Claude Code submits more than typed words through UserPromptSubmit. A task
notification, a slash-command echo and the caveat banner a local command
prints all arrive as user turns, reaching `.prompt` indistinguishable from
something the operator wrote. scribe_autoinject.sh believed all of them.
The cost that matters is not the wasted embedding — it is the log row. Every
such call counts in the denominator of every prompt-boundary surface, so
delivery rate reads low for a reason unrelated to retrieval; and each refusal
lands in `near_misses`, where a later tuning decision reads it as demand.
Measured while taking milestone 399's acceptance (#3898): 15 of the top 20
`preference_slot` near-misses were `<task-notification>` blocks, all matching
ONE record — #140 "Let each action land before starting the next" — all within
thousandths of the 0.70 floor. A notification that an action finished really
does resemble a preference about letting actions land. Lowering the floor to
serve that apparent demand would have injected that record into every
notification: the instrument arguing for the wrong fix, which is #379 again.
scribe_skip_prompt is a PREFIX test, not a substring one, and that is the
whole safety argument. A real prompt may contain one of these tags — an
operator pasting a transcript, or a `<system-reminder>` after typed words —
and must still be retrieved against. Nothing an operator types begins with a
client envelope. The compaction-resume injection is deliberately NOT filtered:
it is machine-written, but it summarises real work, and a resumed session is
where recalling a rule earns its keep.
Client-side only, so no server contract moves and lagging plugin caches keep
working. The tags are Claude Code protocol constructs, identical on every
install — instance-agnostic under rule 115.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01821k5B3Ysecp9fNYs92Kuy
bvandeusenandClaude Opus 5
1f7ff7b215