shumengya/sproutclaw
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(ai): preserve OpenRouter cached token semantics

Browse Source
This commit is contained in:
Armin Ronacher
2026-05-16 11:50:24 +02:00
parent b4ee3aaeea
commit 87881ca686
2 changed files with 17 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
packages/ai/src/providers/openai-completions.ts
View File

@@ -997,17 +997,17 @@ function parseChunkUsage(
model: Model<"openai-completions">, model: Model<"openai-completions">,
): AssistantMessage["usage"] { ): AssistantMessage["usage"] {
const promptTokens = rawUsage.prompt_tokens || 0; const promptTokens = rawUsage.prompt_tokens || 0;
const reportedCachedTokens = rawUsage.prompt_tokens_details?.cached_tokens ?? rawUsage.prompt_cache_hit_tokens ?? 0; const cacheReadTokens = rawUsage.prompt_tokens_details?.cached_tokens ?? rawUsage.prompt_cache_hit_tokens ?? 0;
const cacheWriteTokens = rawUsage.prompt_tokens_details?.cache_write_tokens || 0; const cacheWriteTokens = rawUsage.prompt_tokens_details?.cache_write_tokens || 0;
// Normalize to pi-ai semantics: // Follow documented OpenAI/OpenRouter semantics: cached_tokens is cache-read
// - cacheRead: hits from cache created by previous requests only // tokens (hits). OpenAI does not document or emit cache_write_tokens, but
// - cacheWrite: tokens written to cache in this request // OpenRouter-compatible providers can include it as a separate write count.
// Some OpenAI-compatible providers (observed on OpenRouter) report cached_tokens // OpenRouter's own provider/tests affirm the separate mapping:
// as (previous hits + current writes). In that case, remove cacheWrite from cacheRead. // https://github.com/OpenRouterTeam/ai-sdk-provider/pull/409
const cacheReadTokens = // Do not subtract writes from cached_tokens, otherwise spec-compliant
cacheWriteTokens > 0 ? Math.max(0, reportedCachedTokens - cacheWriteTokens) : reportedCachedTokens; // providers are under-reported. DS4 mirrors this contract too:
// https://github.com/antirez/ds4/pull/29
const input = Math.max(0, promptTokens - cacheReadTokens - cacheWriteTokens); const input = Math.max(0, promptTokens - cacheReadTokens - cacheWriteTokens);
// OpenAI completion_tokens already includes reasoning_tokens. // OpenAI completion_tokens already includes reasoning_tokens.
const outputTokens = rawUsage.completion_tokens || 0; const outputTokens = rawUsage.completion_tokens || 0;

14
packages/ai/test/openai-completions-tool-choice.test.ts
View File

@@ -819,7 +819,7 @@ describe("openai-completions tool_choice", () => {
expect(response.usage.totalTokens).toBe(43); expect(response.usage.totalTokens).toBe(43);
}); });
it("preserves prompt_tokens_details.cache_write_tokens from chunk usage", async () => { it("preserves prompt_tokens_details cache read/write fields from chunk usage", async () => {
mockState.chunks = [ mockState.chunks = [
{ {
id: "chatcmpl-cache-write", id: "chatcmpl-cache-write",
@@ -853,13 +853,14 @@ describe("openai-completions tool_choice", () => {
{ apiKey: "test" }, { apiKey: "test" },
).result(); ).result();
expect(response.usage.input).toBe(50); // cached_tokens is documented as cache reads; cache_write_tokens is separate.
expect(response.usage.cacheRead).toBe(20); expect(response.usage.input).toBe(20);
expect(response.usage.cacheRead).toBe(50);
expect(response.usage.cacheWrite).toBe(30); expect(response.usage.cacheWrite).toBe(30);
expect(response.usage.totalTokens).toBe(105); expect(response.usage.totalTokens).toBe(105);
}); });
it("preserves prompt_tokens_details.cache_write_tokens from choice usage fallback", async () => { it("preserves prompt_tokens_details cache read/write fields from choice usage fallback", async () => {
mockState.chunks = [ mockState.chunks = [
{ {
id: "chatcmpl-cache-write-choice", id: "chatcmpl-cache-write-choice",
@@ -898,8 +899,9 @@ describe("openai-completions tool_choice", () => {
{ apiKey: "test" }, { apiKey: "test" },
).result(); ).result();
expect(response.usage.input).toBe(50); // cached_tokens is documented as cache reads; cache_write_tokens is separate.
expect(response.usage.cacheRead).toBe(20); expect(response.usage.input).toBe(20);
expect(response.usage.cacheRead).toBe(50);
expect(response.usage.cacheWrite).toBe(30); expect(response.usage.cacheWrite).toBe(30);
expect(response.usage.totalTokens).toBe(105); expect(response.usage.totalTokens).toBe(105);
}); });