0.9.1 — 2026-05-11

Release Notes

Patch release rolling up the post-0.9.0 daemon-stall fix (#55) and the
daemon-health deep-probe + SDK admin wrappers follow-up (#56), plus the
docs/CI groundwork from #51, #52, #54.

Added

• daemon.health({deep: true}) shim probe. Adds ShimRequest::Health
  CBOR variant + shim-side handler returning
  ShimHealthInfo { uptime_ms, requests_served, last_request_at_ms }. The
  daemon fans out probes concurrently with per-shim and overall budgets
  (defaults 1 s / 3 s; env-overridable via LOOM_PROBE_TIMEOUT_MS and
  LOOM_DEEP_HEALTH_BUDGET_MS). Payload is now typed
  Vec<ShimDeepHealth> (replacing the prior placeholder
  Vec<serde_json::Value>) with a typed
  ProbeStatus { Ok, Timeout, Error } enum. The new shim variant
  requires same-tree daemon+shim shipping; version-mismatched shims
  surface as probe_status: "timeout". daemon.health continues to
  require the existing socket-auth token handshake — same auth surface
  as before; no anonymous reachability. Known gap: no per-connection
  rate limit on daemon.health itself (tracked as follow-up).
• ShimState.restart_count / last_restart_at_ms. Daemon-side
  bookkeeping for shim respawn events, bumped in get_or_spawn only
  when a prior state entry exists (avoids overcounting under
  open-breaker rejection). Exposed via the deep-health payload.
• SDK wrappers (TS + Python) for the admin RPCs. Hand-written
  Session.kill() (sync + async parity in Python), killSession()
  (TypeScript) / kill_session() (Python sync) admin free functions
  with documented "ADMIN ESCAPE HATCH — 5 s ceiling then SIGKILL"
  warnings, and daemonHealth() (TS) / daemon_health() (Python sync)
  free functions. JSON-RPC request id allocator switched from
  hardcoded 1 to monotonic per-connection; transports moved to
  id-keyed response demultiplexing (persistent reader task in async
  Python; persistent data listener in TS) to support request.cancel
  correlation while another call is in flight. Sync Python keeps
  single-in-flight and gains an AsyncSession-redirecting doc note for
  cancellation. TS adds AbortSignal support on call() via
  LoomAbortError (name === "AbortError"); Python async transport
  transparently emits request.cancel on asyncio.CancelledError and
  re-raises so asyncio.wait_for / TaskGroup / asyncio.timeout
  compose cleanly.
• SessionScope primitive (loom-core/src/session_scope/).
  Per-session structured-concurrency parent: tokio_util::sync::CancellationToken
  paired with a tokio::task::JoinSet. All session-lifetime spawns become children;
  drain(grace) cancels cooperatively then force-aborts survivors. Replaces five
  fire-and-forget tokio::spawn sites that previously leaked JoinHandles and
  saturated the daemon runtime after 4–6 sequential client sessions.
• Per-request server-side deadline. LOOM_REQUEST_TIMEOUT_MS (default
  30000) wraps router.dispatch so a hung shim or stuck dispatcher can't hold
  a connection task. Returns the typed request-timeout envelope on expiry.
• request.cancel RPC. Connection-scoped: cancels an in-flight request on
  the same connection by JSON-RPC id. Returns {cancelled: bool}. The
  cancelled request returns the typed request-cancelled envelope.
• session.kill RPC. Admin escape hatch for stuck sessions. Performs the
  abort flow plus blocks on shim teardown with a 5 s ceiling, then SIGKILL
  (shutdown_process already escalates SIGTERM(2s) → SIGKILL(1s) inside the
  ceiling).
• daemon.health RPC. Operational snapshot: active_sessions,
  shim_breaker_states, otel_exporter status. Shallow path is
  non-blocking; {deep: true} slot reserved for a follow-up shim probe.
• LoomErrorCode::RequestTimeout (wire string request-timeout) and
  LoomErrorCode::RequestCancelled (wire string request-cancelled).
  Both SemVer-minor additions per the existing wire-stability commitment.
• Multi-session stress test at
  loom-core/tests/multi_session_stall_repro.rs
  — 100 sequential session create+close cycles complete in under 1 s; the
  user-facing success criterion from the original investigation prompt.

Changed

• ShimManager::record_failure circuit-breaker eviction no longer fires
  tokio::spawn(shutdown_process(p)) fire-and-forget. Spawns into a
  cleanup_tasks: JoinSet<()> field with opportunistic try_join_next
  reaping on every record_failure and shutdown_session.
• CoreBridge::close_session_raw (daemon-side) similarly tracks
  host.shutdown_session background tasks in a cleanup_tasks JoinSet
  instead of leaking them via bare tokio::spawn.
• Session.budget_timer is now a SessionScope child with
  tokio::select! { _ = cancel.cancelled() => {} _ = sleep(budget_ms) => ... }
  so a closed session's timer exits cooperatively before its sleep
  completes, never tripping the kill callback after lifecycle exit.

Removed

• Session.task_handle field — was dead code (declared but never
  populated). Replaced by the always-populated Session.scope: Arc<SessionScope>.
• Session.budget_timer field — superseded by the SessionScope-owned
  cancellation-aware timer.
• ReceiptMarshaller.queue_depth field + the docstring claiming
  "Background queue depth 256; full → synchronous append on the calling task".
  The field was dead and the docstring lied — there was no bounded channel,
  no try_send, no backpressure. Removed both.

Install loom-cli 0.9.1

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/mentiora-ai/loom/releases/download/v0.9.1/loom-cli-installer.sh | sh

Install prebuilt binaries via Homebrew

brew install mentiora-ai/loom/loom

Download loom-cli 0.9.1

File	Platform	Checksum
loom-cli-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
loom-cli-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
loom-cli-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
loom-cli-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum