Skip to content

YC (S26) | Open Computer History | Continuously record your company computer work, map your workflows, help you find work worth automating, and power your agents' context

fix(ai): keep cache markers off media content (#6385)

Ezra Ellette8696fbe7851f
diff --git a/packages/ai-gateway/src/providers/openai.ts b/packages/ai-gateway/src/providers/openai.ts
index 659107043..8b742d8b2 100644
--- a/packages/ai-gateway/src/providers/openai.ts
+++ b/packages/ai-gateway/src/providers/openai.ts
@@ -25,9 +25,6 @@ const GPT56_MAX_READ_BREAKPOINTS = 50;
const GPT56_HISTORY_BREAKPOINTS = GPT56_MAX_READ_BREAKPOINTS - 1;
const CACHEABLE_CONTENT_TYPES = new Set([
'text',
- 'image_url',
- 'input_audio',
- 'file',
'refusal',
]);
@@ -50,14 +47,6 @@ function isCacheableContentPart(part: any): boolean {
return typeof part.text === 'string' && part.text.length > 0;
case 'refusal':
return typeof part.refusal === 'string' && part.refusal.length > 0;
- case 'image_url':
- return typeof part.image_url === 'string'
- ? part.image_url.length > 0
- : Boolean(part.image_url?.url);
- case 'input_audio':
- return Boolean(part.input_audio?.data);
- case 'file':
- return Boolean(part.file?.file_id || part.file?.filename || part.file?.file_data);
default:
return false;
}
diff --git a/packages/ai-gateway/src/test/openai-models.unit.test.ts b/packages/ai-gateway/src/test/openai-models.unit.test.ts
index 2dcb62a7c..44d853f71 100644
--- a/packages/ai-gateway/src/test/openai-models.unit.test.ts
+++ b/packages/ai-gateway/src/test/openai-models.unit.test.ts
@@ -412,32 +412,33 @@ describe('OpenAI API accounting and routing', () => {
expect(marked).toEqual([true, false, true, false, false, true, false, true, false, false, true]);
});
- it('marks the last supported multipart image or file block without changing it', async () => {
+ it('marks preceding multipart text without adding cache fields to media blocks', async () => {
const imagePart = {
type: 'image_url',
image_url: { url: 'data:image/png;base64,AA', detail: 'high' },
};
const filePart = { type: 'file', file: { file_id: 'file_123' } };
+ const audioPart = { type: 'input_audio', input_audio: { data: 'AA', format: 'wav' } };
const params: any = {
model: 'gpt-5.6-luna',
messages: [
{ role: 'system', content: 'stable' },
{ role: 'user', content: [{ type: 'text', text: 'look' }, imagePart, { type: 'text', text: '' }] },
{ role: 'assistant', content: 'seen' },
- { role: 'user', content: [filePart, { type: 'text', text: '' }] },
+ { role: 'user', content: [{ type: 'text', text: 'read' }, filePart, { type: 'text', text: '' }] },
+ { role: 'assistant', content: 'read' },
+ { role: 'user', content: [{ type: 'text', text: 'listen' }, audioPart, { type: 'text', text: '' }] },
],
};
await applyGpt56PromptCaching(params, true);
- expect(params.messages[1].content[1]).toEqual({
- ...imagePart,
- prompt_cache_breakpoint: { mode: 'explicit' },
- });
- expect(params.messages[3].content[0]).toEqual({
- ...filePart,
- prompt_cache_breakpoint: { mode: 'explicit' },
- });
+ expect(params.messages[1].content[0].prompt_cache_breakpoint).toEqual({ mode: 'explicit' });
+ expect(params.messages[1].content[1]).toEqual(imagePart);
+ expect(params.messages[3].content[0].prompt_cache_breakpoint).toEqual({ mode: 'explicit' });
+ expect(params.messages[3].content[1]).toEqual(filePart);
+ expect(params.messages[5].content[0].prompt_cache_breakpoint).toEqual({ mode: 'explicit' });
+ expect(params.messages[5].content[1]).toEqual(audioPart);
});
it('recomputes GPT-5.6 breakpoints idempotently after rewritten history', async () => {