Skip to content

v0.3.0

Choose a tag to compare

@mjbommar mjbommar released this 22 Apr 00:51
· 25 commits to main since this release

[0.3.0] — 2026-04-21

Added

Hosted deployment profile + operator profiling.

  • kernel-lore-mcp serve now accepts --mode local|hosted, and the
    chosen runtime posture is logged explicitly at startup along with
    bind/port, data dir, and the active slow-path thresholds.
  • Python-side hosted logging now emits structured slow-path events for
    MCP requests, tool calls, inner runtime, and cost-class admission
    delay, so operators can distinguish queueing from tool-body work
    during live incidents.
  • New environment knobs:
    KLMCP_SLOW_REQUEST_MS, KLMCP_SLOW_TOOL_MS, and
    KLMCP_SLOW_QUEUE_WAIT_MS.

Repeatable hosted-load gate.

  • Added scripts/bench/bench_hosted_adversarial.py: a real HTTP/MCP
    harness that boots the server in hosted mode, floods cheap queries,
    saturates moderate and expensive tools, polls /status concurrently,
    and emits a JSON report comparing client-observed latency with
    server-side histograms.
  • Wired the harness into CI and added
    docs/ops/public-launch-checklist.md so launch readiness is a
    repeatable gate, not an operator memory exercise.

Changed

The would-be 0.2.3 hosted-hardening line ships here.

  • kernel-lore-sync now self-heals poisoned local shard repos by
    deleting zero-ref / unopenable repos and recloning them during sync.
  • Hosted lore_regex is permanently narrowed to list-scoped,
    anchored, metadata-only regex shapes; unsafe full-corpus / prose /
    patch scans now reject fast with hosted_restriction.
  • lore_corpus_stats is now cached by generation and warmed at server
    startup so the steady-state path stays off the timeout cliff.
  • Sync rebuilds paths/vocab.txt automatically, eliminating the
    post-sync manual step for lore_path_mentions.
  • /metrics now records end-to-end request latency, queue-wait, tool
    latency, and non-ok statuses including rate_limited even when a
    call is rejected before the tool body runs.

Fixed

  • Hosted default logs are quiet enough to follow a real incident:
    third-party INFO chatter is suppressed by default while our own
    warnings/errors remain visible.
  • The latency gap seen in client stress runs is now measurable rather
    than anecdotal: the hosted-load harness and the new metrics/logging
    surface make queueing / transport inflation directly visible.