[diegosouzapw]

Closes #3890

Problem

With compression and/or memory enabled, requests to caching providers (Anthropic-family) miss the prompt cache on every turn, multiplying cost (the reporter saw ~5x). Disabling both restored a 90-99% hit rate.

Root causes (two, both addressed)

1. Memory injection poisoned the cacheable prefix. injectMemory prepended the retrieved memories at index 0 of the message array. Retrieved memories vary per user query (retrieveMemories({ query: lastUserQuery })), so the very front of the prompt changed every turn → every cache_control breakpoint downstream missed.
2. The cache-aware skipSystemPrompt flag was dropped. getCacheAwareStrategy() computes skipSystemPrompt/deterministicOnly, but selectCompressionStrategy() can only return a mode string, so those flags were discarded — the system prompt (part of the cacheable prefix) could still be compressed under caching.

Fix

1. Memory (src/lib/memory/injection.ts + memorySkillsInjection.ts): when the request carries cache_control breakpoints (detectCachingContext(...).hasCacheControl), insert the memory message just before the last user message instead of at index 0 — the cacheable prefix (system prompt + prior turns) stays byte-stable while memory still contextualizes the current turn. Non-caching requests keep the existing leading-injection behavior.
2. Compression (strategySelector.ts + chatCore.ts): new resolveCacheAwareConfig() applies the dropped skipSystemPrompt decision by forcing preserveSystemPrompt on for caching requests, even when the operator disabled it. Non-caching requests are untouched.

Both changes only ever preserve more of the prefix under caching — they cannot make caching worse.

Validation (TDD — Hard Rule #18)

• tests/unit/memory-cache-safe-injection.test.ts (new): proves memory is inserted before the last user message under caching (prefix byte-stable across two turns), keeps legacy index-0 behavior when caching is off, and falls back safely when there is no user message.
• tests/unit/compression/strategySelector-cache-aware.test.ts (new): proves resolveCacheAwareConfig forces preserveSystemPrompt for caching requests and leaves non-caching requests untouched.

All green; typecheck:core clean, lint 0 errors, no new cycles, regression suites green (memory-glm-injection 22/22, cachingAware 17/17). file-size re-baselined (chatCore 5815→5823).

@xenstar — the unit tests prove the request no longer injects per-query content into the cacheable prefix; please confirm the cache-hit rate recovers on v3.8.26 with your RTK/Caveman + memory setup, since upstream cache-hit rates can only be observed live.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.