[cnuss]

What

DetectHostCapacity now has a Linux implementation in pkg/agent/hostcapacity_linux.go. The chip is /proc/device-tree/model with its trailing NUL stripped when the board has a device tree, otherwise the first model name or Model line in /proc/cpuinfo, otherwise empty. It goes through the same printable-prefix filter and 128-byte limit as the darwin reader, so the controller accepts whatever it reports. Memory comes from NewHostMemoryProvider().TotalMemory(), the LinuxMemoryProvider from #2531, so a cgroup limit below MemTotal is what the MetalNode shows, the same figure the memory budget uses.

To keep the dead-code guard passing with GOOS=linux, hostCapacityFromSysctl and its test move from the shared files into hostcapacity_darwin.go and hostcapacity_darwin_test.go. The stub for other OSes (!darwin && !linux) now returns an empty HostCapacity directly. macOS behaviour is unchanged.

On main the agent still exits at the Metal check on Linux, so for now this code runs only in tests. #2520 makes it reachable.

Why

Off macOS, DetectHostCapacity reports nothing, so a Linux host's MetalNode would show no chip and no memory in its capacity.

Part of #2481.

Fixes #2518

How

hostChipFromProc takes a readFile function, so the tests can use a fake /proc. They cover a device-tree model with its trailing NUL, the cpuinfo fallback when there is no device tree, and no /proc at all, which gives an empty chip rather than a guess. TestDetectHostCapacityOnLinux runs the real reader and checks that memory is non-zero and no Metal fields are set.

On a real board: a Raspberry Pi 5 (Debian 13, arm64) ran the agent built from the epic branch (7df0d2b9, which carries this slice). Its MetalNode capacity showed chip: "Raspberry Pi 5 Model B Rev 1.1", which matches /proc/device-tree/model, and memory: 8255824Ki, which matches MemTotal. That systemd user unit has no MemoryMax, so nothing capped it.

Under a cgroup limit: a small program calling agent.DetectHostCapacity() in a golang:1.27.4 container under --memory 3g printed MemoryBytes:3221225472, which is the cgroup's memory.max, while /proc/meminfo showed MemTotal: 8122332 kB (about 8 GB). The chip was empty there, because that VM has no device-tree model and its cpuinfo has no model name line.

What I ran on this branch:

• go build ./... and GOOS=linux go build ./...: both build.
• GOOS=linux bin/golangci-lint run ./... and GOOS=darwin bin/golangci-lint run ./...: 0 issues. for each.
• make lint-deadcode and GOOS=linux make lint-deadcode: 0 issues. for each.
• go test -count=1 ./pkg/agent/ ./api/... and go test -count=1 ./cmd/metal-agent/ on macOS: all ok.
• In a golang:1.27.4 container, go test -count=1 ./pkg/agent/ -run 'TestHostChipFromProc|TestDetectHostCapacityOnLinux|TestMetalVersionString|TestLinuxMemory': all pass.
• make chart-crds and make docs-reference: running them a second time changed nothing more, and make check-docs-reference exits 0.
• make gate BASE_REF=origin/main: regen, drift, fmt, copyright, vet and lint passed. The test step's only failure was TestOMLXBindHostOverrideReachesDaemon, a known flake in a path this PR doesn't touch, which passed 3/3 rerun alone.

Checklist

• Tests added/updated
• make test passes locally
• make lint passes locally
• Commit messages follow conventional commits
• All commits are signed off (git commit -s) per DCO
• AI assistance (if any) is disclosed above, per CONTRIBUTING.md
• Documentation updated (if user-facing change)

Miscellaneous fixes

• The MetalNodeCapacity field comments in api/v1alpha1/metalnode_types.go described only a Mac. They now describe any host. memoryBytes is unified memory on a Mac and RAM under the agent's cgroup limit on Linux. chip is the CPU brand string on macOS and the device-tree model or cpuinfo model name on Linux. gpuCores and metalVersion are reported on macOS only. The regenerated CRD, the chart's copy and docs/site/reference/crds.md carry the new descriptions.

Assisted-by: Claude Code (Claude Opus 5.5) split this out of #2482, wrote the description and ran the checks above in my session.

[codecov]

Codecov Report

❌ Patch coverage is 88.23529% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
pkg/agent/hostcapacity_linux.go	88.23%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!