perf(cache): static/volatile system split — cache the core across the whole session#47
Merged
Merged
Conversation
… whole session Step 3 of the layered-context architecture. #44 caches the system block within a task, but across conversation TURNS the per-turn injections (memory, retrieval, dir-tree) change, so the whole system block cache-misses every turn — taking the stable instruction core down with it. This splits the system prompt at a static/volatile boundary so the byte-stable core gets its own cache breakpoint and stays a HIT for the entire session: - Message gains a transient `cacheBoundary` (char offset where the stable core ends). Set in buildInitialMessages right after the identity + base prompt + provider guidance, BEFORE the volatile injections. Not persisted; rebuilt each run. - Zero-leak by design: ONLY the Anthropic provider reads cacheBoundary (convertMessages surfaces it; withCacheBreakpoints splits system into [core (cached) | volatile (uncached)]). Every other provider just reads `content` — no sentinel string, no changes to them. - Still ≤4 breakpoints (core + last tool + rolling last message). No boundary, or one out of range ⇒ the previous single-block behavior. +3 tests (split marks only the core; fallback when boundary absent/out-of-range; ≤4 breakpoints). Full suite 1339 green; tsc clean.
QodeXcli
added a commit
that referenced
this pull request
Jun 30, 2026
…turn hit) (#48) Follow-up to the static/volatile split (#47). #47 put the boundary right after the base prompt, so ALL injections — including session-STABLE ones (code-style profile, failure lessons) — landed in the uncached volatile tail and re-billed every turn. Now injections route into two buffers: - stableTail (code style, failure lessons) — byte-identical across turns → folded INTO the cached core, so they're a cache HIT for the whole session. - volatileTail (auto-retrieval, dep-graph, episodic recall) — genuinely query-dependent → stays after the boundary, uncached. The cache boundary now lands between them. Pure content regrouping (no message bloat — volatile stays in the regenerated system prompt, never persisted to history) — guidance simply precedes per-turn context now. Also helps LOCAL backends: a larger byte-stable prefix means Ollama/ llama.cpp KV prefix-cache hits more across turns (local "turbo cache"). Full suite 1339 green; tsc clean. The failure-lessons taskKey side effect is preserved. Co-authored-by: Louise Lau <QodeXcli@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
Step 3 of the layered-context architecture (the "continue" after #44–#46). #44 caches the system block within a task, but across conversation turns the per-turn injections (memory, retrieval, dir-tree) change — so the whole system block cache-misses every turn, dragging the stable instruction core (~2k tokens of identity + base prompt, byte-identical every turn) down with it and re-billing it.
What
Split the system prompt at a static/volatile boundary so the byte-stable core gets its own cache breakpoint and stays a HIT for the whole session, not just within one task.
Message.cacheBoundary(transient, not persisted) — char offset where the stable core ends. Set inbuildInitialMessagesright after identity + base prompt + provider guidance, before the volatile injections.cacheBoundary:convertMessagessurfaces it,withCacheBreakpointssplitssysteminto[core (cached) | volatile (uncached)]. Every other provider just readscontent— no sentinel string in the prompt, no changes to them.Honest scope
This nails cross-turn caching of the core. It does not fully cache the conversation history across turns — volatile injections still sit before the message prefix in cache order, so the history re-caches when they change. Fully fixing that means moving volatile after the history (a behavior-affecting restructure) — deliberately deferred.
Tests
+3 (split marks only the core block; fallback when boundary absent/0/past-end; ≤4 breakpoints). Full suite 1339 green,
tscclean.