The new TestUpAttachedTerminatesOnExternalStop (#13990) occasionally
observed `up` exit 1 instead of 0 after an external stop/down cancels
a service's restart backoff. The cause: once the monitor detects
termination and cancels globalCtx, a log-streaming goroutine still
blocked on logs?follow=1 for the backoff container gets a
context.Canceled error, which appendErr collected unfiltered into the
final errors.Join.
That race could only surface once #13990 made the monitor loop
actually return in this scenario -- before that fix, up hung forever
and never reached this code path.
appendErr now drops an error that is nothing more than fallout from
our own shutdown (errors.Is(err, context.Canceled) once
u.globalCtx.Err() is set), the same way a SIGINT-triggered teardown's
residual goroutine errors are already expected to be harmless. A
genuine error is still reported regardless of shutdown state.
Signed-off-by: Nicolas De Loof <nicolas.deloof@gmail.com>
compose commit --dry-run delegated ContainerCommit to the real client
and actually created the image. It now returns a dry-run ID without
touching the engine.
Auditing the rest of the client surface: 39 more mutating operations
were delegated — none that compose itself issues (verified caller by
caller), but any future code path reaching one would mutate the engine
under --dry-run silently. They now refuse with an explicit error.
The classification — read-only delegated / faked / refused — was only
implied by a method's position under a comment; it is now declared in
dryrunclient_test.go, which parses the client and fails on any method
missing, misclassified, or new and unclassified: extending APIClient
requires an explicit dry-run decision.
The commit e2e scenario gains a --dry-run step asserting the image is
NOT created (new ImageAbsent check).
Epic #14074, B.7.
Signed-off-by: Nicolas De Loof <nicolas.deloof@gmail.com>
Unit tests alone didn't cover the exact failure: an attached up
hanging when the project is stopped/removed externally while a
service sits in restart backoff. The new test reproduces that
sequence end to end, so a future regression on this path breaks
loudly instead of slipping through as a passing unit suite. A
RestartCount guard narrows the race with the backoff expiring, so a
missed window fails loudly instead of passing silently through the
pre-existing die-based path.
Signed-off-by: Guillaume Lours <glours@users.noreply.github.com>
Since 2.39.3 the monitor relies solely on die events to stop tracking
containers, but a container stopped while in restart backoff emits only
stop and destroy events, so an attached `up` hangs forever after an
external `stop`/`down`. Handle destroy as terminal, and on stop inspect
the container to distinguish a definitive stop from the transient one
emitted during a ContainerRestart, which keeps watch sync+restart
(#13161) working.
A plain stop and its die aren't ordered, so evicting as soon as a stop
looks definitive can end the loop before an already-queued die is
drained, losing the exit code --exit-code-from and
--abort-on-container-exit depend on. Eviction on stop now requires a
die already observed for that container; otherwise it defers to the
pending die.
Fixes#13985
Signed-off-by: Guillaume Lours <glours@users.noreply.github.com>
`docker compose port SERVICE PORT` picks the first matching publisher
as its answer, but containerPublishers()'s sort.Slice does not
guarantee tie order, so a dual-stack publish (same target port on
both an IPv4 and an IPv6 address) could non-deterministically return
either one. Switch to sort.SliceStable to preserve the daemon's
original order for ties.
Also drops a dead length check and a duplicated host:port formatter.
Signed-off-by: Guillaume Lours <glours@users.noreply.github.com>
A rebuild whose build context was snapshotted before a newer change to
one of its own services is doomed: its output will be replaced by the
trailing rebuild anyway, and meanwhile a sync+restart change landing in
that window was silently lost — the sync went to the doomed container,
the restart was skipped, and the replacement image predated the edit.
The scheduler now gives each run its own cancellable context and
upholds one invariant: every request ends up covered by a rebuild whose
snapshot postdates it. A request naming an actively-rebuilt service
interrupts the run and folds its whole active set back into pending
(interruption kills the other services' rebuild too), so the trailing
rebuild redoes everything on a fresh snapshot. Requests for other
services keep coalescing without interrupting. handleWatchBatch drops a
restart only when the rebuild is pending (that one converges on its
own) and folds it into a Request when the rebuild is in flight —
recreate + start is the restart intent, converged. The run loop also
stops draining once the watch context is cancelled, which silences the
spurious "Build failed: context canceled" on Ctrl-C.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Nicolas De Loof <nicolas.deloof@gmail.com>
compose watch triggered one rebuild per debounced batch, so a burst of
edits arriving faster than the debounce window (e.g. from a coding
agent editing several files) queued up several rebuilds in a row
instead of a single consolidated one.
Rebuilds now run asynchronously through a small scheduler that
coalesces requests received while a rebuild is in flight into one
trailing rebuild, without adding latency to an isolated edit. This
also closes a race where a plain restart could run concurrently with
an asynchronous rebuild of the same service.
Signed-off-by: Guillaume Lours <glours@users.noreply.github.com>
The github.com/AlecAivazis/survey/v2 module was archived and is no longer
maintained. Looking at the code, we didn't really use most of its features;
survey mostly provided terminal input handling around a simple yes/no
question.
Replace the interactive confirmation implementation with a small local
prompt reader while keeping the exported prompt types and existing
non-terminal prompt behavior unchanged. Preserve the confirmation hint,
default handling, retries for invalid input, and Ctrl+C handling.
Interactive prompts continue to use raw terminal input so Ctrl+C is handled
by the prompt instead of requiring a second interrupt to terminate the
process.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
A plain line printed outside any operation lands below the tracked
block, so the block is no longer directly above the cursor. Reset the
screen so the next operation repaints a fresh block instead of moving
the cursor over rows that shifted — the same reason unsuspend resets
after buildkit output.
Suggested-by: Guillaume Lours <705411+glours@users.noreply.github.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Nicolas De Loof <nicolas.deloof@gmail.com>
- layoutFrame clips its output to the terminal height: height==1 no
longer overflows with header + more-marker
- the two status-to-color switches merge into eventColor(status, def)
- the clear-suspension sequence extracts into unsuspend()
- tests drive a manually-advanced fake clock through the injected
now(), and the cancel test loses its settling sleep — Done waits for
the refresh goroutine itself
- roots, per-parent children and the completed count are maintained
incrementally by apply(), so a 100ms tick with an unchanged tree no
longer rescans every task
- the frame<0 clamp in spinGlyph drops: time comes from the writer's
single clock and never regresses; the invariant is documented on the
model
- Start's retire-previous-cycle comment spells out that cancellation
only guarantees the old goroutine will exit, not that it already has
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Nicolas De Loof <nicolas.deloof@gmail.com>
Transposes the applicable findings from the review of the ttyWriter fix
(#14119) to the rewritten renderer:
- Done waits for the refresh goroutine to exit (outside the lock), so
nothing can repaint after Done returns and output printed right after
an operation cannot be garbled by a stray frame
- a nested Start retires the previous cycle instead of leaking its
refresh goroutine until the parent context is cancelled
- build suspension no longer leaks across cycles: Start resets it and
starts a fresh block below whatever buildkit wrote
- drop the dead strings.ToLower in startDependencies: the unexported
start only reads projectName when no Project is passed
The remaining findings are structural to this rewrite: Done is a single
critical section, and per-cycle context cancellation replaces the
doneSignal handshake.
Signed-off-by: Nicolas De Loof <nicolas.deloof@gmail.com>
The "[+] op N/M" header was the one line bypassing renderSegs, so on a
very narrow terminal it could wrap and desync the cursor arithmetic the
rest of the design guarantees against. Route it through the same clip
and cover degenerate widths (8, 12 cells) in the invariant test.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Nicolas De Loof <nicolas.deloof@gmail.com>
The TTY renderer accumulated display-corruption fixes (truncation of
details, then progress sizes, then task ids; timer alignment; rune-based
measurement) that each patched one symptom of the same structural gap:
nothing guaranteed a rendered line fits the terminal, and once a line
wraps, cursor arithmetic desyncs and the block corrupts.
Replace it with three separable units:
- tty_model.go: pure event reducer with an injected clock, preserving
first-parent-wins updates, monotonic progress and header counters
- tty_layout.go: pure (model, size, now) -> lines function; all widths
are measured in terminal cells (go-runewidth, so CJK is correct) on
plain text before coloring, and every line is clipped to the terminal
width by construction, status text included
- tty_screen.go: diff-based repaint; unchanged rows are skipped,
identical frames write nothing, a frame is a single Write; a shrinking
terminal abandons the block instead of moving the cursor over reflowed
rows
The writer coordinates them behind a mutex and stops the refresh
goroutine through context cancellation, so Done cannot block when the
operation context was cancelled first (Ctrl-C during pull). The spinner
frame is derived from the clock instead of advancing on every call, and
truncation can no longer split multi-byte runes. Visual output is
unchanged: the snapshot test reproduces the previous renderer's golden
output character for character.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Nicolas De Loof <nicolas.deloof@gmail.com>
Compose silently ignored several spec-valid attributes it doesn't
honor outside Swarm mode: deploy.update_config, credential_spec,
ports[].mode: host, and others had no visible effect on a plain
`docker compose up`.
Detection is wired into compose-go's loader as a fail-closed
allowlist: every attribute path the schema declares, minus the
handful this runtime deliberately doesn't implement, plus a few
value-conditional checks for attributes only unsupported for one
specific value. A newly-specified attribute gets flagged
automatically, with no corresponding code change needed here.
Depends on an unreleased compose-go change; go.mod carries a
temporary replace to a fork branch until it's released.
Signed-off-by: Guillaume Lours <glours@users.noreply.github.com>
The fake clock is only used by tests to control the debounce timer,
but requires carrying a clock abstraction through production code.
Use the standard time package for the debounce ticker and use
testing/synctest to control time in tests instead.
This removes the clock from composeService and BatchDebounceEvents,
and eliminates the clockwork dependency from the watch implementation.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
Providers could not see the definition of the service they manage:
options had to be duplicated between the compose file and the provider,
or the provider had to re-resolve the model on its own.
A provider may now emit {"type": "get-service-config"} on stdout;
compose answers on the provider's stdin with one JSON line holding the
resolved canonical configuration of the provider's own service,
straight from the in-memory model. The message can be repeated; each
occurrence is answered with one line. Detection is by construction: a
compose that predates the message aborts on it and never writes to
stdin, so the provider treats EOF as 'unsupported, upgrade compose'.
The example provider demonstrates the round trip, backed by an e2e
scenario; unit tests drive executePlugin against a helper-process
provider and cover the injection.
Signed-off-by: Nicolas De Loof <nicolas.deloof@gmail.com>
display.Mode resolves "auto" against stderr — the stream Compose's own
progress renders to — so `compose build >/dev/null` in an interactive
terminal resolves to "tty". Bake reuses that mode for a display writing
to stdout, and buildkit's progressui hard-fails in tty mode when the
stream is not a console.
Pass the unresolved "auto" to bake unless the user explicitly asked for
tty: progressui then probes the actual output stream and degrades to
plain by itself when stdout is redirected. An explicit --progress=tty
keeps failing loudly on a non-terminal, as `buildx bake` does. This
also revives the BUILDKIT_PROGRESS override in doBuildBake, which only
applies to AutoMode and could no longer trigger.
Fixes#14182
Signed-off-by: Nicolas De Loof <nicolas.deloof@gmail.com>
full diff: https://github.com/containerd/containerd/compare/v2.3.4...v2.3.5
Security Updates
- [**CVE-2026-53495**](https://github.com/containerd/containerd/security/advisories/GHSA-7jxh-36q5-gcqv)
- [**GHSA-rp3h-jf77-q9p4**](https://github.com/containerd/containerd/security/advisories/GHSA-rp3h-jf77-q9p4)
Image Distribution
- Apply hardening to strip sensitive authentication headers when fetching descriptor URLs
Runtime
- Avoid hangs and data races when streaming container standard I/O in CRI
- Fix missing error messages in OpenTelemetry trace attributes
- Fix user and group lookup failures in container rootfs containing symlinked /etc/passwd or /etc/group
- Fix configuration loading error when drop-in configuration files have a higher version than the root configuration
- Avoid containerd startup hangs when loading shims
- Add context to error when shim delete times out
- Fix Windows Server 2022 container compatibility on host builds newer than the latest LTSC
Snapshotters
- Fix unpack failure for EROFS images containing the erofs OS feature
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
`down --rmi` removed every dangling image for the project unconditionally,
while its tagged-image counterpart (ImagesToPrune) already skips a service
no longer in the project unless --remove-orphans is set. `down --rmi local`
on a project with an orphaned service therefore kept the tagged image but
deleted its dangling ones - an inconsistent result for one command.
Extract the "list a project's dangling images, remove what a keep predicate
doesn't spare, tolerate individual failures" logic - which down.go and
watch.go's pruneDanglingImagesOnRebuild had each partially and divergently
implemented - into one shared composeService.removeDanglingImages,
parallelized via errgroup like the rest of the codebase's bulk removals.
down.go's keep now mirrors ImagesToPrune's own RemoveOrphans/GetService
check; watch.go keeps its existing built-IDs check, behavior unchanged.
The removal fan-out now also honors --parallel via eg.SetLimit(s.maxConcurrency),
matching push.go/pull.go, instead of always running unbounded. Fixed
TestWatch_Sync's composeService{} literal to set maxConcurrency: -1
(NewComposeService's own default) - the zero value means "allow zero
concurrent goroutines" to errgroup, not "unlimited", and was silently
relying on maxConcurrency being unused before this change.
Signed-off-by: hiroto.toyoda <hiroto.toyoda@dena.com>
`down --rmi` removed every dangling image for the project unconditionally,
while its tagged-image counterpart (ImagesToPrune) already skips a service
no longer in the project unless --remove-orphans is set. `down --rmi local`
on a project with an orphaned service therefore kept the tagged image but
deleted its dangling ones - an inconsistent result for one command.
Extract the "list a project's dangling images, remove what a keep predicate
doesn't spare, tolerate individual failures" logic - which down.go and
watch.go's pruneDanglingImagesOnRebuild had each partially and divergently
implemented - into one shared composeService.removeDanglingImages,
parallelized via errgroup like the rest of the codebase's bulk removals.
down.go's keep now mirrors ImagesToPrune's own RemoveOrphans/GetService
check; watch.go keeps its existing built-IDs check, behavior unchanged.
Signed-off-by: hiroto.toyoda <hiroto.toyoda@dena.com>
`bridge convert -o <dir>` unconditionally ran `os.RemoveAll` on the
output directory and discarded any error, so a typo (`-o .`, `-o
$PWD`, `-o ~`) silently destroyed unrelated user data.
An empty or missing output directory is still created silently, but a
non-empty one now requires explicit confirmation (prompted
interactively, or via the new `--yes` flag for scripts/CI) before
being wiped.
Signed-off-by: Guillaume Lours <glours@users.noreply.github.com>
full diff: https://github.com/golang/crypto/compare/v0.55.0...v0.56.0
We have tagged version v0.56.0 of golang.org/x/crypto in
order to address the following security issues:
- ssh: prevent DoS on deadlocked established channel
Previously, after a channel has been established, a
malicious peer could send crafted messages that would
deadlock the entire connection.
Now, we handle all RFC 4254 channel messages; global
requests are handled explicitly. Then, treat all other
messages as a protocol error and tear the connection
down instead of buffering and blocking.
Thanks to Will Mortensen for reporting this issue.
This is CVE-2026-56855 and Go issue https://go.dev/issue/81317.
- ssh: prevent DoS on deadlocked undecided channel
Previously, a channel registered in the mux's chanList is
not usable until it is established. A malicious peer was
able flood the channel's incomingRequests, deadlocking the
entire connection.
Now, we add an atomic established state, set when a channel
becomes usable. Until such a time, handlePacket drops every
packet other than the open confirmation/failure, without
blocking and without tearing down the connection.
Thanks to Will Mortensen for reporting this issue.
This is CVE-2026-78662 and Go issue https://go.dev/issue/81316.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
v0.0.6 dropped the Egress policyType/block from generated
NetworkPolicy manifests since it blocked CoreDNS traffic; update
the fixtures to match.
Signed-off-by: Ricardo Branco <rbranco@suse.de>
v0.0.3 had a mis-published multi-arch manifest, causing exec format
errors on non-amd64 architectures.
Fixed upstream in docker/compose-bridge-transformer@665f67e.
Signed-off-by: Ricardo Branco <rbranco@suse.de>
OTel SDK/exporter errors already go through docker/cli's own
otel.SetErrorHandler(debug.OTELErrorHandler), installed by the plugin
framework at startup. Our own handler, installed at package init time,
was always clobbered before any command ran (plugin.Run() sets its handler
after all inits execute), so it never had any effect - verified by tracing
the exact call order and reproducing the flag behavior against a built
binary.
Drop that dead handler along with the COMPOSE_OTEL_DEBUG env var it
introduced. The one genuinely-discarded error (the tracer shutdown/flush
error in wrapRunE) now goes through logrus like everything else in the
codebase: it's already gated by --debug/-D via logrus's own level, so no
explicit check is needed.
Signed-off-by: hiroto.toyoda <hiroto.toyoda@dena.com>
Split SetRecursive into the existing per-GOOS watcher files instead of
a runtime check + nolint: the non-Windows stub never calls the
always-failing SetRecursive, and Windows failures now propagate as a
real error instead of silently downgrading to non-recursive.
Addresses review from thaJeztah on #14164.
Signed-off-by: Guillaume Lours <glours@users.noreply.github.com>