Allow Sandbox container limits overrides

Containerfile

@@ -164,6 +164,17 @@ ENV PATH="${JAVA_HOME}/bin:${PNPM_HOME}/bin:${GOROOT}/bin:${GOPATH}/bin:${CARGO_
 # NODE_PATH is the fallback once that climb reaches the filesystem root.
 ENV NODE_PATH=/usr/local/lib/node_modules
 
+# GRADLE_OPTS: bounds the Gradle daemon's own heap. Gradle's default heap
+# auto-sizing goes off whatever memory it sees as "available" -- with no
+# cap it sizes against the Makefile's --memory container ceiling in full,
+# leaving nothing for the separate Kotlin-compiler daemon Gradle forks
+# alongside it (see the RUN_FLAGS comment in the Makefile for the full
+# story: an uncapped build here doesn't just OOM the container, it can
+# starve the whole podman-machine VM). This only bounds the Gradle daemon
+# itself -- a project needing the Kotlin daemon capped too should set
+# `kotlin.daemon.jvmargs` in its own gradle.properties.
+ENV GRADLE_OPTS="-Xmx1024m -XX:MaxMetaspaceSize=256m"
+
 # Fixed HOME: the container is run with --userns=keep-id, so the runtime UID
 # is whatever host user invoked `podman run`, which has no /etc/passwd entry
 # inside the image. A plain ENV HOME is unaffected by that UID mapping, so we

Makefile

@@ -18,9 +18,36 @@ STATE_DIR    := $(HOME)/.local/state/opencode
 # inotify file-change events don't propagate into the container. Not a
 # blocker for this sandbox, just a rough edge if a tool relies on watching
 # /workspace for live changes.
+#
+# --memory/--memory-swap/--cpus: on macOS, `podman machine`'s Linux VM has a
+# FIXED resource allocation (check with `podman machine inspect`; currently
+# 4 GiB / 6 CPUs, bumped up from the 2 GiB default -- see `podman machine
+# set --memory <MB> podman-machine-default`, which requires the machine to
+# be stopped first). Without a cgroup cap here, a memory-hungry build tool
+# (Gradle's daemon and its separate Kotlin-compiler daemon are the worst
+# offenders -- both size their heap off "available" memory with no idea
+# they're sharing a small VM) can consume the VM's *entire* RAM. That
+# doesn't just OOM-kill the container -- it starves the VM's own guest
+# kernel, which stalls the very processes Podman needs to stay responsive
+# (its API socket, gvproxy, SSH), making the whole machine appear to hang.
+# Capping the container forces a contained OOM kill instead. --memory-swap
+# equal to --memory disables extra swap for the container specifically, so
+# it fails fast rather than thrashing. Sized to leave ~1 GiB / 2 CPUs of
+# headroom below the VM's own allocation for the VM/podman/gvproxy itself --
+# adjust alongside `podman machine set` if you resize the machine again.
+# Override any of these per-invocation by exporting OPENCODE_SANDBOX_MEMORY,
+# OPENCODE_SANDBOX_MEMORY_SWAP, and/or OPENCODE_SANDBOX_CPUS before running
+# `make opencode`/`make shell`.
+OPENCODE_SANDBOX_MEMORY      ?= 3072m
+OPENCODE_SANDBOX_MEMORY_SWAP ?= 3072m
+OPENCODE_SANDBOX_CPUS        ?= 4
+
 RUN_FLAGS := \
 	--rm -it \
 	--userns=keep-id \
+	--memory=$(OPENCODE_SANDBOX_MEMORY) \
+	--memory-swap=$(OPENCODE_SANDBOX_MEMORY_SWAP) \
+	--cpus=$(OPENCODE_SANDBOX_CPUS) \
 	--network $(NETWORK_NAME) \
 	--network-alias $(WORKDIR_NAME) \
 	-v $(CURDIR):/workspace \

README.md

@@ -86,10 +86,11 @@ Bumping a pinned LSP version means editing the corresponding `ARG`/`FROM` tag in
   Because these are real host paths shared by every container regardless of workdir, logging into OpenCode once carries over to every project you use the sandbox in.
 
 - Both run targets join a shared, auto-created Podman network (`opencode-sandbox-net`) under a `--network-alias` derived from the current directory's name, so a container started from another dev-container setup on the same network can reach this one by that name (and vice versa).
+- Both run targets also cap the container to `--memory=3072m --memory-swap=3072m --cpus=4`. Without this, a heavy JVM build tool (Gradle's daemon plus its separate Kotlin-compiler daemon are the worst offenders — both auto-size their heap off "available" memory with no idea they're sharing a fixed-size VM) can consume the entire `podman machine`'s RAM. That starves the VM's own guest kernel, which stalls the processes Podman itself needs to stay responsive — so the *whole machine* hangs, not just the container. The cap forces a contained OOM kill instead. Sized to leave headroom below the VM's own allocation for the VM/podman/gvproxy itself — adjust both together if you resize the machine (`podman machine set --memory <MB> --cpus <N> podman-machine-default`, machine must be stopped first). The `Containerfile` also sets a default `GRADLE_OPTS=-Xmx1024m` so the Gradle daemon itself doesn't try to claim the whole container ceiling; a project whose Kotlin-compiler daemon also needs bounding should set `kotlin.daemon.jvmargs` in its own `gradle.properties`. These defaults can be overridden per-invocation by exporting `OPENCODE_SANDBOX_MEMORY`, `OPENCODE_SANDBOX_MEMORY_SWAP`, and/or `OPENCODE_SANDBOX_CPUS` before running `make opencode`/`make shell` (e.g. `OPENCODE_SANDBOX_MEMORY=4096m OPENCODE_SANDBOX_CPUS=6 make shell`).
 
 ## macOS-specific notes
 
-- Podman on macOS runs everything inside a Linux VM (`podman machine`, backend `libkrun`/`applehv`, virtiofs file sharing). The Makefile detects this and starts a stopped machine automatically before `build`, `opencode`, or `shell`.
+- Podman on macOS runs everything inside a Linux VM (`podman machine`, backend `libkrun`/`applehv`, virtiofs file sharing). The Makefile detects this and starts a stopped machine automatically before `build`, `opencode`, or `shell`. This machine has a fixed CPU/memory allocation (`podman machine inspect`) — bump it (see above) if JVM-heavy builds are still tight even with the container's own cap.
 - SELinux volume-relabeling suffixes (`:Z`/`:z`) are intentionally omitted — macOS's host filesystem has no SELinux concept, so they'd be a no-op.
 - Known virtiofs limitation: inotify/file-watch events from the host don't propagate into the container. Not a blocker for normal use, just worth knowing if a tool inside the container tries to watch `/workspace` for live host-side changes.