feat: enhance responses API context management with server-side compaction

caozhiyuan · caozhiyuan · commit 371d53128233 · 2026-06-01T21:10:04.000+08:00
diff --git a/src/routes/responses/handler.ts b/src/routes/responses/handler.ts
@@ -122,10 +122,9 @@ export const handleResponses = async (c: Context) => {
     )
   }
 
-  applyResponsesApiContextManagement(
-    payload,
-    selectedModel?.capabilities.limits.max_prompt_tokens,
-  )
+  // Smaller than the client compaction threshold, use server-side compaction to maintain cache hit rate
+  const maxPromptTokens = selectedModel?.capabilities.limits.max_prompt_tokens
+  applyResponsesApiContextManagement(payload, maxPromptTokens, 0.8)
 
   debugJson(logger, "Translated Responses payload:", payload)
 

Original file line number	Diff line number	Diff line change
`@@ -122,10 +122,9 @@ export const handleResponses = async (c: Context) => {`
`122`	`122`	`)`
`123`	`123`	`}`
`124`	`124`
`125`		`- applyResponsesApiContextManagement(`
`126`		`- payload,`
`127`		`- selectedModel?.capabilities.limits.max_prompt_tokens,`
`128`		`- )`
	`125`	`+ // Smaller than the client compaction threshold, use server-side compaction to maintain cache hit rate`
	`126`	`+ const maxPromptTokens = selectedModel?.capabilities.limits.max_prompt_tokens`
	`127`	`+ applyResponsesApiContextManagement(payload, maxPromptTokens, 0.8)`
`129`	`128`
`130`	`129`	`debugJson(logger, "Translated Responses payload:", payload)`
`131`	`130`