Module sysinfo 

Sysinfo data collection loop: collects CPU, memory, and network metrics and publishes them via the MPS broker. Sampling interval is configurable via the telemetry:interval setting (0.1s–2.0s, default 1.0s).

Structs

NetState 🔒

Network I/O tracking state for rate calculations.

Constants

BYTES_PER_GB 🔒

BYTES_PER_MB 🔒

DEFAULT_INTERVAL_SECS 🔒

DISK_FULL_REFRESH 🔒

How often the disk list and each volume’s capacity/free space are re-read. Every tick only refreshes I/O counters (the read/write rates need a delta per tick). A full refresh re-lists every mounted volume and, on macOS, asks each for its “available for important usage” capacity, which goes through CacheDelete: with a few DMGs mounted that was thousands of system calls and log lines a minute, for numbers that move slowly.

HOLDER_BADGES 🔒

HOLDER_JANITOR 🔒

HOLDER_TELEMETRY 🔒

MAX_INTERVAL_SECS 🔒

MIN_INTERVAL_SECS 🔒

NEVER_STRETCHED 🔒

A budget no run exceeds: the source keeps its interval whatever a run costs.

PANE_JANITOR_MS 🔒

How often the panes’ processes are scanned with no badge watching: enough to evict a pane whose shell exited without normal cleanup.

PERSIST_COUNT 🔒

SCHEDULER_COALESCE_MS 🔒

Sources due within this much of each other run in one pass.

Statics

PROCESS_START_MS 🔒

Baseline reading of suspend_aware_now_ms captured at process start by mark_process_start (main.rs).

Functions

available_commit_gb

Available system commit headroom in GB (Windows: GlobalMemoryStatusEx → ullAvailPageFile). Returns None where there’s no cheap commit figure (non-Windows) or the read fails — callers treat None as “no limit / admit”. This is the admission-control signal for Pillar 3 (commit-aware agent spawn): see agents::runner::admit_spawn and SPEC_WIN10_PAGEFILE_OOM_CRASH.

get_commit_data 🔒

Read the Windows commit budget via GlobalMemoryStatusEx. Emits mem:commit:used and mem:commit:total (in GB). No-op on non-Windows — keys are simply absent from the payload.

get_cpu_data 🔒

Collect CPU usage (total + per-core).

get_disk_data 🔒

Collect disk I/O rates (in MB/s). sysinfo Disk::usage() returns deltas (bytes since last refresh) so we divide by elapsed time to get rates.

get_interval_secs 🔒

Read the telemetry interval from config, clamped to [MIN, MAX].

get_mem_data 🔒

Collect memory metrics (in GB).

get_pagefile_volume_data 🔒

Windows watch volume: the drive backing the page file. Emits the platform-neutral disk:watch:free_gb / :total_gb / :free_pct, plus disk:pagefile_system_managed (1.0/0.0), which stays Windows-named because it is a page-file fact and its ABSENCE is what tells the status bar it is not on Windows (see the colour tri-state in SPEC_STATUSBAR_DISK_PILL_CROSS_PLATFORM_2026_09_18 §3.3).

get_watch_volume_data 🔒

Unix watch volume: the mount backing AgentMux’s data dir — where the store, logs and packaged builds land, so it is the volume whose filling up actually breaks this app.

mark_process_start

Record this process’s start for uptime reporting. Idempotent; the first call wins, so a stray second call can’t restart the clock.

pick_watch_mount 🔒

Pick the mount whose path is the longest prefix of target.

run_sysinfo_loop

Run the sysinfo collection loop. Uses tokio::time::interval for steady tick rate regardless of refresh duration. Interval is re-read from config each tick and the timer is reset if it changes.

suspend_aware_now_ms 🔒

uptime_secs

Seconds since mark_process_start. Falls back to initializing on first read, so a caller that never marked startup reports a truthful (if late-based) 0 rather than a wrong number or a panic. saturating_sub makes a decrease structurally impossible even if a platform counter ever misbehaved.

write_watch_summary 🔒

Summarise one already-collected volume as the watch target.