Module cache_telemetry 

Read-cache reliability telemetry.

The subjective “re-reads feel unreliable” signal is turned into data here: every event that wipes or invalidates a fully-delivered entry — and thus forces the next read to re-send the whole file instead of the cheap [unchanged] stub — increments a process-global counter grouped by cause.

Counters are monotonic AtomicU64s. They are surfaced ONLY in the ctx_cache status diagnostic, never inside a cacheable tool-output body, so output determinism (#498) is preserved.

Structs

Snapshot

Immutable snapshot of the re-delivery counters.

Functions

record_compaction

Fully-delivered entries whose delivery flag was reset by a host compaction.

record_conversation_mismatch

A re-read that fell back to full content because the reading conversation differed from the one the entry was delivered to (conversation scoping, #954).

record_eviction

Fully-delivered entries evicted under RAM / token-budget pressure.

record_idle

Fully-delivered entries dropped by an idle-TTL cache clear.

record_raw_cap

Framed output was larger than raw content — cap_to_raw() fell back to verbatim to prevent negative savings. Tracked so the dashboard can distinguish “no savings” from “savings prevented inflation”.

snapshot

Reads the current counters into a consistent snapshot.