continue-fix-volatile-msg

name	continue-fix-volatile-msg
description	Continue inherits the Cline-family volatile-message cache bug. Same fix shape.
target_harness	Continue
target_repo	continuedev/continue
target_files	["packages/openai-adapters/src/apis/Anthropic.ts","core/llm/llms/Bedrock.ts"]
target_commit	main (last push 2026-05-26)
estimated_savings	~30% wasted cache write premium eliminated per turn

Continue: fix volatile-message cache thrash

Target

packages/openai-adapters/src/apis/Anthropic.ts and core/llm/llms/Bedrock.ts in continuedev/continue. The relevant function is addCacheControlToLastTwoUserMessages (Anthropic) and _addCachingToLastTwoUserMessages (Bedrock).

Symptom

Same as cline-fix-volatile-msg and roo-fix-volatile-msg: Continue caches the last 2 USER messages, including the current user turn which changes every request. Pays write premium for zero reads on that breakpoint every turn.

Fix

Rename and rewire the function to target stable messages (the previous assistant turn) instead of user messages:

Anthropic

--- a/packages/openai-adapters/src/apis/Anthropic.ts
+++ b/packages/openai-adapters/src/apis/Anthropic.ts
@@
-function addCacheControlToLastTwoUserMessages(messages: Message[]) {
-  const userIdxs = messages
-    .map((m, i) => (m.role === "user" ? i : -1))
-    .filter(i => i >= 0);
-  const lastTwo = userIdxs.slice(-2);
-  for (const idx of lastTwo) {
+function addCacheControlToLastStableMessage(messages: Message[]) {
+  // Cache the last STABLE message (not the volatile current user turn).
+  // Stable = the message before the user's in-flight request.
+  if (messages.length < 2) return;
+  const idx = messages.length - 2;
+  {
     const m = messages[idx];
     // ...existing logic to attach cache_control to last content block
   }
 }

Update the caller in _convertBody() to use the new name.

Bedrock

Same shape in core/llm/llms/Bedrock.ts — replace the _addCachingToLastTwoUserMessages body with the _addCachingToLastStableMessage logic, using cachePoint instead of cache_control.

Verify

Same procedure as the Cline skill — see cline-fix-volatile-msg/SKILL.md#verify. Wire-level assertion is identical: turn-2 cache_creation_input_tokens should drop to ~0, cache_read_input_tokens should cover the full prefix.

Background

This is the third instance of the same copy-paste bug (Cline → Roo → Continue). Apply this skill alongside continue-enable-defaults so users don't get the bug surfaced by the default-on change.

Full audit: audits/continue.md.