Overview: show the host and its live stats — CPU, memory, microVMs, disk (closes #205)

[ParallelEntrepreneur]

What changed

GET /api/status now returns a top-level host object, and the overview renders it.

Backend (crates/colonizer/src/runtime.rs, main.rs, sessions.rs)

• New Host facts probe in runtime.rs: id (stable per-install uuid), hostname, cpu_cores, memory_total_bytes/memory_used_bytes, load (1/5/15 min), uptime_secs, disk_total_bytes/disk_used_bytes/disk_free_bytes (of the data dir), and checked_at. Measurements come from /proc/meminfo, /proc/loadavg, /proc/uptime, /proc/sys/kernel/hostname (fallback: bounded hostname exec), std::thread::available_parallelism(), and one bounded df -kP <data_dir> exec. No new crate dependencies.
• Cached exactly like Runtime: same 10 s TTL constant, ?fresh=1 honoured, lock held across the probe so concurrent polls share one run, new host_cache slot on App (initialized in both the real and test constructors). A warm cache spawns nothing per request; the probe joins the existing status() tokio::join!.
• Omit-don't-fake: every measurable field is Option with skip_serializing_if, so a machine that can't measure something (e.g. macOS has no /proc) omits the key rather than zero-filling it.
• host_json() (pure, unit-tested) merges in the request-time facts: microvms_live (sessions matching the same busy predicate as queue admission — is_live() || Publishing, factored as SessionStatus::busy()), microvms_ceiling (orgs::global_max_parallel, the exact value admission uses), and kvm_ok (rides runtime.kvm, omitted off-Linux) — so a host that cannot boot colonies is visibly broken, not idle.
• Host id: persisted at <config_dir>/host_id (0600, created once). Deliberately not telemetry's install_id, which is ephemeral by design (forgotten when the live map is switched off). Everything is keyed by that id so a second machine can be aggregated later without an API rewrite — the single-host host object becomes a list entry, not a reshape.
• Session gains boot_cpus/boot_memory, recorded from the actual BootSpec in boot_inner right after it is built (so they survive a later boot-phase failure). The container-level #[serde(default)] keeps existing sessions.json loadable. Guest CPU %/RSS are not shipped: microsandbox/agentd expose no such metrics (agentd serves only health/events/pty/shutdown) — noted in a code comment next to the fields rather than faked.

Frontend (web/src)

• types.ts: new HostInfo interface matching the wire shape exactly; HarnessStatus.host; Session.boot_cpus/boot_memory (and boot_timing, already on the wire, now typed for rendering).
• New cockpit/host.ts: pure formatters (formatBytes re-exports the existing diskSize, formatUptime, formatLoad, formatBootMs) plus hostFacts()/colonyFacts() which drop any segment whose inputs are missing — the strip never renders zeros or undefined.
• OverviewView.tsx: a compact host strip above the org grid — hostname · IconServer live/ceiling · IconCpu cores+load · IconMemory used/total · disk free/total · uptime · checked <ago> — flex flex-wrap at font-mono text-[11px]/text-faint, degrades cleanly at the 300 px minimum card width; warns when kvm_ok === false. Colony rows gained a second meta line with only the facts that exist (boot cpus/memory, boot total, mesh ip, agent); pre-change colonies render exactly as before.
• Prop threaded from App.tsx's existing 30 s /api/status poll through Cockpit.tsx — no new fetch or poll.
• mock.ts: mockHost() so ?mock=1 exercises the strip; ?runtime=mac omits kvm_ok, ?runtime=kvm reports it false.

Docs: docs/protocol.md §4 documents the host contract (always-present id/checked_at/microVM counts; every other key omitted when unmeasurable) and the new Session fields.

Size — over the soft ceiling, deliberately

git diff --stat: 10 tracked files, +706/−24, plus 2 new files (web/src/cockpit/host.ts 102 lines, host.test.ts 155) — ≈963 changed lines across 12 files. The task is inherently full-stack (new probe + cache + wire contract + Session persistence + UI + mock + protocol docs), and roughly 40 % of the Rust diff is tests the repo's conventions call for (parser unit tests, cache TTL/fresh tests, host_json shape, host_id persistence, old-sessions.json round-trip, a real status-handler assertion). Trimming tests to hit 400 would have meant shipping unmeasured parsers.

Sibling-overlap heads-up: hot shared files touched here are crates/colonizer/src/main.rs (App struct + status()), crates/colonizer/src/sessions.rs (Session struct/Default/create + SessionStatus::busy()), web/src/types.ts, web/src/mock.ts, web/src/cockpit/OverviewView.tsx, web/src/cockpit/Cockpit.tsx (1 line), docs/protocol.md. New scaffolding kept minimal: no new crate or module registration — the only new files are the two web/src/cockpit/host.*, and the only new persisted state is <config_dir>/host_id. usage.rs moved by 2 lines (mechanical: new fields in a test Session literal).

Verification

• cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings: clean.
• cargo test --workspace: 447 passed, 0 failed — includes a real end-to-end status-handler test producing a genuine host payload on this machine.
• npm test (web): 306 passed (15 files), including the new host.test.ts suite.
• npm run build: tsc --noEmit + vite clean (only the pre-existing >900 kB chunk warning).
• An independent review pass re-ran everything and fixed one doc bug (the protocol.md example had shown raw df 1K-block counts instead of bytes).

Reviewer pointers

• status() computes the host value before the json! literal, because the literal moves runtime while host_json borrows runtime.kvm — easy to regress.
• Confirm the live count matches admission: SessionStatus::busy() (sessions.rs) vs the predicate in queue::has_room.
• Unit scaling in the parsers: meminfo kB ×1024, df -kP 1K-blocks ×1024, used = total − available guarded against available > total.
• Known nits (accepted, not fixed): df -P output would garble if the filesystem device name contains spaces (all-None fallback, no crash); /proc reads are blocking inside the async probe, matching the existing probe_kvm precedent; ?runtime=old in the mock still includes a host object (mock-only realism gap).

Closes #205

---

🤖 Generated by Colonizer in a microVM