When a scheduled task gets cancelled via `docket.cancel()`, the cancel
Lua
script removes it from the stream/queue but leaves the `known` field in
the
runs hash. Later, when `_schedule_all_automatic_perpetual_tasks` (or any
`docket.add()`) tries to reschedule the same key, the schedule script
sees
`known` and returns EXISTS. The task stays silently unscheduled until
`execution_ttl` expires (default 15 min).

The fix adds an `HDEL runs_key 'known' 'stream_id'` to the cancel
script,
clearing the dedup markers while keeping `state=cancelled` and
`completed_at`
for observability. The schedule script's existing fallback (`state ==
'running'`
→ EXISTS) already handles cancelled tasks correctly once `known` is
gone.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

`build_cluster_image` does a check-then-build that can race when
parallel
test sessions try to build the same image. The `images.get()` check
passes
for both, then one build succeeds and the other gets a `BuildError` with
`AlreadyExists`. This showed up as every test failing in a CI cluster
job:


https://github.com/chrisguidry/docket/actions/runs/22020705636/job/63639829973

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

…ob (#338)

Two changes to improve CI reliability:

**Serialize cluster image builds with file lock**

The `AlreadyExists` fix in #337 handled one symptom of parallel xdist
workers racing to build the same cluster image, but there's a second
failure mode showing up in CI:


https://github.com/chrisguidry/docket/actions/runs/22025132964/job/63640478732

When concurrent builds target the same tag, the Docker SDK's `build()`
completes successfully in the daemon, then tries to inspect the
resulting image by its short ID. If another worker's build re-tagged the
image in the meantime, the first image ID gets orphaned and the inspect
404s. This knocked out 485 of 686 tests in the cluster job.

Rather than catching yet another exception type, this serializes the
builds with `fcntl.flock` so only one worker builds at a time. The
others wait and find it already built. Eliminates both the
`AlreadyExists` and `ImageNotFound` races structurally.

**Split CLI tests into separate CI job**

Cluster CI jobs consistently run right at the 4-minute timeout, and when
any test runs slightly slow the whole job gets cancelled. This has been
showing up in roughly a third of recent CI runs:


https://github.com/chrisguidry/docket/actions/runs/22025359927/job/63641245074

The 91 CLI tests are subprocess-based and don't exercise
backend-specific behavior — they spawn `python -m docket ...` processes
and check output. Running them against every Python x Backend
combination (30 matrix entries) is wasted effort.

This moves CLI tests to their own job that varies by Python version but
uses a single Redis backend (8.0). The main test matrix now passes
`--ignore=tests/cli` so cluster/valkey/memory jobs only run the tests
that actually care about the backend. Local `pytest` runs are
unaffected.

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Docket's worker loops rely on Redis blocking behavior that fakeredis
doesn't implement — `get_message()` and `xread()` return immediately
instead of blocking, turning the loops into CPU spinners. The
`xreadgroup` loop already had a workaround; this applies the same
pattern to the cancellation listener and strike monitor.

Also fixes a few things that prevent running on Windows at all: signal
handler registration (raises `NotImplementedError` on Windows instead
of working), `fcntl` import (doesn't exist on Windows — swapped for
cross-platform locking with `msvcrt`), and Docker SDK imports in
conftest that fail when Docker isn't installed.

Adds a Windows CI job (memory:// only, 3 Python versions) and a
regression test that directly measures idle CPU usage.

Closes #339

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

redis-py 7.2.0 deprecated `lib_name`/`lib_version` on
`Connection.__init__`
in favor of `DriverInfo`, but their own `RedisCluster` still passes
`lib_name="redis-py"` internally, which triggers a `DeprecationWarning`.
Our `filterwarnings = ["error"]` promotes that to an exception, breaking
cluster tests. We pinned `<7.2` in #340 as a stopgap.

This adds a targeted warning filter for that specific deprecation so we
can run with redis-py 7.2.0 unpinned. Also bumps the Redis version in CI
from 8.0 to 8.6 to track the latest release.

Closes #341

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

mkdocs-material has entered maintenance mode in favor of Zensical, which
is built by the same team and reads mkdocs.yml directly. This swaps the
docs dependencies and build command accordingly.

Also removes a stale top-level sitecustomize.py that snuck in with #170,
adds site/ to .gitignore, and updates the README with current
Redis/Valkey test versions and a note on working with docs locally.

Closes #273

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>