kunobi-ninja/kache v0.0.14

on GitHub

What's Changed

Fire-and-forget daemon protocol

• Upload and prefetch requests write to the kernel socket buffer and close immediately — no waiting for a response
• Eliminates read-timeout failures when the daemon's Tokio runtime is saturated during S3 key-cache population at startup

Client-epoch auto-restart

• Upload and BuildStarted requests carry client_epoch (binary mtime)
• Daemon detects when client binary is newer and schedules graceful shutdown for restart
• Replaces the stats-request-based version check that added an extra round-trip prone to timeouts

BFS prefetch ordering

• Wrapper uses cargo metadata with full dependency graph, then Kahn's algorithm for topological sort
• Crate names emitted in compilation order (leaves first) so daemon prefetches what cargo needs soonest
• S3KeyCache reverse index (crate_name → cache keys) enables prefetch on cold CI runners
• Session timeout increased from 30s to 120s to reduce redundant build-started triggers

Dual-layer logging

• Stderr layer (KACHE_LOG, default warn) for build output
• File layer (KACHE_LOG_FILE, default info) at ~/Library/Logs/kache/kache.log (macOS) or <cache_dir>/kache.log (Linux)
• kache daemon log and kache daemon status show diagnostic log path

Additional hardening

• Structured tracing fields on upload path for observability
• Daemon exits cleanly when another instance is already running (prevents launchd restart loops)
• Prefetch cancellation properly drains all remaining keys from the downloading set