onlyterp/prompt-cache-skills/continue-enable-defaults

Continue: enable caching defaults

Target

packages/openai-adapters/src/apis/Anthropic.ts and core/llm/llms/Bedrock.ts in continuedev/continue.

Symptom

Continue's caching is gated on three config flags:

• cacheBehavior.cacheConversation (default: false)
• cacheBehavior.cacheSystemMessage (default: false)
• completionOptions.promptCaching (default: false)

A user has to know all three exist and set them in config.yaml. Most don't. Net effect: the median user gets no caching.

Open issue #5172 ("Anthropic prompt caching doesn't work") is mostly this — users don't realize they need to configure it.

Fix (two-part)

Part A — flip the default strategy

In Anthropic.ts:

--- a/packages/openai-adapters/src/apis/Anthropic.ts
+++ b/packages/openai-adapters/src/apis/Anthropic.ts
@@
-  const cachingStrategy = CACHING_STRATEGIES[this.config.cachingStrategy ?? "systemAndTools"];
+  // Default to caching system + tools. Users on supported models pay
+  // the 1.25x write premium on the first call and get 0.1x on every
+  // subsequent call within 5 minutes — strict win above ~3 reads.
+  const cachingStrategy = CACHING_STRATEGIES[this.config.cachingStrategy ?? "systemAndTools"];

(systemAndTools already IS the documented default in the code path above, but the user-facing config schema treats it as optional with unclear default. Audit your codebase to confirm the actual fallback behavior matches the documented one. If not, fix.)

Part B — auto-enable conversation message caching for supported models

@@
-  if ((this.config.cachingStrategy ?? "systemAndTools") !== "none") {
-    addCacheControlToLastTwoUserMessages(result.messages);
+  // Auto-enable for any model that declares supportsPromptCache.
+  // User can still opt out with cachingStrategy: "none".
+  const strategy = this.config.cachingStrategy ?? "systemAndTools";
+  if (strategy !== "none" && this.modelSupportsPromptCache()) {
+    // NOTE: this is currently buggy — see continue-fix-volatile-msg
+    // skill for the volatile-message fix that should land alongside.
+    addCacheControlToLastTwoUserMessages(result.messages);
   }

Same change in Bedrock.ts — auto-enable cachePoint for models with supportsPromptCache: true unless user explicitly opts out.

Part C — document the change in CHANGELOG

Users coming from older Continue versions will see bills shift (downward, but still a change). Add a CHANGELOG entry calling out that caching is now default-on for supported models.

Verify

1. With NO cacheBehavior or completionOptions.promptCaching set

in config.yaml, start a Continue chat with Claude.

1. Capture wire.
2. Confirm outbound request body contains cache_control markers

without any user config touching them.

1. Confirm second-turn response usage.cache_read_input_tokens > 0.

Background

Caching that requires a config flag to enable is caching that doesn't get used. The 1.25x write premium pays for itself after ~3 reads, which any multi-turn chat clears trivially. There's no scenario where "caching off" is the right default for an agent CLI on a model that supports it.

This skill is best applied alongside continue-fix-volatile-msg (same Cline-family copy-paste bug present in Continue too) so users don't get the cache thrash without knowing.

Full audit: audits/continue.md.