`instructions = [...]` (the per-workspace config-driven block),
the user memory file (`/memory`), and the current session goal
(`/goal`) were being rendered at position 2.5 in the system
prompt — inside the static prefix layer that DeepSeek's KV
prefix cache hits.
Any edit to those files invalidated every cached byte from that
position onward. A `# foo` memory quick-add (or a `/goal` update)
on turn 5 meant the engine had to re-tokenize and re-charge the
full static suffix — skills block, context management, compact
template, environment, ~thousands of tokens — on turn 6.
Relocate the three blocks to position 6, immediately above the
previous-session handoff block, where the volatile-content
boundary already lives. The static prefix above the boundary
(mode, project context, env, skills, context management, compact
template) now stays cached across turns regardless of how often
the user edits their memory file or shifts session goals.
Resolved a 3-way merge against the v0.8.32 `translation_enabled`
addition (PR #1462). The new translation-output instruction stays
at position 2.3a (inside the static prefix layer) because it's a
per-session flag — `/translate` is a session toggle, not a
turn-by-turn knob, so the prompt-prefix bytes don't drift
mid-session.
Harvested from PR #1345 by @Duducoco
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Hunter Bown
a9412d7df8