeAdapter({ adapter: "ollama-native", baseUrl: "http://127.0.0.1:11434", models: ["m"] });
+const ndjson = (frames: unknown[]) => new Response(frames.map(frame => JSON.stringify(frame)).join("\n") + "\n",
+  { headers: { "content-type": "application/x-ndjson" } });
+const final = (usage: Record<string, unknown>) => ({
+  model: "m", message: { role: "assistant", content: "hi" }, done: true, done_reason: "stop", ...usage,
+});
+
+async function streamedUsage(usage: Record<string, unknown>) {
+  const events: AdapterEvent[] = [];
+  for await (const event of adapter().parseStream(ndjson([final(usage)]), createTestTranslatorBudget())) events.push(event);
+  const done = events.at(-1);
+  expect(done?.type).toBe("done");
+  return (done as Extract<AdapterEvent, { type: "done" }>).usage;
+}
+
+describe("ollama-native cached prompt usage (#8101)", () => {
+  test("streamed final record maps the cached subset without changing the input total", async () => {
+    expect(await streamedUsage({ prompt_eval_count: 2000, prompt_eval_cached_count: 1500, eval_count: 100 }))
+      .toEqual({ inputTokens: 2000, outputTokens: 100, cachedInputTokens: 1500, cacheReadInputTokens: 1500 });
+  });
+
+  test("buffered responses carry the same cached count", async () => {
+    const events = await adapter().parseResponse!(ndjson([final({ prompt_eval_count: 11, prompt_eval_cached_count: 8, eval_count: 18 })]),
+      createTestTranslatorBudget());
+    expect(events.at(-1)).toMatchObject({ type: "done", usage: { inputTokens: 11, cachedInputTokens: 8, cacheReadInputTokens: 8 } });
+  });
+
+  test("an explicit zero stays a known zero", async () => {
+    expect(await streamedUsage({ prompt_eval_count: 40, prompt_eval_cached_count: 0, eval_count: 2 }))
+      .toEqual({ inputTokens: 40, outputTokens: 2, cachedInputTokens: 0, cacheReadInputTokens: 0 });
+  });
+
+  test.each([
+    ["omitted", {}],
+    ["negative", { prompt_eval_cached_count: -1 }],
+    ["fractional", { prompt_eval_cached_count: 1.5 }],
+    ["non-numeric", { prompt_eval_cached_count: "8" }],
+    ["larger than the prompt", { prompt_eval_cached_count: 41 }],
+  ])("%s cached count reports no cache field", async (_label, extra) => {
+    expect(await streamedUsage({ prompt_eval_count: 40, eval_count: 2, ...extra })).toEqual({ inputTokens: 40, outputTokens: 2 });
+  });
+
+  test("a cached count without a prompt count is ignored", async () => {
+    expect(await streamedUsage({ prompt_eval_cached_count: 5, eval_count: 2 })).toEqual({ inputTokens: 0, outputTokens: 2 });
+  });
+});
+