Repository navigation
Release Notes
Changed
-
Quiet agent sessions hold less memory. Each agent connected to anvil's MCP
server kept a worker process running for as long as it stayed connected, even
when it had not called anvil for hours. After two minutes with no MCP traffic
the worker now stops, and it starts again in well under a tenth of a second on
the next call, so an idle connection costs only its small host process. Set
ANVIL_MCP_WORKER_IDLE_SECSto change the window, or tooffto keep the
worker running. -
Idle worktrees stop holding a save-time driver. Each registered worktree
kept a background save-time driver, with its own copy of the code graph,
running for as long as the daemon ran, even when nothing had been edited for
hours. After 30 minutes by default with no saves and no MCP session, the
driver now stops, and one small watcher process looks after every idle
worktree. The first edit starts the driver again, and it checks every file
changed while it was stopped, whether or not the edit came through anvil's
MCP.anvil statusshows such a driver as
dormant (profile balanced; wakes on next change). The user setting
daemon.save_time_dormancychooses how soon:eager(10 minutes),balanced
(30, the default),relaxed(60) oroffto keep every driver running. Set
it in your usersettings.yamland runanvil start;anvil doctorshows
the profile in use.ANVIL_SAVE_TIME_DRIVER_IDLE_MINSstill overrides the
window; the daemon reads it when it starts, so set it in the daemon's
environment and restart the daemon. When the daemon starts, a worktree nobody
has used for longer than the profile's age (12 hours underbalanced, 5 days
underrelaxed, and always undereager) goes dormant after about a minute
instead of a full idle window, andanvil doctornames those worktrees.After upgrading, restart your editors and agent sessions. An
anvil mcp servefrom an earlier release that is still running asks for
status in a way that wakes every dormant driver one after another, so it can
report the daemon as slow or unresponsive. On a machine with many registered
worktrees, expect a short burst of CPU and disk activity after a daemon
restart while each driver re-checks the files it watches. Before rolling
back to 0.14.1-beta or earlier, delete the rows marked
"configured_only": truefromregistered-worktrees.json(beside the fence
store, under anvil's state directory); an older daemon would treat them as
ordinary registrations and keep worktrees you have removed from your
configuration. An older release also cannot read the new graph cache, so it
rebuilds every worktree's graph again, as does upgrading once more. -
A worktree too large to scan in time is no longer re-scanned in a loop.
When the first scan of a worktree overran its time budget, every MCP warm-up
started another full scan about every 30 seconds, holding a CPU core. After a
timeout anvil now waits before scanning that worktree again: one minute at
first, doubling to at most 30 minutes, and a scan that completes clears the
wait. Removing the cause (for example, excluding a large generated directory)
does not shorten the wait;anvil mcp refresh --daemon restartclears it. -
The code graph is rebuilt once after the upgrade. The graph cache now
records how each call was resolved, so a cache written by an earlier release
is discarded and every registered worktree is scanned from scratch the first
time it is used. Nothing needs to be done. -
Graph-context results are more exact about calls.
anvil_find_callers
reports how each caller was resolved (same file, through an import, or one of
several overloads). A graph-context page cursor from an earlier release is
refused rather than resumed, so an agent part-way through a paged result
starts the query again. -
anvil conformancedoes not evaluate partial clones. A clone made with
--filter=blob:none(or another partial-clone filter) would fetch missing
objects over the network while anvil read them. It is now reported as not
evaluated, as a shallow clone already was. Use a full clone. -
The bundled agent skill tells agents to edit with their edit tool. The
anvil-developer-functionsskill installed byanvil skill installnow says
that anvil's write checks do not write the file: make the edit with the normal
edit tool, treat a shell write as a write to validate, and retry a failed edit
instead of rewriting the file from the shell. Runanvil updateor
anvil skill installto refresh installed copies.
Added
anvil doctorreports what the graph cache is using. A new check shows
the size of the shared graph bases and snapshots and warns when at least 256
MiB could be reclaimed (withanvil graph-base gc), when a registered
worktree's merge base cannot be found, or when a worktree set to register on
start no longer exists (withanvil workspace unregister --persist <path>).
Fixed
-
Local trace files keep less of what passes through them. With
ANVIL_TRACE_SINK=file=<path>, messages that are not one of anvil's own
diagnostic sentences, error chains and nested debug values are now redacted or
dropped before they are written, and each retained payload is bounded. -
Unused graph bases are reclaimed again. Graph-base clean-up and
persistence read the repository's default branch wrongly on every worktree, so
every clean-up pass kept every base and the cache grew without limit on
machines with many worktrees. Clean-up now finds the default branch, and its
warning names any worktree it still cannot resolve. -
The first scan of a large repository finishes. The initial scan redid
whole-graph work after every file, so on a large repository it overran its
time budget every time, the graph never became ready, and MCP checks timed out
behind it. That work now runs once at the end of the scan. -
A save can no longer go unchecked when the watch worker is finishing. A
save that arrived just asanvil watch's worker decided to exit could be left
queued with no worker to check it. It is now always handed to a worker, and a
second worker is not started. -
anvil checksuppression matchesanvil gate, and the counts are true.
An@anvil-ignoreon the previous line was still treated as a live error by
anvil check, which reportedsuppressed 0whileanvil gatewithheld the
same match. Suppressed findings now stay in the result without blocking, and
JSON rows carrysuppressedReason; if you countwarnings[], skip rows that
have it. Withheld secret matches are now included insummary.suppressedand
summary.total. High-confidence secret shapes still cannot be ignored inline.
A comment on thepassof anexcept: passno longer hides PY-010, so code
relying on that will see the warning again; put the ignore on the line above
theexcept. -
TypeScript helpers inside an exported function are no longer public API.
Every symbol inside an exported declaration, including local helpers in a
function body, was treated as public, so adding a helper read as a public API
change. Only the export itself, and the methods of an exported class, are now
public. -
A relative import that climbs above the repository root stays unresolved.
An import such as../../../../afromsrc/main.tswas bound to a root-level
a.ts. It is now unresolved. -
A deeply nested file can no longer crash a check. A small file with about
1,000 nested namespaces or modules could overflow a worker's stack and abort
the process. A file nested more than 1,024 levels deep is now skipped, like a
file that does not parse. -
Concurrent fence and unblock writes no longer drop each other. Two
save-time fence updates at the same moment could lose the earlier one. -
Concurrent network-switch changes no longer erase each other. Two
network-feature switch changes made at the same moment, such as
anvil updates off, could leave only one of them recorded. -
anvil watchreports more accurately.- It warns when it cannot watch new directories, for example when a large
Linux checkout hits the inotify watch limit; before, those directories went
unwatched silently. - A comment-only file excluded by
--patternsor--excludeno longer counts
as tracked. - The fallback message repeats the value you set for
ANVIL_WATCH_DAEMON(for
examplefalseorOFF) rather than always printing0.
- It warns when it cannot watch new directories, for example when a large
-
The airgap now stops signed-in account activity. While you are signed in,
anvil records which core commands your account used. That request ignored the
airgap, includingANVIL_AIRGAPPED=1, so an airgapped install could still
open a socket to the anvil API. It now stops when the airgap is on, when
ANVIL_ACCOUNT_ACTIVITY=offorDO_NOT_TRACK=1is set, or when you sign out.
A request already scheduled checks the switch again before it sends. It is
registered as theaccount-activitynetwork feature. Theanvil airgapped
command that lists those features ships behind its own release switch and is
off by default in this build, so useANVIL_AIRGAPPED=1to airgap a process. -
A worktree no longer gets two save-time drivers when a driver start is
slow. If the daemon could not take a worktree's driver start lock within two
seconds, for example while the wake sentinel was starting the same driver, it
started a second one anyway. Two drivers then ran duplicate save-time checks
and wrote the same findings log and PID record. The daemon now adopts the
driver the other start recorded. If the lock stays held, it starts nothing:
anvil statusshows the driver as failed until the worktree is registered
again (anvil startor a bareanvilthere); a read such asanvil status
does not retry, and a dormant worktree's driver is started by the sentinel
instead. -
An architecture file that breaks while the save-time driver is running is
reported. If.anvil/architecture.yamlstops loading after the driver has
started, boundary enforcement stops, the findings log records the reload
failure, andanvil statusshows
driver: degraded (architecture config unreadable). Fixing the file resumes
enforcement. A driver started against a file that is already broken still
refuses to start. -
A drift snapshot no longer treats a partial boundary walk as a clean
count. A source file that did not parse used to disappear from
anvil drift snapshot. The snapshot could still say the boundary check ran
and report zero violations, hiding an import that lived only in that file.
Those files are now named inmetrics.boundary_unparsed_files, and the
boundary count is marked incomplete. Imports in files that did parse are still
recorded. Compare, the drift dashboard, and the drift trend treat a partial or
unknown count as a lower bound: they do not report an exact boundary delta, a
stable or improving trend, or a measured zero from it. An older snapshot with
no list still means the coverage is unknown. -
MCP servers no longer outlive the agent that started them. When an editor
or agent exited without closing anvil's MCP connection cleanly, its
anvil mcp serveprocess and the worker behind it could keep running until
you killed them. On macOS and Linux they now notice their parent has gone and
exit within a few seconds, even mid-request. A fix or suppression interrupted
that way, or by a crash, no longer leaves the file reporting "locked by
another writer" for later sessions.
Install eddacraft-anvil 0.14.2-beta
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/eddacraft/anvil/releases/download/v0.14.2-beta/eddacraft-anvil-installer.sh | shInstall prebuilt binaries via powershell script
powershell -ExecutionPolicy Bypass -c "irm https://github.com/eddacraft/anvil/releases/download/v0.14.2-beta/eddacraft-anvil-installer.ps1 | iex"Install prebuilt binaries via Homebrew
brew install eddacraft/tap/anvilDownload eddacraft-anvil 0.14.2-beta
| File | Platform | Checksum |
|---|---|---|
| eddacraft-anvil-aarch64-apple-darwin.tar.xz | Apple Silicon macOS | checksum |
| eddacraft-anvil-x86_64-apple-darwin.tar.xz | Intel macOS | checksum |
| eddacraft-anvil-aarch64-pc-windows-msvc.zip | ARM64 Windows | checksum |
| eddacraft-anvil-x86_64-pc-windows-msvc.zip | x64 Windows | checksum |
| eddacraft-anvil-aarch64-unknown-linux-gnu.tar.xz | ARM64 Linux | checksum |
| eddacraft-anvil-x86_64-unknown-linux-gnu.tar.xz | x64 Linux | checksum |