chore: bump all crates to 1.60.1

The 1.60.0 debs were built against glibc 2.39 and never installed on the
bookworm fleet; rather than force-moving the tag, the rebuilt release
ships as 1.60.1. Nothing was published to crates.io at 1.60.0.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

.github/workflows/build-deb.yml

@@ -18,7 +18,15 @@ env:
 jobs:
   build-deb:
     name: Build .deb for ${{ matrix.target }}
-    runs-on: ubuntu-latest
+    # PINNED, not ubuntu-latest: the binaries link against this runner's
+    # (multiarch) glibc, so the runner sets the MINIMUM glibc the .debs demand
+    # at install time. ubuntu-latest moved to 24.04/glibc 2.39 and the packages
+    # stopped installing on Debian 12/RPi OS bookworm (glibc 2.36). 22.04 links
+    # 2.35, which bookworm satisfies. The "enforce glibc floor" step below
+    # turns any future violation into a red build instead of a fleet-wide apt
+    # failure — if this pin ever has to move past bookworm's glibc, that step
+    # is the contract to renegotiate first.
+    runs-on: ubuntu-22.04
     strategy:
       matrix:
         include:
@@ -159,6 +167,22 @@ jobs:
           mkdir -p debs
           cp target/${{ matrix.target }}/debian/*.deb debs/
 
+      - name: Enforce glibc floor (Debian 12 / RPi OS bookworm fleet)
+        run: |
+          # The fleet's oldest supported glibc. A .deb that demands newer libc6
+          # than this will not install on the Pis — fail HERE, not at apt time.
+          FLOOR="2.36"
+          fail=0
+          for deb in debs/*.deb; do
+            req=$(dpkg-deb -f "$deb" Depends | sed -n 's/.*libc6 (>= \([0-9.]*\)).*/\1/p' | head -1)
+            echo "$deb -> libc6 >= ${req:-none}"
+            if [ -n "$req" ] && [ "$(printf '%s\n' "$req" "$FLOOR" | sort -V | tail -1)" != "$FLOOR" ]; then
+              echo "::error::$deb requires libc6 >= $req, exceeding the fleet floor $FLOOR (bookworm). The build runner's glibc is too new — see the runs-on pin comment."
+              fail=1
+            fi
+          done
+          exit $fail
+
       - name: List generated packages
         run: ls -lh debs/
 

.gitignore

@@ -1,6 +1,7 @@
-/target
+# Any crate's build directory, including standalone sub-crates
+# (zellij_socktop_plugin, socktop_wasm_test) that live outside the workspace.
+target/
 .vscode/
-/socktop-wasm-test/target
 /.cargo/
 
 # Documentation files from development sessions (context-specific, not for public repo)

CHANGELOG.md

@@ -0,0 +1,70 @@
+# Changelog
+
+## 1.60.1 — unreleased
+
+Identical to 1.60.0 plus rebuilt Debian packages: the 1.60.0 debs were linked
+against glibc 2.39 (a GitHub runner migration) and would not install on
+Debian 12 / Raspberry Pi OS bookworm. CI now pins the build environment and
+gates every package against the fleet's glibc floor. 1.60.0 was never
+published to crates.io.
+
+Everything since `v1.50.0`. Applies to all three crates (`socktop`, `socktop_agent`, `socktop_connector`), which move to 1.60.1 together.
+
+### Security
+
+- **Certificate pinning is now real.** With `--verify-hostname` off (the default), the client previously accepted *any* server certificate — the `--tls-ca` file was never consulted. The presented certificate must now be byte-identical to one in the pinned PEM (multi-cert files supported for rotation). If you use TLS, update the client: earlier versions are MITM-able despite the pinning documentation. (housekeeping-p2)
+- `key.pem` is created with mode 0600 (was world-readable 0644); agents also tighten existing keys on startup. (housekeeping-p2)
+- The agent's per-PID caches now evict (60s age / 64 entries); previously they grew without bound. (housekeeping-p2)
+
+### Performance
+
+- Agent CPU on GPU machines cut ~6× (measured 23.5 → 4.0 ms/s at default polling): GPU collection moved to a dedicated worker thread that keeps the NVML session open instead of re-initializing it every 1.5 s on the async runtime. (housekeeping-p2)
+- `journalctl` no longer blocks the agent's async workers. (housekeeping-p2)
+- Cached "no temp sensor / no GPU" results count as fresh — no more per-request rescans on hosts without them. (housekeeping-p2)
+- Nagle disabled on all connection paths (small request/response frames). (housekeeping-p2)
+
+### TUI
+
+- **Compact layout for small windows**: when the window is too short for the Disks pane, Disks is dropped, Memory/Swap go side by side, GPU collapses to one line (omitted if absent), and the reclaimed rows keep the CPU graph and per-core bars visible. `--compact` pins it. (#37)
+- **Width-aware text**: header, CPU title, and process table shed detail by priority as the terminal narrows instead of overwriting each other; process Name column is now the last to go, not the first. Fixed sort-header clicks landing up to 4 columns off. (#38)
+- **Responsive input**: keys and mouse are handled within ~30 ms instead of queueing for a full metrics interval. (housekeeping-p2)
+- **No more freezes**: all requests carry a 5 s timeout; a dead connection shows the reconnect modal (with working `q`) instead of hanging the UI. Consecutive timeouts surface a persistent "agent not responding" error. (housekeeping-p2)
+- Old agents without the per-process endpoints once again show "Agent Update Required" instead of a reconnect loop. (housekeeping-p2)
+- Journal pane distinguishes "no entries" from "no journal access" (e.g. user-run/demo agents) and shows journalctl's hint plus the fix. (housekeeping-p2)
+- Scatter-plot axes align correctly for large CPU-time values. (housekeeping-p2)
+- Demo mode explains how to install `socktop_agent` when the binary is missing. (#36)
+
+### Correctness
+
+- Process/child CPU times were sent as ms but displayed as µs — values rendered 1000× too small in the details modal. (housekeeping-p2)
+- Non-Linux per-process CPU% no longer truncates multi-core usage (clamp after divide). (housekeeping-p2)
+- Journal timestamps are real RFC 3339 UTC with numeric sorting (additive `timestamp_us`). (housekeeping-p2)
+- Partition detection uses `/sys/block` on Linux — whole-disk filesystems (`nvme0n1`, `zram1`) are no longer misclassified as partitions. (housekeeping-p2)
+- Network rates use agent-side sample timestamps (additive `sampled_at_ms`), eliminating rate sawtooth from TTL-cached snapshots; falls back to the client clock with older agents. (housekeeping-p2)
+- The details modal's Command/exe/cwd fields are populated again (dropped by an earlier refresh optimization). (housekeeping-p2)
+- Non-ASCII device names no longer panic the disk pane. (housekeeping-p2)
+
+### Wire format (additive only — old/new client-agent pairs keep working)
+
+- `Metrics.sampled_at_ms` (epoch ms of actual collection)
+- `JournalEntry.timestamp_us` (epoch µs), `JournalEntry.timestamp` now RFC 3339
+- `JournalResponse.notice` (journal-access hint)
+
+### Internal / packaging
+
+- ratatui 0.28 → 0.30 (#33); aws-lc-rs advisories patched (#34); Debian packaging for the agent (#25); assorted dependabot bumps.
+- ~3,100 lines of dead code removed, including an orphaned pre-refactor copy of the connector.
+- `socktop` consumes `socktop_connector` via a path+version dep — connector changes are testable in-repo before publishing.
+- wasm examples build against the in-repo connector; note `zellij_socktop_plugin` has pre-existing compile errors and needs its own rework.
+
+### Process kill (PR #40)
+
+- **Kill a local process from the TUI** (`t` on a selected process, or inside Process Details): btop-style Terminate/Force-kill confirmation. Local agents only — the signal is sent by socktop itself with its own privileges, never over the wire; remote agents never show the option. PID-reuse guarded (the confirmed name must still own the PID at signal time).
+- **Agent no longer reports dead processes**: a long-lived sysinfo `System` accumulated every process ever seen (21k+ entries on a 289-process host), inflating memory, per-poll work, and the process count — and keeping killed processes on screen forever. Update agent and client together on machines where the kill feature will be used.
+- Killed rows leave the list when the process actually exits and cannot be resurrected by cached agent snapshots; details views for dead processes close themselves, including through parent-navigation chains.
+- Selection hint no longer vanishes for long process names; confirmation/info dialogs size to their content.
+
+### Upgrade notes
+
+- **Release/publish order**: `socktop_connector` → `socktop` → agent packages.
+- Clients older than 1.60 work against 1.60 agents and vice versa; the security fix is client-side, so prioritize client updates where TLS is used.

Cargo.lock

@@ -2412,7 +2412,7 @@ dependencies = [
 
 [[package]]
 name = "socktop"
-version = "1.50.0"
+version = "1.60.1"
 dependencies = [
  "anyhow",
  "assert_cmd",
@@ -2422,7 +2422,7 @@ dependencies = [
  "ratatui",
  "serde",
  "serde_json",
- "socktop_connector 1.50.0 (registry+https://github.com/rust-lang/crates.io-index)",
+ "socktop_connector",
  "sysinfo",
  "tempfile",
  "tokio",
@@ -2432,7 +2432,7 @@ dependencies = [
 
 [[package]]
 name = "socktop_agent"
-version = "1.50.2"
+version = "1.60.1"
 dependencies = [
  "anyhow",
  "assert_cmd",
@@ -2442,6 +2442,7 @@ dependencies = [
  "futures-util",
  "gfxinfo",
  "hostname",
+ "nvml-wrapper",
  "once_cell",
  "prost",
  "prost-build",
@@ -2463,7 +2464,7 @@ dependencies = [
 
 [[package]]
 name = "socktop_connector"
-version = "1.50.0"
+version = "1.60.1"
 dependencies = [
  "flate2",
  "futures-util",
@@ -2484,27 +2485,6 @@ dependencies = [
  "web-sys",
 ]
 
-[[package]]
-name = "socktop_connector"
-version = "1.50.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "61ea6a5733e71da6d5c94d23265b85f7041305bca51e6c33e7104464444047bc"
-dependencies = [
- "flate2",
- "futures-util",
- "prost",
- "prost-build",
- "protoc-bin-vendored",
- "rustls",
- "rustls-pemfile",
- "serde",
- "serde_json",
- "thiserror 2.0.17",
- "tokio",
- "tokio-tungstenite 0.24.0",
- "url",
-]
-
 [[package]]
 name = "stable_deref_trait"
 version = "1.2.1"

README.md

@@ -416,6 +416,7 @@ Tip: If only the binary changed, restart is enough. If the unit file changed, ru
 
 ```json
 {
+  "sampled_at_ms": 1786752000123,
   "cpu_total": 12.4,
   "cpu_per_core": [11.2, 15.7],
   "mem_total": 33554432,
@@ -475,7 +476,7 @@ socktop --tls-ca /path/to/agent/cert.pem wss://HOST:8443/ws
 Notes:
 - Do not copy the private key off the server; only the cert.pem is needed by clients.
 - When --tls-ca/-t is supplied, the client auto‑upgrades ws:// to wss:// to avoid protocol mismatch.
-- Hostname (SAN) verification is DISABLED by default (the cert is still pinned). Use `--verify-hostname` to enable strict SAN checking.
+- Hostname (SAN) verification is DISABLED by default; instead the client PINS the certificate: the agent must present a cert byte-identical to one in your `--tls-ca` file (expiry is ignored in this mode — you pinned that exact cert). Use `--verify-hostname` to switch to strict chain + SAN validation instead.
 - You can run multiple clients with different cert paths by passing --tls-ca per invocation.
 
 ---

scripts/install.sh

@@ -0,0 +1,246 @@
+#!/usr/bin/env bash
+# Build socktop + socktop_agent from source and install them.
+#
+# Works on Linux (x86_64, arm64/armv7, riscv64) and macOS. Handles fresh
+# installs and upgrades; if a systemd socktop-agent service is present, its
+# binary is replaced in place and the service restarted.
+#
+#   ./scripts/install.sh                     # build HEAD of the repo you're in
+#   ./scripts/install.sh --ref v1.60.0       # build a tag/branch (clones if needed)
+#   ./scripts/install.sh --ref master           # or any branch
+#   ./scripts/install.sh --prefix ~/.local/bin --no-service
+#
+set -euo pipefail
+
+REPO_URL="https://github.com/jasonwitty/socktop.git"
+REF=""
+PREFIX=""
+NO_SERVICE=0
+SRC_DIR="${SOCKTOP_SRC_DIR:-$HOME/.cache/socktop-src}"
+
+while [ $# -gt 0 ]; do
+  case "$1" in
+    --ref)       REF="$2"; shift 2 ;;
+    --prefix)    PREFIX="$2"; shift 2 ;;
+    --no-service) NO_SERVICE=1; shift ;;
+    -h|--help)   grep '^#' "$0" | sed 's/^# \{0,1\}//'; exit 0 ;;
+    *) echo "unknown argument: $1" >&2; exit 2 ;;
+  esac
+done
+
+say()  { printf '\033[1;36m==>\033[0m %s\n' "$*"; }
+warn() { printf '\033[1;33mwarn:\033[0m %s\n' "$*" >&2; }
+die()  { printf '\033[1;31merror:\033[0m %s\n' "$*" >&2; exit 1; }
+
+# The entire remainder runs inside main(), invoked on the LAST line. This
+# makes the script safe against being MODIFIED WHILE RUNNING: when executed
+# from the clone it manages, the git checkout below replaces this very file,
+# and bash reads scripts lazily by byte offset — without this wrapper it
+# resumes parsing the NEW file at the OLD offset and executes an arbitrary
+# tail of it (observed: the fresh-service path ran on a host whose unit
+# already existed). With main(), the whole script is parsed before any of
+# it executes.
+main() {
+
+OS="$(uname -s)"
+ARCH="$(uname -m)"
+
+# ---------- toolchain ----------
+command -v git >/dev/null || die "git is required"
+if ! command -v cargo >/dev/null; then
+  # rustup may be installed but not on PATH in this shell
+  [ -f "$HOME/.cargo/env" ] && . "$HOME/.cargo/env"
+fi
+if ! command -v cargo >/dev/null; then
+  say "Rust toolchain not found — installing via rustup (stable, default profile)"
+  curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --profile minimal
+  . "$HOME/.cargo/env"
+fi
+command -v cc >/dev/null || warn "no C compiler found (apt: build-essential / brew: xcode-select --install) — the build may fail"
+case "$ARCH" in
+  riscv64*)
+    # protoc-bin-vendored ships no riscv64 binary; the build falls back to
+    # the system protoc (see build.rs).
+    command -v protoc >/dev/null || die "riscv64 needs a system protoc: sudo apt install protobuf-compiler"
+    ;;
+esac
+
+# ---------- source ----------
+# If run from inside a socktop checkout and no --ref given, build that tree
+# as-is (whatever is checked out, including local changes).
+if [ -z "$REF" ] && git rev-parse --show-toplevel >/dev/null 2>&1 \
+   && grep -qs '^name = "socktop"' "$(git rev-parse --show-toplevel)/socktop/Cargo.toml" 2>/dev/null; then
+  SRC_DIR="$(git rev-parse --show-toplevel)"
+  say "Building the current checkout: $SRC_DIR ($(git -C "$SRC_DIR" describe --always --dirty 2>/dev/null))"
+else
+  REF="${REF:-master}"
+  if [ ! -d "$SRC_DIR/.git" ]; then
+    say "Cloning $REPO_URL -> $SRC_DIR"
+    git clone "$REPO_URL" "$SRC_DIR"
+  fi
+  say "Checking out $REF"
+  git -C "$SRC_DIR" fetch --tags origin
+  git -C "$SRC_DIR" checkout -q "$REF"
+  # fast-forward when REF is a branch
+  git -C "$SRC_DIR" merge --ff-only "origin/$REF" >/dev/null 2>&1 || true
+fi
+
+# ---------- build ----------
+say "Building release binaries (this can take a while on SBCs)"
+( cd "$SRC_DIR" && cargo build --release -p socktop -p socktop_agent )
+CLIENT="$SRC_DIR/target/release/socktop"
+AGENT="$SRC_DIR/target/release/socktop_agent"
+
+# ---------- install ----------
+if [ -z "$PREFIX" ]; then
+  PREFIX="/usr/local/bin"
+fi
+SUDO=""
+if [ ! -w "$PREFIX" ]; then
+  if command -v sudo >/dev/null; then SUDO="sudo"; else
+    PREFIX="$HOME/.local/bin"; mkdir -p "$PREFIX"
+    warn "no sudo — installing to $PREFIX (ensure it is on your PATH)"
+  fi
+fi
+say "Installing to $PREFIX"
+$SUDO install -m 755 "$CLIENT" "$PREFIX/socktop"
+$SUDO install -m 755 "$AGENT" "$PREFIX/socktop_agent"
+
+# Update every other copy on PATH as well. A stale `cargo install` in
+# ~/.cargo/bin would otherwise SHADOW the fresh binary (~/.cargo/bin
+# usually precedes /usr/local/bin on PATH), leaving `socktop --version`
+# stuck on the old release after a "successful" install.
+update_path_copies() {
+  local name="$1" src="$2" copy dir
+  # type -ap lists every match on PATH (bash builtin, symlinks not resolved)
+  for copy in $(type -ap "$name" | sort -u); do
+    [ "$copy" = "$PREFIX/$name" ] && continue
+    dir="$(dirname "$copy")"
+    say "Updating additional copy on PATH: $copy"
+    if [ -w "$copy" ] || [ -w "$dir" ]; then
+      install -m 755 "$src" "$copy"
+    else
+      # Non-fatal: an un-updatable extra copy shouldn't kill the install,
+      # but the user must know it may shadow the fresh binary.
+      $SUDO install -m 755 "$src" "$copy"         || warn "could not update $copy — it may shadow $PREFIX/$name"
+    fi
+  done
+}
+update_path_copies socktop "$CLIENT"
+update_path_copies socktop_agent "$AGENT"
+
+# ---------- systemd service (Linux only) ----------
+# System-level operations (unit files, users, service control) need root no
+# matter where the binaries were installed — decide independently of PREFIX.
+SYS_SUDO=""
+if [ "$(id -u)" -ne 0 ]; then
+  if command -v sudo >/dev/null; then SYS_SUDO="sudo"; else SYS_SUDO="__none__"; fi
+fi
+if [ "$SYS_SUDO" = "__none__" ] && [ "$NO_SERVICE" -eq 0 ]; then
+  warn "no sudo available — skipping systemd service management"
+  NO_SERVICE=1
+fi
+if [ "$OS" = "Linux" ] && [ "$NO_SERVICE" -eq 0 ] && command -v systemctl >/dev/null; then
+  if systemctl cat socktop-agent.service >/dev/null 2>&1; then
+    # UPGRADE: the unit file is the operator's (SSL, tokens, ports may be
+    # configured there) — never overwrite it. Only the binary it points at
+    # is replaced, then the service is restarted.
+    say "Existing socktop-agent.service found — preserving unit file, refreshing binary"
+    UNIT_BIN="$(systemctl show -p ExecStart socktop-agent.service 2>/dev/null \
+                | sed -n 's/.*path=\([^ ;]*\).*/\1/p' | head -1)"
+    if [ -n "$UNIT_BIN" ] && [ "$UNIT_BIN" != "$PREFIX/socktop_agent" ]; then
+      $SYS_SUDO systemctl stop socktop-agent.service
+      $SYS_SUDO install -m 755 "$AGENT" "$UNIT_BIN"
+      $SYS_SUDO systemctl start socktop-agent.service
+    else
+      $SYS_SUDO systemctl restart socktop-agent.service
+    fi
+  else
+    # FRESH INSTALL: unit + the system user it runs as + its state dir,
+    # then enable and start. Mirrors the deb package's postinst and
+    # https://www.socktop.io/assets/docs/installation/agent-service.html
+    say "No socktop-agent.service found — installing and enabling it"
+
+    if ! getent group socktop >/dev/null; then
+      $SYS_SUDO groupadd --system socktop
+    fi
+    if ! getent passwd socktop >/dev/null; then
+      NOLOGIN="$(command -v nologin || echo /usr/sbin/nologin)"
+      $SYS_SUDO useradd --system -g socktop -d /var/lib/socktop -M -s "$NOLOGIN" socktop
+    fi
+    $SYS_SUDO mkdir -p /var/lib/socktop
+    $SYS_SUDO chown socktop:socktop /var/lib/socktop
+    $SYS_SUDO chmod 755 /var/lib/socktop
+
+    UNIT_TMP="$(mktemp)"
+    if [ -f "$SRC_DIR/docs/socktop-agent.service" ]; then
+      cp "$SRC_DIR/docs/socktop-agent.service" "$UNIT_TMP"
+    else
+      # Fallback for refs that predate docs/socktop-agent.service
+      cat > "$UNIT_TMP" <<'UNIT'
+[Unit]
+Description=Socktop agent
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+ExecStart=/usr/local/bin/socktop_agent --port 3000
+Environment=RUST_LOG=info
+# Optional auth:
+# Environment=SOCKTOP_TOKEN=changeme
+# TLS (self-signed cert on first run, default port 8443):
+# Environment=SOCKTOP_ENABLE_SSL=1
+Restart=on-failure
+User=socktop
+Group=socktop
+NoNewPrivileges=true
+
+[Install]
+WantedBy=multi-user.target
+UNIT
+    fi
+    # Pick the agent port: 3000 by default, but NEVER bind onto a port that
+    # something else already holds (e.g. Gitea/Umami and friends love 3000)
+    # — that puts the fresh service straight into a crash-restart loop.
+    AGENT_PORT=""
+    for p in 3000 3001 3010 3231 3232; do
+      if ! ss -tln 2>/dev/null | awk '{print $4}' | grep -q ":${p}\$"; then
+        AGENT_PORT="$p"
+        break
+      fi
+    done
+    if [ -z "$AGENT_PORT" ]; then
+      AGENT_PORT=3000
+      warn "no free port among the defaults — using 3000; edit the unit if the service fails to start"
+    elif [ "$AGENT_PORT" != "3000" ]; then
+      warn "port 3000 is already in use by another service — configuring the agent on port $AGENT_PORT"
+    fi
+
+    # Point ExecStart at wherever this run installed the agent, on the chosen port.
+    sed -i.bak -e "s|^ExecStart=[^ ]*socktop_agent|ExecStart=$PREFIX/socktop_agent|" \
+               -e "s|--port [0-9]*|--port $AGENT_PORT|" "$UNIT_TMP"
+    rm -f "$UNIT_TMP.bak"
+
+    $SYS_SUDO install -o root -g root -m 0644 "$UNIT_TMP" /etc/systemd/system/socktop-agent.service
+    rm -f "$UNIT_TMP"
+    $SYS_SUDO systemctl daemon-reload
+    $SYS_SUDO systemctl enable --now socktop-agent.service
+    say "Service installed — agent URL: ws://$(hostname):$AGENT_PORT/ws"
+    say "To enable TLS or a token, edit /etc/systemd/system/socktop-agent.service, then: sudo systemctl daemon-reload && sudo systemctl restart socktop-agent"
+  fi
+  sleep 1
+  systemctl --no-pager -l status socktop-agent.service | head -5 || true
+fi
+
+say "Installed:"
+"$PREFIX/socktop" --version
+"$PREFIX/socktop_agent" --version
+say "Active on PATH: $(type -p socktop || true) / $(type -p socktop_agent || true)"
+socktop --version
+
+}
+
+# exit in the same parse unit as the call: after main returns, bash must not
+# read another byte from this (possibly replaced) file.
+main "$@"; exit $?

socktop/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "socktop"
-version = "1.50.0"
+version = "1.60.1"
 authors = ["Jason Witty <jasonpwitty+socktop@proton.me>"]
 description = "Remote system monitor over WebSocket, TUI like top"
 edition = "2024"
@@ -11,7 +11,7 @@ repository = "https://github.com/jasonwitty/socktop"
 
 [dependencies]
 # socktop connector for agent communication
-socktop_connector = "1.50.0"
+socktop_connector = { version = "1.60.1", path = "../socktop_connector" }
 
 tokio = { workspace = true }
 futures-util = { workspace = true }
@@ -22,8 +22,10 @@ ratatui = { workspace = true }
 crossterm = { workspace = true }
 unicode-width = { workspace = true }
 anyhow = { workspace = true }
-dirs-next = { workspace = true }
+# Local process signalling only (src/proc_kill.rs). The TUI never gathers its
+# own metrics — everything on screen comes from the agent over the connector.
 sysinfo = { workspace = true }
+dirs-next = { workspace = true }
 
 [dev-dependencies]
 assert_cmd = "2.0"

socktop/src/local.rs

@@ -0,0 +1,85 @@
+//! Detection of whether the connected agent is running on this same machine.
+//!
+//! Process-kill is only offered for *local* agents. The reasoning is a
+//! security one: the PIDs shown in the UI are reported by the agent, and when
+//! the user asks to kill one, socktop sends the signal with its OWN local OS
+//! privileges (a direct syscall — never over the network; see [`crate::proc_kill`]).
+//! A PID is therefore only meaningful — and only safe to act on — when the
+//! agent lives on this machine. If we acted on a remote agent's PIDs we would
+//! be signalling whatever unrelated *local* process happened to share that
+//! number.
+//!
+//! An address is considered local when it is loopback, or when we can bind an
+//! ephemeral socket to it: a bind only succeeds for an address assigned to one
+//! of this host's own network interfaces, so it also covers the case of an
+//! agent reached over this machine's LAN IP. Detection fails closed — any
+//! parse/resolution failure, or any resolved address that is not local,
+//! disables the feature.
+
+use std::net::{IpAddr, ToSocketAddrs, UdpSocket};
+
+/// Returns true only if the agent reached at `ws_url` is on this machine.
+pub fn agent_is_local(ws_url: &str) -> bool {
+    let Ok(parsed) = url::Url::parse(ws_url) else {
+        return false;
+    };
+    match parsed.host() {
+        // IP literals can be checked directly without any name resolution.
+        Some(url::Host::Ipv4(ip)) => ip_is_local(IpAddr::V4(ip)),
+        Some(url::Host::Ipv6(ip)) => ip_is_local(IpAddr::V6(ip)),
+        // A hostname (e.g. "localhost", or a LAN name) must resolve, and every
+        // address it resolves to must be local. ws=80, wss=443 are the known
+        // default ports; an explicit port in the URL is honored.
+        Some(url::Host::Domain(domain)) => {
+            let port = parsed.port_or_known_default().unwrap_or(0);
+            match (domain, port).to_socket_addrs() {
+                Ok(addrs) => {
+                    let mut saw_any = false;
+                    for addr in addrs {
+                        saw_any = true;
+                        if !ip_is_local(addr.ip()) {
+                            return false;
+                        }
+                    }
+                    saw_any
+                }
+                Err(_) => false,
+            }
+        }
+        None => false,
+    }
+}
+
+/// An address is local if it is loopback, or if we can bind an ephemeral
+/// socket to it (only possible for an address on one of our own interfaces).
+/// Port 0 requests an ephemeral port and sends no traffic.
+fn ip_is_local(ip: IpAddr) -> bool {
+    ip.is_loopback() || UdpSocket::bind((ip, 0)).is_ok()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::agent_is_local;
+
+    #[test]
+    fn loopback_hosts_are_local() {
+        assert!(agent_is_local("ws://127.0.0.1:3000/ws"));
+        assert!(agent_is_local("ws://localhost:3000/ws"));
+        assert!(agent_is_local("ws://[::1]:3000/ws"));
+        assert!(agent_is_local("wss://127.0.0.1/ws"));
+    }
+
+    #[test]
+    fn public_addresses_are_not_local() {
+        // 8.8.8.8 is not assigned to any local interface.
+        assert!(!agent_is_local("ws://8.8.8.8:3000/ws"));
+        // Documentation-range address, guaranteed not bound locally.
+        assert!(!agent_is_local("ws://203.0.113.1:3000/ws"));
+    }
+
+    #[test]
+    fn garbage_fails_closed() {
+        assert!(!agent_is_local("not a url"));
+        assert!(!agent_is_local(""));
+    }
+}

socktop/src/main.rs

@@ -2,6 +2,8 @@
 
 mod app;
 mod history;
+mod local;
+mod proc_kill;
 mod profiles;
 mod retry;
 mod types;
@@ -321,10 +323,15 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
 
     let is_tls = url.starts_with("wss://");
     let has_token = url.contains("token=");
+    // Only enable local process-kill when the agent is verified to be on this
+    // machine; otherwise on-screen PIDs refer to a remote host and acting on
+    // them locally would signal the wrong process. See local::agent_is_local.
+    let is_local = local::agent_is_local(&url);
     let mut app = App::new()
         .with_intervals(metrics_interval_ms, processes_interval_ms)
         .with_status(is_tls, has_token)
-        .with_compact(parsed.compact);
+        .with_compact(parsed.compact)
+        .with_local(is_local);
     if parsed.dry_run {
         return Ok(());
     }
@@ -404,7 +411,12 @@ async fn run_demo_mode(
         }
         Err(e) => return Err(e.into()),
     };
-    let mut app = App::new().with_compact(compact);
+    // Demo mode runs the real agent on loopback, so its PIDs are real local
+    // processes — enable the local process-kill feature, gated the same way as
+    // the normal connect path (loopback resolves local).
+    let mut app = App::new()
+        .with_compact(compact)
+        .with_local(local::agent_is_local(&url));
     // Demo mode connects to localhost, so disable hostname verification
     tokio::select! { res=app.run(&url,None,false)=>{ drop(child); res } _=tokio::signal::ctrl_c()=>{ drop(child); Ok(()) } }
 }

socktop/src/proc_kill.rs

@@ -0,0 +1,181 @@
+//! Local process termination.
+//!
+//! Signals are sent by socktop itself, using this process's own OS privileges,
+//! via a direct `sysinfo` call. Nothing is transmitted to the agent — the
+//! agent and connector have no kill capability at all. This code path is only
+//! reachable once the agent has been verified to be local (see
+//! [`crate::local`]), which guarantees the PID refers to a process on this
+//! machine.
+
+use sysinfo::{ProcessRefreshKind, ProcessesToUpdate, Signal, System};
+
+/// The signals socktop can send. Deliberately limited to the two btop-style
+/// primaries; no arbitrary-signal chooser.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum KillSignal {
+    /// SIGTERM — polite request to terminate.
+    Term,
+    /// SIGKILL — forceful, cannot be caught.
+    Kill,
+}
+
+impl KillSignal {
+    fn as_sysinfo(self) -> Signal {
+        match self {
+            KillSignal::Term => Signal::Term,
+            KillSignal::Kill => Signal::Kill,
+        }
+    }
+
+    /// Human-facing label for confirmation/result messages.
+    pub fn label(self) -> &'static str {
+        match self {
+            KillSignal::Term => "SIGTERM",
+            KillSignal::Kill => "SIGKILL",
+        }
+    }
+}
+
+/// Is `pid` still a live local process?
+///
+/// A zombie counts as gone: after a kill the entry can linger until the parent
+/// reaps it, and showing a row for a process that no longer runs is exactly the
+/// staleness this check exists to avoid.
+pub fn process_exists(pid: u32) -> bool {
+    let spid = sysinfo::Pid::from_u32(pid);
+    let mut sys = System::new();
+    sys.refresh_processes_specifics(
+        ProcessesToUpdate::Some(&[spid]),
+        false,
+        ProcessRefreshKind::nothing(),
+    );
+    match sys.process(spid) {
+        Some(p) => p.status() != sysinfo::ProcessStatus::Zombie,
+        None => false,
+    }
+}
+
+/// Send `signal` to local process `pid`. Returns `Ok(())` on success, or an
+/// `Err` with a human-readable reason (process gone, PID reused, permission
+/// denied, signal unsupported on this platform).
+///
+/// `expected_name`, when given, is compared against the process that owns the
+/// PID **right now**: the PID came from an agent snapshot and the confirmation
+/// dialog can sit open indefinitely, so by signal time the kernel may have
+/// recycled the number for an unrelated process. Both names come from the
+/// same sysinfo source, so a live, unchanged target compares equal.
+pub fn kill_local_process(
+    pid: u32,
+    expected_name: Option<&str>,
+    signal: KillSignal,
+) -> Result<(), String> {
+    let spid = sysinfo::Pid::from_u32(pid);
+
+    // Refresh just this one PID — we don't need a full process scan to signal it.
+    let mut sys = System::new();
+    sys.refresh_processes_specifics(
+        ProcessesToUpdate::Some(&[spid]),
+        false,
+        ProcessRefreshKind::nothing(),
+    );
+
+    let Some(proc_) = sys.process(spid) else {
+        return Err(format!("Process {pid} no longer exists"));
+    };
+
+    if let Some(expected) = expected_name {
+        let current = proc_.name().to_string_lossy();
+        if current != expected {
+            return Err(format!(
+                "PID {pid} now belongs to \"{current}\", not \"{expected}\" — \
+                 not signalling. Reselect the process and try again."
+            ));
+        }
+    }
+
+    match proc_.kill_with(signal.as_sysinfo()) {
+        Some(true) => Ok(()),
+        Some(false) => Err(format!(
+            "Could not send {} to PID {pid} (permission denied?)",
+            signal.label()
+        )),
+        None => Err(format!(
+            "{} is not supported on this platform",
+            signal.label()
+        )),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::process::Command;
+    use std::time::{Duration, Instant};
+
+    /// The path that matters: a real, live, local process must actually receive
+    /// the signal. Exercises the `refresh_processes_specifics` lookup as well —
+    /// if that call does not populate the process map, `sys.process()` returns
+    /// None and a live PID is reported as "no longer exists".
+    #[test]
+    fn signals_a_real_child_process() {
+        let mut child = Command::new("sleep")
+            .arg("30")
+            .spawn()
+            .expect("spawn sleep for the test");
+        let pid = child.id();
+
+        let result = kill_local_process(pid, Some("sleep"), KillSignal::Term);
+
+        // Reap on every path before asserting, so a failing assert cannot leak a
+        // 30s sleep and cannot trip clippy's zombie_processes lint.
+        let deadline = Instant::now() + Duration::from_secs(5);
+        let mut exited = false;
+        while Instant::now() < deadline {
+            if matches!(child.try_wait(), Ok(Some(_))) {
+                exited = true;
+                break;
+            }
+            std::thread::sleep(Duration::from_millis(20));
+        }
+        if !exited {
+            let _ = child.kill();
+        }
+        let _ = child.wait();
+
+        assert!(result.is_ok(), "kill_local_process returned {result:?}");
+        assert!(
+            exited,
+            "SIGTERM was reported sent but the child never exited"
+        );
+    }
+
+    /// The reuse guard: a live PID whose owner does not match the name the
+    /// user confirmed must NOT be signalled. This also proves the name is
+    /// populated under ProcessRefreshKind::nothing() — if it weren't, the
+    /// matching-name test above would fail instead.
+    #[test]
+    fn refuses_a_pid_owned_by_a_different_process() {
+        let mut child = Command::new("sleep")
+            .arg("30")
+            .spawn()
+            .expect("spawn sleep");
+        let pid = child.id();
+
+        let result = kill_local_process(pid, Some("firefox"), KillSignal::Term);
+
+        let _ = child.kill();
+        let _ = child.wait();
+
+        let err = result.expect_err("signalled a process under the wrong name");
+        assert!(err.contains("firefox") && err.contains("sleep"), "{err}");
+    }
+
+    #[test]
+    fn reports_a_pid_that_is_gone() {
+        let mut child = Command::new("true").spawn().expect("spawn true");
+        let pid = child.id();
+        child.wait().expect("reap");
+        // The PID is now free; signalling it must fail cleanly, not panic.
+        assert!(kill_local_process(pid, None, KillSignal::Term).is_err());
+    }
+}

socktop/src/ui/cpu.rs

@@ -589,6 +589,7 @@ mod render_tests {
 
     fn fake_metrics(cores: Vec<f32>) -> Metrics {
         Metrics {
+            sampled_at_ms: None,
             cpu_total: 0.0,
             cpu_per_core: cores,
             mem_total: 1024,

socktop/src/ui/disks.rs

@@ -1,7 +1,8 @@
 //! Disk cards with per-device gauge and title line.
 
 use crate::types::Metrics;
-use crate::ui::util::{disk_icon, human, truncate_middle};
+use crate::ui::fit::truncate_middle_cols;
+use crate::ui::util::{disk_icon, human};
 use ratatui::{
     layout::{Constraint, Direction, Layout, Rect},
     style::Style,
@@ -69,7 +70,7 @@ pub fn draw_disks(f: &mut ratatui::Frame<'_>, area: Rect, m: Option<&Metrics>) {
             "{}{}{}{}  {} / {}  ({}%)",
             indent,
             disk_icon(&d.name),
-            truncate_middle(&d.name, (slot.width.saturating_sub(6)) as usize / 2),
+            truncate_middle_cols(&d.name, slot.width.saturating_sub(6) / 2),
             temp_str,
             human(used),
             human(d.total),

socktop/src/ui/fit.rs

@@ -43,6 +43,46 @@ pub fn truncate_cols(s: &str, max: u16) -> String {
     out
 }
 
+/// Shortens `s` to at most `max` columns by cutting the MIDDLE, marking the
+/// cut with `…` — device names like `/dev/nvme0n1p1` keep their distinctive
+/// prefix and suffix. Column- and char-boundary-safe; the byte-slicing
+/// predecessor in `util.rs` panicked on non-ASCII names.
+pub fn truncate_middle_cols(s: &str, max: u16) -> String {
+    if cols(s) <= max {
+        return s.to_string();
+    }
+    if max <= 1 {
+        return truncate_cols(s, max);
+    }
+    // Reserve one column for the ellipsis; split the rest left/right.
+    let left_budget = (max - 1) / 2;
+    let right_budget = max - 1 - left_budget;
+
+    let mut left_end = 0; // byte index
+    let mut used = 0u16;
+    for (i, ch) in s.char_indices() {
+        let w = cols(ch.encode_utf8(&mut [0u8; 4]));
+        if used + w > left_budget {
+            break;
+        }
+        used += w;
+        left_end = i + ch.len_utf8();
+    }
+
+    let mut right_start = s.len();
+    let mut used = 0u16;
+    for (i, ch) in s.char_indices().rev() {
+        let w = cols(ch.encode_utf8(&mut [0u8; 4]));
+        if used + w > right_budget || i < left_end {
+            break;
+        }
+        used += w;
+        right_start = i;
+    }
+
+    format!("{}…{}", &s[..left_end], &s[right_start..])
+}
+
 /// Picks the first (richest) candidate pair that fits side by side in `width` columns
 /// with at least `gap` columns between them.
 ///
@@ -108,6 +148,23 @@ mod tests {
         assert_eq!(truncate_cols("🔒ab", 2), "…");
     }
 
+    /// Middle truncation keeps both ends — the parts that identify a device —
+    /// and must never exceed the budget or split a character.
+    #[test]
+    fn truncate_middle_keeps_both_ends_within_budget() {
+        assert_eq!(truncate_middle_cols("/dev/nvme0n1p1", 20), "/dev/nvme0n1p1");
+        let out = truncate_middle_cols("/dev/nvme0n1p1", 9);
+        assert_eq!(cols(&out), 9);
+        assert!(out.starts_with("/dev"), "{out}");
+        assert!(out.ends_with("1p1"), "{out}");
+        assert!(out.contains('…'), "{out}");
+        // Non-ASCII names must not panic (the old byte-slicing version did).
+        for max in 0..12u16 {
+            let out = truncate_middle_cols("диск-🗄️-данные", max);
+            assert!(cols(&out) <= max.max(1), "{out:?} exceeds {max}");
+        }
+    }
+
     #[test]
     fn pick_pair_takes_the_richest_that_fits() {
         let candidates = [

socktop/src/ui/gpu.rs

@@ -249,6 +249,7 @@ mod render_tests {
 
     fn metrics(gpus: Option<Vec<GpuInfo>>) -> Metrics {
         Metrics {
+            sampled_at_ms: None,
             cpu_total: 0.0,
             cpu_per_core: vec![],
             mem_total: 1024,

socktop/src/ui/modal.rs

@@ -1,6 +1,10 @@
 //! Modal window system for socktop TUI application
 
-use super::theme::MODAL_DIM_BG;
+use super::fit;
+use super::theme::{
+    BTN_EXIT_BG_ACTIVE, BTN_RETRY_BG_ACTIVE, MODAL_BG, MODAL_BORDER_FG, MODAL_DIM_BG, MODAL_FG,
+    MODAL_TITLE_FG,
+};
 use crossterm::event::KeyCode;
 use ratatui::{
     Frame,
@@ -26,6 +30,10 @@ pub struct ModalManager {
     pub help_scroll_offset: usize,
 }
 
+/// Key hints shown under the confirmation buttons. Also sets the minimum
+/// width of that dialog — sizing from the question alone clipped this line.
+const CONFIRM_HINT: &str = "Tab ← → choose  ·  Enter run  ·  Esc cancel";
+
 impl ModalManager {
     pub fn new() -> Self {
         Self {
@@ -83,6 +91,52 @@ impl ModalManager {
         }
         m
     }
+    /// Close the details view for `pid` WHEREVER it sits in the stack.
+    /// Returns whether anything was closed.
+    ///
+    /// Not just the top: killing from inside the details view stacks the
+    /// "Signal sent" Info modal on top of it, and a SIGKILL victim is usually
+    /// confirmed dead on the very next tick — while that Info is still up. A
+    /// top-only check missed the close, and since a gone PID is processed
+    /// once, the details view stayed open (frozen on the dead process's last
+    /// sample) with nothing left to ever close it.
+    ///
+    /// Per-PID matching keeps the parent-navigation property: only the dead
+    /// process's view goes; parent views underneath are other processes that
+    /// may still be alive and close themselves the same way.
+    pub fn close_process_details(&mut self, pid: u32) -> bool {
+        let was_top =
+            matches!(self.stack.last(), Some(ModalType::ProcessDetails { pid: p }) if *p == pid);
+        let before = self.stack.len();
+        self.stack
+            .retain(|m| !matches!(m, ModalType::ProcessDetails { pid: p } if *p == pid));
+        if self.stack.len() == before {
+            return false;
+        }
+        // Mirror pop_modal's focus bookkeeping when the top changed.
+        if was_top && let Some(next) = self.stack.last() {
+            self.active_button = match next {
+                ModalType::ConnectionError { .. } => ModalButton::Retry,
+                ModalType::ProcessDetails { .. } => ModalButton::Ok,
+                ModalType::About => ModalButton::Ok,
+                ModalType::Help => ModalButton::Ok,
+                ModalType::Confirmation { .. } => ModalButton::Confirm,
+                ModalType::Info { .. } => ModalButton::Ok,
+            };
+        }
+        true
+    }
+
+    /// PID of the uppermost ProcessDetails view, looking through any
+    /// Info/Confirmation stacked above it. What the user will land on when
+    /// transient modals are dismissed.
+    pub fn topmost_process_details(&self) -> Option<u32> {
+        self.stack.iter().rev().find_map(|m| match m {
+            ModalType::ProcessDetails { pid } => Some(*pid),
+            _ => None,
+        })
+    }
+
     pub fn update_connection_error_countdown(&mut self, new_countdown: Option<u64>) {
         if let Some(ModalType::ConnectionError {
             auto_retry_countdown,
@@ -110,6 +164,16 @@ impl ModalManager {
                 self.prev_button();
                 ModalAction::None
             }
+            // Kill the process being viewed. `t` rather than `k` because `k`
+            // scrolls the thread table in this modal — and using the same key
+            // here as on the processes pane means one thing to remember.
+            KeyCode::Char('t') | KeyCode::Char('T') => {
+                if let Some(ModalType::ProcessDetails { pid }) = self.stack.last() {
+                    ModalAction::KillSelected(*pid)
+                } else {
+                    ModalAction::None
+                }
+            }
             KeyCode::Char('r') | KeyCode::Char('R') => {
                 if matches!(self.stack.last(), Some(ModalType::ConnectionError { .. })) {
                     ModalAction::RetryConnection
@@ -241,7 +305,16 @@ impl ModalManager {
                 ModalAction::Dismiss
             }
             (Some(ModalType::Confirmation { .. }), ModalButton::Confirm) => ModalAction::Confirm,
-            (Some(ModalType::Confirmation { .. }), ModalButton::Cancel) => ModalAction::Cancel,
+            (Some(ModalType::Confirmation { .. }), ModalButton::ConfirmForce) => {
+                ModalAction::ConfirmForce
+            }
+            (Some(ModalType::Confirmation { .. }), ModalButton::Cancel) => {
+                // Pop here so Enter-on-Cancel behaves like Esc (which pops in
+                // handle_key); the app's Cancel handler can then assume the
+                // modal is already gone.
+                self.pop_modal();
+                ModalAction::Cancel
+            }
             (Some(ModalType::Info { .. }), ModalButton::Ok) => {
                 self.pop_modal();
                 ModalAction::Dismiss
@@ -253,12 +326,29 @@ impl ModalManager {
         self.active_button = match (&self.stack.last(), &self.active_button) {
             (Some(ModalType::ConnectionError { .. }), ModalButton::Retry) => ModalButton::Exit,
             (Some(ModalType::ConnectionError { .. }), ModalButton::Exit) => ModalButton::Retry,
-            (Some(ModalType::Confirmation { .. }), ModalButton::Confirm) => ModalButton::Cancel,
+            // Confirmation cycles through three: the safe affirmative, the
+            // escalated one, then cancel.
+            (Some(ModalType::Confirmation { .. }), ModalButton::Confirm) => {
+                ModalButton::ConfirmForce
+            }
+            (Some(ModalType::Confirmation { .. }), ModalButton::ConfirmForce) => {
+                ModalButton::Cancel
+            }
             (Some(ModalType::Confirmation { .. }), ModalButton::Cancel) => ModalButton::Confirm,
             _ => self.active_button.clone(),
         };
     }
     fn prev_button(&mut self) {
+        // Confirmation has three buttons, so stepping back is not the same as
+        // stepping forward; everything else is a two-way toggle.
+        if let Some(ModalType::Confirmation { .. }) = self.stack.last() {
+            self.active_button = match self.active_button {
+                ModalButton::Confirm => ModalButton::Cancel,
+                ModalButton::ConfirmForce => ModalButton::Confirm,
+                _ => ModalButton::ConfirmForce,
+            };
+            return;
+        }
         self.next_button();
     }
 
@@ -280,6 +370,150 @@ impl ModalManager {
         );
     }
 
+    /// Wrap `text` to at most `width` columns on word boundaries, so a dialog
+    /// can be sized from its content instead of guessing.
+    fn wrap_cols(text: &str, width: u16) -> Vec<String> {
+        let mut lines = Vec::new();
+        let mut current = String::new();
+        for word in text.split_whitespace() {
+            let candidate = if current.is_empty() {
+                word.to_string()
+            } else {
+                format!("{current} {word}")
+            };
+            if fit::cols(&candidate) <= width || current.is_empty() {
+                current = candidate;
+            } else {
+                lines.push(std::mem::take(&mut current));
+                current = word.to_string();
+            }
+        }
+        if !current.is_empty() {
+            lines.push(current);
+        }
+        if lines.is_empty() {
+            lines.push(String::new());
+        }
+        lines
+    }
+
+    /// A centered box just big enough for `message` plus `footer_rows` of
+    /// buttons/hints. Never exceeds the screen, and never gets so narrow that
+    /// the title is clipped.
+    ///
+    /// `min_content_w` is the width the footer needs. Sizing from the message
+    /// alone clipped the key-hint line, which is longer than most questions.
+    fn dialog_rect(area: Rect, message: &str, footer_rows: u16, min_content_w: u16) -> Rect {
+        // 2 border columns + 2 columns of breathing room on each side.
+        const CHROME_W: u16 = 6;
+        const MAX_TEXT_W: u16 = 64;
+        const MIN_TEXT_W: u16 = 24;
+
+        let avail_text = area.width.saturating_sub(CHROME_W).max(1);
+        let text_w = fit::cols(message)
+            .min(MAX_TEXT_W)
+            .min(avail_text)
+            .max(MIN_TEXT_W.min(avail_text));
+        let lines = Self::wrap_cols(message, text_w);
+        let widest = lines
+            .iter()
+            .map(|l| fit::cols(l))
+            .max()
+            .unwrap_or(text_w)
+            .max(min_content_w.min(avail_text));
+
+        let width = (widest + CHROME_W).min(area.width);
+        // borders + blank + message + blank + footer
+        let height = (lines.len() as u16 + footer_rows + 4).min(area.height);
+        Rect {
+            x: area.x + (area.width.saturating_sub(width)) / 2,
+            y: area.y + (area.height.saturating_sub(height)) / 2,
+            width,
+            height,
+        }
+    }
+
+    /// Shared chrome for the small dialogs: themed border, centered message
+    /// with real padding, and the footer row(s) returned for the caller to
+    /// fill with buttons.
+    ///
+    /// The old versions laid their content out over `area` rather than the
+    /// block's inner rect, which put the first line of text on top of the
+    /// border and pushed the buttons against the frame.
+    fn render_dialog_frame(
+        f: &mut Frame,
+        area: Rect,
+        title: &str,
+        message: &str,
+        footer_rows: u16,
+    ) -> Rect {
+        let block = Block::default()
+            .title(
+                Line::from(format!(" {title} ")).style(
+                    Style::default()
+                        .fg(MODAL_TITLE_FG)
+                        .add_modifier(Modifier::BOLD),
+                ),
+            )
+            .borders(Borders::ALL)
+            .border_style(Style::default().fg(MODAL_BORDER_FG))
+            .style(Style::default().bg(MODAL_BG));
+        let inner = block.inner(area);
+        f.render_widget(block, area);
+
+        // Pad one column each side so text never touches the border.
+        let padded = Rect {
+            x: inner.x + 1,
+            y: inner.y,
+            width: inner.width.saturating_sub(2),
+            height: inner.height,
+        };
+        let rows = Layout::default()
+            .direction(Direction::Vertical)
+            .constraints([
+                Constraint::Length(1),           // breathing room under the title
+                Constraint::Min(1),              // message
+                Constraint::Length(1),           // gap above the footer
+                Constraint::Length(footer_rows), // buttons / hints
+            ])
+            .split(padded);
+
+        f.render_widget(
+            Paragraph::new(message)
+                .style(Style::default().fg(MODAL_FG))
+                .alignment(Alignment::Center)
+                .wrap(Wrap { trim: true }),
+            rows[1],
+        );
+        rows[3]
+    }
+
+    /// One button, sized to its label and centered in `area`.
+    fn render_button(f: &mut Frame, area: Rect, label: &str, active: bool, accent: Color) {
+        let style = if active {
+            Style::default()
+                .bg(accent)
+                .fg(MODAL_BG)
+                .add_modifier(Modifier::BOLD)
+        } else {
+            Style::default().fg(accent)
+        };
+        let text = format!(" {label} ");
+        let w = fit::cols(&text).min(area.width);
+        let btn = Rect {
+            x: area.x + (area.width.saturating_sub(w)) / 2,
+            y: area.y,
+            width: w,
+            height: 1,
+        };
+        f.render_widget(
+            Paragraph::new(text)
+                .style(style)
+                .alignment(Alignment::Center),
+            btn,
+        );
+    }
+
     fn render_modal_content(&mut self, f: &mut Frame, modal: &ModalType, data: ProcessModalData) {
         let area = f.area();
         // Different sizes for different modal types
@@ -296,6 +530,13 @@ impl ModalManager {
                 // Help modal uses medium size
                 self.centered_rect(70, 80, area)
             }
+            // Confirmation and Info are one-question dialogs. A fixed 70%x50%
+            // box left a short question floating in a mostly-empty pane, so
+            // these size themselves to their content instead.
+            ModalType::Confirmation { message, .. } => {
+                Self::dialog_rect(area, message, 3, fit::cols(CONFIRM_HINT))
+            }
+            ModalType::Info { message, .. } => Self::dialog_rect(area, message, 1, 16),
             _ => {
                 // Other modals use smaller size
                 self.centered_rect(70, 50, area)
@@ -340,86 +581,64 @@ impl ModalManager {
         confirm_text: &str,
         cancel_text: &str,
     ) {
-        let chunks = Layout::default()
+        // Three buttons + a key hint line.
+        let footer = Self::render_dialog_frame(f, area, title, message, 3);
+        let rows = Layout::default()
             .direction(Direction::Vertical)
-            .constraints([Constraint::Min(1), Constraint::Length(3)])
-            .split(area);
-        let block = Block::default()
-            .title(format!(" {title} "))
-            .borders(Borders::ALL)
-            .style(Style::default().bg(Color::Black));
-        f.render_widget(block, area);
-        f.render_widget(
-            Paragraph::new(message)
-                .style(Style::default().fg(Color::White))
-                .alignment(Alignment::Center)
-                .wrap(Wrap { trim: true }),
-            chunks[0],
-        );
-        let buttons = Layout::default()
+            .constraints([
+                Constraint::Length(1), // buttons
+                Constraint::Length(1), // spacer
+                Constraint::Length(1), // key hints
+            ])
+            .split(footer);
+
+        let cols = Layout::default()
             .direction(Direction::Horizontal)
-            .constraints([Constraint::Percentage(50), Constraint::Percentage(50)])
-            .split(chunks[1]);
-        let confirm_style = if self.active_button == ModalButton::Confirm {
-            Style::default()
-                .bg(Color::Green)
-                .fg(Color::Black)
-                .add_modifier(Modifier::BOLD)
-        } else {
-            Style::default().fg(Color::Green)
-        };
-        let cancel_style = if self.active_button == ModalButton::Cancel {
-            Style::default()
-                .bg(Color::Red)
-                .fg(Color::Black)
-                .add_modifier(Modifier::BOLD)
-        } else {
-            Style::default().fg(Color::Red)
-        };
-        f.render_widget(
-            Paragraph::new(confirm_text)
-                .style(confirm_style)
-                .alignment(Alignment::Center),
-            buttons[0],
+            .constraints([
+                Constraint::Ratio(1, 3),
+                Constraint::Ratio(1, 3),
+                Constraint::Ratio(1, 3),
+            ])
+            .split(rows[0]);
+
+        Self::render_button(
+            f,
+            cols[0],
+            confirm_text,
+            self.active_button == ModalButton::Confirm,
+            BTN_RETRY_BG_ACTIVE,
         );
+        Self::render_button(
+            f,
+            cols[1],
+            "Force kill",
+            self.active_button == ModalButton::ConfirmForce,
+            MODAL_TITLE_FG,
+        );
+        Self::render_button(
+            f,
+            cols[2],
+            cancel_text,
+            self.active_button == ModalButton::Cancel,
+            BTN_EXIT_BG_ACTIVE,
+        );
+
         f.render_widget(
-            Paragraph::new(cancel_text)
-                .style(cancel_style)
+            Paragraph::new(CONFIRM_HINT)
+                .style(Style::default().fg(MODAL_FG).add_modifier(Modifier::DIM))
                 .alignment(Alignment::Center),
-            buttons[1],
+            rows[2],
         );
     }
 
     fn render_info(&self, f: &mut Frame, area: Rect, title: &str, message: &str) {
-        let chunks = Layout::default()
-            .direction(Direction::Vertical)
-            .constraints([Constraint::Min(1), Constraint::Length(3)])
-            .split(area);
-        let block = Block::default()
-            .title(format!(" {title} "))
-            .borders(Borders::ALL)
-            .style(Style::default().bg(Color::Black));
-        f.render_widget(block, area);
-        f.render_widget(
-            Paragraph::new(message)
-                .style(Style::default().fg(Color::White))
-                .alignment(Alignment::Center)
-                .wrap(Wrap { trim: true }),
-            chunks[0],
-        );
-        let ok_style = if self.active_button == ModalButton::Ok {
-            Style::default()
-                .bg(Color::Blue)
-                .fg(Color::White)
-                .add_modifier(Modifier::BOLD)
-        } else {
-            Style::default().fg(Color::Blue)
-        };
-        f.render_widget(
-            Paragraph::new("[ Enter ] OK")
-                .style(ok_style)
-                .alignment(Alignment::Center),
-            chunks[1],
+        let footer = Self::render_dialog_frame(f, area, title, message, 1);
+        Self::render_button(
+            f,
+            footer,
+            "Enter — OK",
+            self.active_button == ModalButton::Ok,
+            BTN_RETRY_BG_ACTIVE,
         );
     }
 
@@ -505,6 +724,9 @@ impl ModalManager {
             "  ↑/↓ ............ Select/navigate processes",
             "  Enter .......... Open Process Details",
             "  x/X ............ Clear selection",
+            "  t .............. Signal selected process — local agent only",
+            "                   (also works inside Process Details; the prompt",
+            "                    offers Terminate/SIGTERM or Force kill/SIGKILL)",
             "  Click header ... Sort by column (CPU/Mem)",
             "  Click row ...... Select process",
             "",
@@ -632,3 +854,254 @@ impl ModalManager {
             .split(vert[1])[1]
     }
 }
+
+#[cfg(test)]
+mod confirm_tests {
+    use super::*;
+
+    fn confirm_modal() -> ModalManager {
+        let mut m = ModalManager::new();
+        m.push_modal(ModalType::Confirmation {
+            title: "Confirm signal".into(),
+            message: "Send a signal to bash (PID 42)?".into(),
+            confirm_text: "Terminate".into(),
+            cancel_text: "Cancel".into(),
+        });
+        m
+    }
+
+    /// The safe option is focused first, so a reflexive Enter terminates rather
+    /// than force-kills.
+    #[test]
+    fn opens_on_the_safe_option() {
+        let mut m = confirm_modal();
+        assert_eq!(m.active_button, ModalButton::Confirm);
+        assert_eq!(m.handle_key(KeyCode::Enter), ModalAction::Confirm);
+    }
+
+    #[test]
+    fn tab_cycles_all_three_buttons_forward() {
+        let mut m = confirm_modal();
+        m.handle_key(KeyCode::Tab);
+        assert_eq!(m.active_button, ModalButton::ConfirmForce);
+        m.handle_key(KeyCode::Tab);
+        assert_eq!(m.active_button, ModalButton::Cancel);
+        m.handle_key(KeyCode::Tab);
+        assert_eq!(m.active_button, ModalButton::Confirm);
+    }
+
+    /// With three buttons, back is not the same as forward — the old
+    /// prev_button just called next_button, which only worked for two.
+    #[test]
+    fn shift_tab_cycles_backward() {
+        let mut m = confirm_modal();
+        m.handle_key(KeyCode::BackTab);
+        assert_eq!(m.active_button, ModalButton::Cancel);
+        m.handle_key(KeyCode::BackTab);
+        assert_eq!(m.active_button, ModalButton::ConfirmForce);
+        m.handle_key(KeyCode::BackTab);
+        assert_eq!(m.active_button, ModalButton::Confirm);
+    }
+
+    #[test]
+    fn force_kill_reports_its_own_action() {
+        let mut m = confirm_modal();
+        m.handle_key(KeyCode::Tab);
+        assert_eq!(m.handle_key(KeyCode::Enter), ModalAction::ConfirmForce);
+    }
+
+    #[test]
+    fn escape_cancels_and_closes() {
+        let mut m = confirm_modal();
+        assert_eq!(m.handle_key(KeyCode::Esc), ModalAction::Cancel);
+        assert!(!m.is_active());
+    }
+
+    /// Enter on Cancel must behave like Esc, including closing the modal.
+    #[test]
+    fn enter_on_cancel_closes_too() {
+        let mut m = confirm_modal();
+        m.handle_key(KeyCode::Tab);
+        m.handle_key(KeyCode::Tab);
+        assert_eq!(m.handle_key(KeyCode::Enter), ModalAction::Cancel);
+        assert!(!m.is_active());
+    }
+
+    /// `t` inside process details asks the app to raise the kill prompt for the
+    /// process being viewed — not for whatever is selected in the list behind it.
+    #[test]
+    fn t_in_process_details_targets_that_pid() {
+        let mut m = ModalManager::new();
+        m.push_modal(ModalType::ProcessDetails { pid: 4242 });
+        assert_eq!(
+            m.handle_key(KeyCode::Char('t')),
+            ModalAction::KillSelected(4242)
+        );
+    }
+
+    /// `k` still scrolls the thread table, which is why `t` is the kill key.
+    #[test]
+    fn k_in_process_details_still_scrolls() {
+        let mut m = ModalManager::new();
+        m.push_modal(ModalType::ProcessDetails { pid: 1 });
+        m.thread_scroll_max = 5;
+        m.handle_key(KeyCode::Char('j'));
+        assert_eq!(m.thread_scroll_offset, 1);
+        assert_eq!(m.handle_key(KeyCode::Char('k')), ModalAction::Handled);
+        assert_eq!(m.thread_scroll_offset, 0);
+    }
+
+    #[test]
+    fn t_elsewhere_is_not_a_kill() {
+        let mut m = ModalManager::new();
+        m.push_modal(ModalType::Help);
+        assert_eq!(m.handle_key(KeyCode::Char('t')), ModalAction::None);
+    }
+
+    /// A one-line question must not be handed a half-screen box.
+    #[test]
+    fn dialog_is_sized_to_its_content() {
+        let screen = Rect::new(0, 0, 120, 40);
+        let r = ModalManager::dialog_rect(
+            screen,
+            "Send a signal to bash (PID 42)?",
+            3,
+            fit::cols(CONFIRM_HINT),
+        );
+        assert!(r.width < screen.width, "dialog took the full width");
+        assert!(r.height <= 12, "dialog was {} rows tall", r.height);
+        assert!(r.height >= 7, "dialog too short to hold its own footer");
+        // Centered to within the rounding of integer division.
+        let center_delta = (r.x + r.width / 2) as i32 - (screen.width / 2) as i32;
+        assert!(center_delta.abs() <= 1, "off-center by {center_delta}");
+    }
+
+    #[test]
+    fn dialog_never_exceeds_a_small_screen() {
+        let screen = Rect::new(0, 0, 20, 8);
+        let long = "a".repeat(400);
+        let r = ModalManager::dialog_rect(screen, &long, 3, fit::cols(CONFIRM_HINT));
+        assert!(r.width <= screen.width && r.height <= screen.height);
+    }
+}
+
+#[cfg(test)]
+mod button_style_tests {
+    use super::*;
+    use ratatui::Terminal;
+    use ratatui::backend::TestBackend;
+
+    /// The focused button must be the highlighted one — the whole point of the
+    /// three-button layout is that you can see which action Enter will run.
+    #[test]
+    fn focus_moves_the_highlight() {
+        let msg = "Send a signal to bash (PID 42)?";
+        let mut m = ModalManager::new();
+        m.push_modal(ModalType::Confirmation {
+            title: "Confirm signal".into(),
+            message: msg.into(),
+            confirm_text: "Terminate".into(),
+            cancel_text: "Cancel".into(),
+        });
+
+        // Background colors present on the button row, per focused button.
+        let bgs = |m: &ModalManager| -> Vec<Color> {
+            let screen = Rect::new(0, 0, 100, 30);
+            let area = ModalManager::dialog_rect(screen, msg, 3, fit::cols(CONFIRM_HINT));
+            let mut t = Terminal::new(TestBackend::new(100, 30)).unwrap();
+            t.draw(|f| {
+                m.render_confirmation(f, area, "Confirm signal", msg, "Terminate", "Cancel")
+            })
+            .unwrap();
+            let buf = t.backend().buffer();
+            // Buttons sit on the first footer row: title, gap, message, gap.
+            let row = area.y + 4;
+            (area.x..area.x + area.width)
+                .map(|x| buf[(x, row)].bg)
+                .collect()
+        };
+
+        let terminate_focused = bgs(&m);
+        assert!(
+            terminate_focused.contains(&BTN_RETRY_BG_ACTIVE),
+            "Terminate should be highlighted when focused"
+        );
+        assert!(
+            !terminate_focused.contains(&BTN_EXIT_BG_ACTIVE),
+            "Cancel must not be highlighted while Terminate has focus"
+        );
+
+        m.handle_key(KeyCode::Tab);
+        m.handle_key(KeyCode::Tab);
+        let cancel_focused = bgs(&m);
+        assert!(
+            cancel_focused.contains(&BTN_EXIT_BG_ACTIVE),
+            "Cancel should be highlighted after two Tabs"
+        );
+        assert!(
+            !cancel_focused.contains(&BTN_RETRY_BG_ACTIVE),
+            "Terminate must not stay highlighted"
+        );
+    }
+}
+
+#[cfg(test)]
+mod close_details_tests {
+    use super::*;
+
+    #[test]
+    fn closes_the_view_for_that_pid() {
+        let mut m = ModalManager::new();
+        m.push_modal(ModalType::ProcessDetails { pid: 4242 });
+        assert!(m.close_process_details(4242));
+        assert!(!m.is_active());
+    }
+
+    #[test]
+    fn leaves_a_different_pid_alone() {
+        let mut m = ModalManager::new();
+        m.push_modal(ModalType::ProcessDetails { pid: 4242 });
+        assert!(!m.close_process_details(1));
+        assert!(m.is_active());
+    }
+
+    /// Walking up to a parent stacks details views. Only the dead process's
+    /// view goes — whichever position it holds — and the survivor stays put.
+    #[test]
+    fn closes_only_the_dead_pids_view_in_a_parent_chain() {
+        let mut m = ModalManager::new();
+        m.push_modal(ModalType::ProcessDetails { pid: 100 }); // parent
+        m.push_modal(ModalType::ProcessDetails { pid: 200 }); // child, on top
+
+        // Parent dies while the child is viewed: its view is removed from
+        // UNDER the top, so closing the child later lands on the process list
+        // instead of a frozen corpse view.
+        assert!(m.close_process_details(100));
+        assert!(matches!(
+            m.current_modal(),
+            Some(ModalType::ProcessDetails { pid: 200 })
+        ));
+        assert!(m.close_process_details(200));
+        assert!(!m.is_active());
+    }
+
+    /// The F1 regression: killing from inside the details view stacks the
+    /// "Signal sent" Info on top, and the death is usually confirmed while
+    /// that Info is still up. The details view must close anyway — a top-only
+    /// check left it open forever, frozen on the dead process.
+    #[test]
+    fn closes_details_beneath_a_stacked_info_modal() {
+        let mut m = ModalManager::new();
+        m.push_modal(ModalType::ProcessDetails { pid: 7 });
+        m.push_modal(ModalType::Info {
+            title: "Signal sent".into(),
+            message: "Sent SIGKILL".into(),
+        });
+
+        assert!(m.close_process_details(7));
+        // The Info survives on top; dismissing it lands on the process list.
+        assert!(matches!(m.current_modal(), Some(ModalType::Info { .. })));
+        m.pop_modal();
+        assert!(!m.is_active());
+    }
+}

socktop/src/ui/modal_process.rs

@@ -95,7 +95,7 @@ impl ModalManager {
         }
 
         // Help line
-        let help_text = vec![Line::from(vec![
+        let mut help_text = vec![Line::from(vec![
             Span::styled(
                 "X ",
                 Style::default()
@@ -129,6 +129,23 @@ impl ModalManager {
             Span::styled("journal", Style::default().add_modifier(Modifier::DIM)),
         ])];
 
+        // Kill from here too — same key as the processes pane, and only shown
+        // when the agent is local, since that is the only case where it works.
+        if data.is_local
+            && let Some(line) = help_text.first_mut()
+        {
+            line.spans.push(Span::styled(
+                "  t ",
+                Style::default()
+                    .fg(PROCESS_DETAILS_ACCENT)
+                    .add_modifier(Modifier::BOLD),
+            ));
+            line.spans.push(Span::styled(
+                "kill",
+                Style::default().add_modifier(Modifier::DIM),
+            ));
+        }
+
         let help = Paragraph::new(Text::from(help_text))
             .alignment(Alignment::Center)
             .style(Style::default());
@@ -472,13 +489,28 @@ impl ModalManager {
             .borders(Borders::ALL);
 
         let content_lines: Vec<Line> = if journal.entries.is_empty() {
-            vec![
+            let mut lines = vec![
                 Line::from(""),
                 Line::from(Span::styled(
                     "No journal entries found for this process",
                     Style::default().add_modifier(Modifier::DIM),
                 )),
-            ]
+            ];
+            // Access limits, not absence of logs: show journalctl's own hint
+            // (typical when the agent runs as an unprivileged user, e.g. demo
+            // mode) plus the practical fix.
+            if let Some(notice) = &journal.notice {
+                lines.push(Line::from(""));
+                lines.push(Line::from(Span::styled(
+                    format!("⚠ {notice}"),
+                    Style::default().fg(Color::Yellow),
+                )));
+                lines.push(Line::from(Span::styled(
+                    "  Run the agent as a service (or a user in the systemd-journal group) for full journal access.",
+                    Style::default().add_modifier(Modifier::DIM),
+                )));
+            }
+            lines
         } else {
             journal
                 .entries
@@ -624,16 +656,29 @@ impl ModalManager {
         // labels + axis title + (top) Y-axis title + legend + spacing.
         let mut lines: Vec<Line> = Vec::with_capacity(plot_height + 6);
 
-        // Y-axis labels and plot content
-        let mut row_buf = String::with_capacity(plot_width);
-        for y in 0..plot_height {
-            let y_value = params.max_system * (1.0 - (y as f64 / (plot_height - 1).max(1) as f64));
-            // 4-char fixed-width label so the axis doesn't shift as digits change.
-            let y_label = if y_value >= 100.0 {
-                format!("{y_value:>4.0}")
+        // Format a CPU-time value: whole ms once past 100, one decimal below.
+        let fmt_ms = |v: f64| {
+            if v >= 100.0 {
+                format!("{v:.0}")
             } else {
-                format!("{y_value:>4.1}")
-            };
+                format!("{v:.1}")
+            }
+        };
+
+        // Y-axis labels, right-aligned to the widest value this frame so the
+        // axis stays a straight line. The old fixed 4-char field predates the
+        // CPU-time unit fix; honest millisecond values (e.g. 136114) blew
+        // through it and skewed the whole axis.
+        let y_values: Vec<String> = (0..plot_height)
+            .map(|y| {
+                fmt_ms(params.max_system * (1.0 - (y as f64 / (plot_height - 1).max(1) as f64)))
+            })
+            .collect();
+        let y_label_w = y_values.iter().map(|s| s.len()).max().unwrap_or(4).max(4);
+
+        let mut row_buf = String::with_capacity(plot_width);
+        for (y, y_value) in y_values.iter().enumerate() {
+            let y_label = format!("{y_value:>y_label_w$}");
 
             // Build the row's char slice into a reusable String buffer.
             row_buf.clear();
@@ -650,8 +695,8 @@ impl ModalManager {
             ]));
         }
 
-        // Add X-axis
-        let x_axis_padding = "     ".to_string(); // Match Y-axis label width
+        // Add X-axis (padding = Y label width + the space before the bar)
+        let x_axis_padding = " ".repeat(y_label_w + 1);
         let x_axis_line = "─".repeat(plot_width + 1);
         lines.push(Line::from(vec![
             Span::styled(x_axis_padding, Style::default()),
@@ -659,13 +704,14 @@ impl ModalManager {
         ]));
 
         // Add X-axis labels
-        let x_label_start = "0.0".to_string();
-        let x_label_mid = format!("{:.1}", params.max_user / 2.0);
-        let x_label_end = format!("{:.1}", params.max_user);
+        let x_label_start = fmt_ms(0.0);
+        let x_label_mid = fmt_ms(params.max_user / 2.0);
+        let x_label_end = fmt_ms(params.max_user);
 
         let spacing = plot_width / 3;
         let x_labels = format!(
-            "     {}{}{}{}{}",
+            "{}{}{}{}{}{}",
+            " ".repeat(y_label_w + 1),
             x_label_start,
             " ".repeat(spacing.saturating_sub(x_label_start.len())),
             x_label_mid,
@@ -677,7 +723,7 @@ impl ModalManager {
 
         // Add axis titles with better visibility
         lines.push(Line::from(vec![Span::styled(
-            "     User CPU Time (ms) →",
+            format!("{}User CPU Time (ms) →", " ".repeat(y_label_w + 1)),
             Style::default()
                 .fg(Color::Yellow)
                 .add_modifier(Modifier::BOLD),

socktop/src/ui/modal_types.rs

@@ -19,6 +19,9 @@ pub struct ProcessModalData<'a> {
     pub history: ProcessHistoryData<'a>,
     pub max_mem_bytes: u64,
     pub unsupported: bool,
+    /// Whether the agent is on this machine. Only used to decide whether the
+    /// `t` kill hint is shown — the kill itself is gated in `App`.
+    pub is_local: bool,
 }
 
 /// Parameters for rendering scatter plot
@@ -64,9 +67,15 @@ pub enum ModalAction {
     RetryConnection,
     ExitApp,
     Confirm,
+    /// Confirmation modal's second affirmative: the same action, escalated.
+    /// Used by the kill prompt for SIGKILL, where `Confirm` means SIGTERM.
+    ConfirmForce,
     Cancel,
     Dismiss,
     SwitchToParentProcess(u32), // Switch to viewing parent process details
+    /// `t` pressed while viewing a process's details — the app decides whether
+    /// the agent is local and, if so, raises the kill confirmation.
+    KillSelected(u32),
 }
 
 #[derive(Debug, Clone, PartialEq)]
@@ -74,6 +83,10 @@ pub enum ModalButton {
     Retry,
     Exit,
     Confirm,
+    /// Escalated affirmative on a Confirmation modal (SIGKILL for the kill
+    /// prompt). Separate button rather than a separate keybinding so the
+    /// destructive option has to be selected deliberately.
+    ConfirmForce,
     Cancel,
     Ok,
 }

socktop/src/ui/processes.rs

@@ -5,13 +5,14 @@ use ratatui::style::Modifier;
 use ratatui::{
     layout::{Constraint, Direction, Layout, Rect},
     style::{Color, Style},
-    text::Span,
+    text::{Line, Span},
     widgets::{Block, Borders, Paragraph, Scrollbar, ScrollbarOrientation, ScrollbarState, Table},
 };
 use std::cmp::Ordering;
 
 use crate::types::Metrics;
 use crate::ui::cpu::{per_core_clamp, per_core_handle_scrollbar_mouse};
+use crate::ui::fit;
 use crate::ui::theme::{
     PROCESS_SELECTION_BG, PROCESS_SELECTION_FG, PROCESS_TOOLTIP_BG, PROCESS_TOOLTIP_FG, SB_ARROW,
     SB_THUMB, SB_TRACK,
@@ -86,6 +87,9 @@ pub struct ProcessDisplayParams<'a> {
     /// Peak cpu_usage from the most recent cache build; used to bold the
     /// busiest process. -1.0 if no cache.
     pub peak_cpu: f32,
+    /// Agent is on this machine, so the `t` kill hint applies. Without it the
+    /// hint would advertise a key that deliberately does nothing.
+    pub is_local: bool,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
@@ -449,16 +453,60 @@ pub fn draw_top_processes(f: &mut ratatui::Frame<'_>, area: Rect, params: Proces
             format!("PID {selected_pid}")
         };
 
-        let tooltip_text = format!("{process_info} | Enter for details • X to unselect");
-        let tooltip_width = tooltip_text.len() as u16 + 2; // Add padding
-        let tooltip_height = 3;
+        // Key hints, built as spans so the keys read as keys. `t` only appears
+        // for a local agent, since that is the only case where it does anything.
+        let key = Style::default()
+            .fg(PROCESS_TOOLTIP_FG)
+            .add_modifier(Modifier::BOLD);
+        let label = Style::default().fg(PROCESS_TOOLTIP_FG);
+        let mut hints: Vec<Span> = vec![
+            Span::styled("⏎", key),
+            Span::styled(" details", label),
+            Span::styled("  ·  ", label),
+        ];
+        if params.is_local {
+            hints.push(Span::styled("t", key));
+            hints.push(Span::styled(" kill", label));
+            hints.push(Span::styled("  ·  ", label));
+        }
+        hints.push(Span::styled("x", key));
+        hints.push(Span::styled(" unselect", label));
 
-        // Position tooltip at bottom-right of the processes area
-        if area.width > tooltip_width + 2 && area.height > tooltip_height + 1 {
+        let hints_w: u16 = hints.iter().map(|s| fit::cols(&s.content)).sum();
+        // One row, borders on both sides, a space of padding each side.
+        let tooltip_height = 3;
+        let max_w = area.width.saturating_sub(2);
+        if max_w > hints_w + 6 && area.height > tooltip_height + 1 {
+            // The process name is the elastic part: truncate it so the hint
+            // always fits. The old version sized the box from the full string
+            // and skipped rendering entirely when a long process name made it
+            // wider than the pane — so the hint silently vanished exactly when
+            // a long-named process was selected.
+            let room_for_info = max_w - hints_w - 6;
+            let info = fit::truncate_cols(&process_info, room_for_info);
+            let mut spans: Vec<Span> = vec![
+                Span::styled(" ", label),
+                Span::styled(
+                    info.clone(),
+                    Style::default()
+                        .fg(PROCESS_TOOLTIP_FG)
+                        .add_modifier(Modifier::BOLD),
+                ),
+                Span::styled("  │  ", label),
+            ];
+            spans.extend(hints);
+            spans.push(Span::styled(" ", label));
+
+            let width = spans
+                .iter()
+                .map(|s| fit::cols(&s.content))
+                .sum::<u16>()
+                .saturating_add(2)
+                .min(area.width);
             let tooltip_area = Rect {
-                x: area.x + area.width.saturating_sub(tooltip_width + 1),
+                x: area.x + area.width.saturating_sub(width + 1),
                 y: area.y + area.height.saturating_sub(tooltip_height + 1),
-                width: tooltip_width,
+                width,
                 height: tooltip_height,
             };
 
@@ -468,11 +516,10 @@ pub fn draw_top_processes(f: &mut ratatui::Frame<'_>, area: Rect, params: Proces
                     .fg(PROCESS_TOOLTIP_FG),
             );
 
-            let tooltip_paragraph = Paragraph::new(tooltip_text)
-                .block(tooltip_block)
-                .wrap(ratatui::widgets::Wrap { trim: true });
-
-            f.render_widget(tooltip_paragraph, tooltip_area);
+            f.render_widget(
+                Paragraph::new(Line::from(spans)).block(tooltip_block),
+                tooltip_area,
+            );
         }
     }
 
@@ -499,7 +546,6 @@ pub fn draw_top_processes(f: &mut ratatui::Frame<'_>, area: Rect, params: Proces
     }
 }
 
-/// Handle keyboard scrolling (Up/Down/PageUp/PageDown/Home/End)
 /// Parameters for process key event handling
 pub struct ProcessKeyParams<'a> {
     pub selected_process_pid: &'a mut Option<u32>,
@@ -509,16 +555,6 @@ pub struct ProcessKeyParams<'a> {
     pub filtered_indices: &'a [usize],
 }
 
-/// LEGACY: Use processes_handle_key_with_selection for enhanced functionality
-#[allow(dead_code)]
-pub fn processes_handle_key(
-    scroll_offset: &mut usize,
-    key: crossterm::event::KeyEvent,
-    page_size: usize,
-) {
-    crate::ui::cpu::per_core_handle_key(scroll_offset, key, page_size);
-}
-
 pub fn processes_handle_key_with_selection(params: ProcessKeyParams) -> bool {
     use crossterm::event::KeyCode;
 
@@ -598,83 +634,6 @@ pub fn processes_handle_key_with_selection(params: ProcessKeyParams) -> bool {
     }
 }
 
-/// Handle mouse for content scrolling and scrollbar dragging.
-/// Returns Some(new_sort) if the header "CPU %" or "Mem" was clicked.
-/// LEGACY: Use processes_handle_mouse_with_selection for enhanced functionality
-#[allow(dead_code)]
-pub fn processes_handle_mouse(
-    scroll_offset: &mut usize,
-    drag: &mut Option<crate::ui::cpu::PerCoreScrollDrag>,
-    mouse: MouseEvent,
-    area: Rect,
-    total_rows: usize,
-) -> Option<ProcSortBy> {
-    // Inner and content areas (match draw_top_processes)
-    let inner = Rect {
-        x: area.x + 1,
-        y: area.y + 1,
-        width: area.width.saturating_sub(2),
-        height: area.height.saturating_sub(2),
-    };
-    if inner.height == 0 || inner.width <= 2 {
-        return None;
-    }
-    let content = Rect {
-        x: inner.x,
-        y: inner.y,
-        width: inner.width.saturating_sub(2),
-        height: inner.height,
-    };
-
-    // Scrollbar interactions (click arrows/page/drag)
-    per_core_handle_scrollbar_mouse(scroll_offset, drag, mouse, area, total_rows);
-
-    // Wheel scrolling when inside the content
-    crate::ui::cpu::per_core_handle_mouse(scroll_offset, mouse, content, content.height as usize);
-
-    // Header click to change sort
-    let header_area = Rect {
-        x: content.x,
-        y: content.y,
-        width: content.width,
-        height: 1,
-    };
-    let inside_header = mouse.row == header_area.y
-        && mouse.column >= header_area.x
-        && mouse.column < header_area.x + header_area.width;
-
-    if inside_header && matches!(mouse.kind, MouseEventKind::Down(MouseButton::Left)) {
-        // Split the header the same way the draw path did, so a click lands on the
-        // column actually on screen even when PID has been dropped.
-        let columns = ProcColumns::for_width(header_area.width);
-        let cols = Layout::default()
-            .direction(Direction::Horizontal)
-            .constraints(columns.constraints())
-            .spacing(COL_SPACING) // must match Table::column_spacing in the draw path
-            .split(header_area);
-        if let Some(cpu) = columns.cpu_index().map(|i| cols[i])
-            && mouse.column >= cpu.x
-            && mouse.column < cpu.x + cpu.width
-        {
-            return Some(ProcSortBy::CpuDesc);
-        }
-        if let Some(mem) = columns.mem_index().map(|i| cols[i])
-            && mouse.column >= mem.x
-            && mouse.column < mem.x + mem.width
-        {
-            return Some(ProcSortBy::MemDesc);
-        }
-    }
-
-    // Clamp to valid range
-    per_core_clamp(
-        scroll_offset,
-        total_rows,
-        (content.height.saturating_sub(1)) as usize,
-    );
-    None
-}
-
 /// Parameters for process mouse event handling
 pub struct ProcessMouseParams<'a> {
     pub scroll_offset: &'a mut usize,
@@ -948,6 +907,7 @@ mod click_tests {
 
     fn metrics() -> Metrics {
         Metrics {
+            sampled_at_ms: None,
             cpu_total: 0.0,
             cpu_per_core: vec![],
             mem_total: 32_000_000_000,
@@ -992,6 +952,7 @@ mod click_tests {
                         filtered_indices: &idxs,
                         cached_rows: &cache,
                         peak_cpu: peak,
+                        is_local: false,
                     },
                 )
             })
@@ -1003,20 +964,29 @@ mod click_tests {
     }
 
     fn click(width: u16, column: u16) -> Option<ProcSortBy> {
+        let m = metrics();
         let mut scroll = 0usize;
         let mut drag = None;
-        processes_handle_mouse(
-            &mut scroll,
-            &mut drag,
-            MouseEvent {
+        let mut sel_pid = None;
+        let mut sel_idx = None;
+        let idxs = [0usize];
+        processes_handle_mouse_with_selection(ProcessMouseParams {
+            scroll_offset: &mut scroll,
+            selected_process_pid: &mut sel_pid,
+            selected_process_index: &mut sel_idx,
+            drag: &mut drag,
+            mouse: MouseEvent {
                 kind: MouseEventKind::Down(MouseButton::Left),
                 column,
                 row: 1,
                 modifiers: KeyModifiers::NONE,
             },
-            Rect::new(0, 0, width, 8),
-            1,
-        )
+            area: Rect::new(0, 0, width, 8),
+            total_rows: 1,
+            metrics: Some(&m),
+            search_box_visible: false,
+            filtered_indices: &idxs,
+        })
     }
 
     /// The hit-test rects are computed by a separate `Layout` call from the one `Table`
@@ -1062,3 +1032,110 @@ mod click_tests {
         }
     }
 }
+
+#[cfg(test)]
+mod tooltip_tests {
+    use super::*;
+    use ratatui::Terminal;
+    use ratatui::backend::TestBackend;
+    use ratatui::layout::Rect;
+    use socktop_connector::{Metrics, ProcessInfo};
+
+    fn metrics(name: &str) -> Metrics {
+        Metrics {
+            sampled_at_ms: None,
+            cpu_total: 0.0,
+            cpu_per_core: vec![],
+            mem_total: 32_000_000_000,
+            mem_used: 0,
+            swap_total: 0,
+            swap_used: 0,
+            hostname: "t".into(),
+            cpu_temp_c: None,
+            disks: vec![],
+            networks: vec![],
+            top_processes: vec![ProcessInfo {
+                pid: 4242,
+                name: name.into(),
+                cpu_usage: 1.5,
+                mem_bytes: 1_000_000,
+            }],
+            gpus: None,
+            process_count: Some(1),
+        }
+    }
+
+    /// Render the pane with a selection and return the whole buffer as text.
+    fn rendered(name: &str, width: u16, is_local: bool) -> String {
+        let m = metrics(name);
+        let mut cache = Vec::new();
+        let peak = rebuild_row_cache(&m, &mut cache);
+        let idxs = [0usize];
+        let mut terminal = Terminal::new(TestBackend::new(width, 12)).unwrap();
+        terminal
+            .draw(|f| {
+                draw_top_processes(
+                    f,
+                    Rect::new(0, 0, width, 12),
+                    ProcessDisplayParams {
+                        metrics: Some(&m),
+                        scroll_offset: 0,
+                        sort_by: ProcSortBy::CpuDesc,
+                        selected_process_pid: Some(4242),
+                        selected_process_index: Some(0),
+                        search_query: "",
+                        search_active: false,
+                        filtered_indices: &idxs,
+                        cached_rows: &cache,
+                        peak_cpu: peak,
+                        is_local,
+                    },
+                )
+            })
+            .unwrap();
+        let buf = terminal.backend().buffer();
+        let mut out = String::new();
+        for y in 0..12 {
+            for x in 0..width {
+                out.push_str(buf[(x, y)].symbol());
+            }
+            out.push('\n');
+        }
+        out
+    }
+
+    #[test]
+    fn hint_offers_kill_for_a_local_agent() {
+        let out = rendered("some-process", 80, true);
+        assert!(out.contains("details"), "no hint rendered at all:\n{out}");
+        assert!(
+            out.contains("kill"),
+            "local agent should offer kill:\n{out}"
+        );
+        assert!(out.contains("unselect"));
+    }
+
+    /// The key does nothing for a remote agent, so advertising it would be a lie.
+    #[test]
+    fn hint_omits_kill_for_a_remote_agent() {
+        let out = rendered("some-process", 80, false);
+        assert!(out.contains("details"), "no hint rendered at all:\n{out}");
+        assert!(
+            !out.contains("kill"),
+            "remote agent must not offer kill:\n{out}"
+        );
+    }
+
+    /// Regression: the hint used to be sized from the full label including the
+    /// process name, and was skipped entirely when that made it wider than the
+    /// pane — so it vanished exactly when a long-named process was selected.
+    #[test]
+    fn hint_survives_a_very_long_process_name() {
+        let long = "/usr/lib/firefox-esr/firefox-esr-with-a-really-long-suffix";
+        let out = rendered(long, 80, true);
+        assert!(
+            out.contains("details") && out.contains("kill"),
+            "hint disappeared for a long process name:\n{out}"
+        );
+    }
+}

socktop/src/ui/util.rs

@@ -22,19 +22,6 @@ pub fn human(b: u64) -> String {
     format!("{tb:.2}TB")
 }
 
-pub fn truncate_middle(s: &str, max: usize) -> String {
-    if s.len() <= max {
-        return s.to_string();
-    }
-    if max <= 3 {
-        return "...".into();
-    }
-    let keep = max - 3;
-    let left = keep / 2;
-    let right = keep - left;
-    format!("{}...{}", &s[..left], &s[s.len() - right..])
-}
-
 pub fn disk_icon(name: &str) -> &'static str {
     let n = name.to_ascii_lowercase();
     if n.contains(':') {

socktop/tests/profiles.rs

@@ -8,6 +8,7 @@ static ENV_LOCK: Mutex<()> = Mutex::new(());
 #[allow(dead_code)] // touch crate
 fn touch() {
     let _ = socktop::types::Metrics {
+        sampled_at_ms: None,
         cpu_total: 0.0,
         cpu_per_core: vec![],
         mem_total: 0,

socktop_agent/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "socktop_agent"
-version = "1.50.2"
+version = "1.60.1"
 authors = ["Jason Witty <jasonpwitty+socktop@proton.me>"]
 description = "Socktop agent daemon. Serves host metrics over WebSocket."
 edition = "2024"
@@ -10,11 +10,10 @@ homepage = "https://github.com/jasonwitty/socktop"
 repository = "https://github.com/jasonwitty/socktop"
 
 [dependencies]
-# Tokio: Use minimal features instead of "full" to reduce binary size
-# Only include: rt-multi-thread (async runtime), net (WebSocket), sync (Mutex/RwLock), macros (#[tokio::test])
-# Excluded: io, fs, process, signal, time (not needed for this workload)
-# Savings: ~200-300KB binary size, faster compile times
-tokio = { version = "1", features = ["rt-multi-thread", "net", "sync", "macros"] }
+# Tokio: minimal features instead of "full" to reduce binary size.
+# rt-multi-thread (runtime), net (WebSocket), sync (Mutex/oneshot),
+# macros (#[tokio::test]), process (async journalctl).
+tokio = { version = "1", features = ["rt-multi-thread", "net", "sync", "macros", "process"] }
 axum = { version = "0.7", features = ["ws", "macros"] }
 sysinfo = { version = "0.37", features = ["network", "disk", "component"] }
 serde = { version = "1", features = ["derive"] }
@@ -24,6 +23,10 @@ futures-util = "0.3.31"
 tracing = { version = "0.1", optional = true }
 tracing-subscriber = { version = "0.3", features = ["env-filter"], optional = true }
 gfxinfo = { version = "0.1.2", optional = true }
+# Direct NVML fallback for distros that ship only libnvidia-ml.so.1 (Debian
+# and derivatives) — gfxinfo's default init dlopens the unversioned name.
+# Same version gfxinfo already pulls in, so this adds no new build cost.
+nvml-wrapper = { version = "0.10", optional = true }
 once_cell = "1.19"
 axum-server = { version = "0.7", features = ["tls-rustls"] }
 rustls = { version = "0.23", features = ["aws-lc-rs"] }
@@ -36,7 +39,7 @@ time = { version = "0.3", default-features = false, features = ["formatting", "m
 
 [features]
 default = ["gpu"]
-gpu = ["gfxinfo"]
+gpu = ["gfxinfo", "nvml-wrapper"]
 logging = ["tracing", "tracing-subscriber"]
 
 [build-dependencies]

socktop_agent/build.rs

@@ -1,13 +1,15 @@
 fn main() {
-    // Vendored protoc for reproducible builds
-    let protoc = protoc_bin_vendored::protoc_bin_path().expect("protoc");
-
     println!("cargo:rerun-if-changed=proto/processes.proto");
 
     // Compile protobuf definitions for processes
     let mut cfg = prost_build::Config::new();
     cfg.out_dir(std::env::var("OUT_DIR").unwrap());
-    cfg.protoc_executable(protoc); // Use the vendored protoc directly
+    // Vendored protoc for reproducible builds where available. It ships no
+    // riscv64 binary, so on such hosts fall through to $PROTOC / PATH
+    // (prost-build's default lookup) — apt: protobuf-compiler.
+    if let Ok(protoc) = protoc_bin_vendored::protoc_bin_path() {
+        cfg.protoc_executable(protoc);
+    }
     // Use local path (ensures file is inside published crate tarball)
     cfg.compile_protos(&["proto/processes.proto"], &["proto"]) // relative to CARGO_MANIFEST_DIR
         .expect("compile protos");

socktop_agent/src/gpu.rs

@@ -1,6 +1,4 @@
 // gpu.rs
-#[cfg(feature = "gpu")]
-use gfxinfo::active_gpu;
 
 #[derive(Debug, Clone, serde::Serialize)]
 pub struct GpuMetrics {
@@ -10,23 +8,118 @@ pub struct GpuMetrics {
     pub mem_total_bytes: u64,
 }
 
+/// Collect metrics for the active GPU. `None` when there is no usable GPU.
+///
+/// Runs on a dedicated worker thread (see `worker`): gfxinfo's handle holds
+/// an `Rc<Nvml>` (not `Send`), and *creating* it runs a full NVML library
+/// init — ~20ms of blocking work that used to execute on the async runtime
+/// for every collection. The worker owns one handle for the process lifetime,
+/// so steady-state collection is just NVML queries. Measured on an RTX 5080
+/// box, re-initing per collect was ~80% of the agent's entire active CPU.
 #[cfg(feature = "gpu")]
-pub fn collect_all_gpus() -> Result<Vec<GpuMetrics>, Box<dyn std::error::Error>> {
-    let gpu = active_gpu()?; // Use ? to unwrap Result
-    let info = gpu.info();
-
-    let metrics = GpuMetrics {
-        name: gpu.model().to_string(),
-        utilization_gpu_pct: info.load_pct() as u32,
-        mem_used_bytes: info.used_vram(),
-        mem_total_bytes: info.total_vram(),
-    };
-
-    Ok(vec![metrics])
+pub async fn collect_all_gpus() -> Option<Vec<GpuMetrics>> {
+    worker::collect().await
 }
 
 #[cfg(not(feature = "gpu"))]
-pub fn collect_all_gpus() -> Result<Vec<GpuMetrics>, Box<dyn std::error::Error>> {
-    // GPU support not available on this platform
-    Ok(vec![])
+pub async fn collect_all_gpus() -> Option<Vec<GpuMetrics>> {
+    None
+}
+
+#[cfg(feature = "gpu")]
+mod worker {
+    use super::GpuMetrics;
+    use once_cell::sync::OnceCell;
+    use std::sync::mpsc;
+
+    type Reply = tokio::sync::oneshot::Sender<Option<Vec<GpuMetrics>>>;
+    static TX: OnceCell<mpsc::Sender<Reply>> = OnceCell::new();
+
+    pub async fn collect() -> Option<Vec<GpuMetrics>> {
+        let tx = TX.get_or_init(spawn);
+        let (reply_tx, reply_rx) = tokio::sync::oneshot::channel();
+        tx.send(reply_tx).ok()?;
+        reply_rx.await.ok().flatten()
+    }
+
+    fn spawn() -> mpsc::Sender<Reply> {
+        let (tx, rx) = mpsc::channel::<Reply>();
+        std::thread::Builder::new()
+            .name("socktop-gpu".into())
+            .spawn(move || run(rx))
+            .expect("spawn gpu worker thread");
+        tx
+    }
+
+    enum Handle {
+        /// gfxinfo's own detection (AMD sysfs, NVIDIA via unversioned NVML).
+        Gfx(Box<dyn gfxinfo::Gpu>),
+        /// Direct NVML with an explicit versioned soname. Debian & friends
+        /// ship only libnvidia-ml.so.1 (the unversioned symlink lives in the
+        /// dev package), so gfxinfo's default dlopen fails there even though
+        /// the driver is fully functional.
+        Nvml(Box<nvml_wrapper::Nvml>),
+    }
+
+    fn probe() -> Option<Handle> {
+        if let Ok(g) = gfxinfo::active_gpu() {
+            return Some(Handle::Gfx(g));
+        }
+        nvml_wrapper::Nvml::builder()
+            .lib_path(std::ffi::OsStr::new("libnvidia-ml.so.1"))
+            .init()
+            .ok()
+            .map(|nvml| Handle::Nvml(Box::new(nvml)))
+    }
+
+    fn collect_from(handle: &Handle) -> Option<Vec<GpuMetrics>> {
+        match handle {
+            Handle::Gfx(gpu) => {
+                let info = gpu.info();
+                Some(vec![GpuMetrics {
+                    name: gpu.model().to_string(),
+                    utilization_gpu_pct: info.load_pct().clamp(0, 100),
+                    mem_used_bytes: info.used_vram(),
+                    mem_total_bytes: info.total_vram(),
+                }])
+            }
+            Handle::Nvml(nvml) => {
+                let device = nvml.device_by_index(0).ok()?;
+                let mem = device.memory_info().ok()?;
+                Some(vec![GpuMetrics {
+                    name: device.name().unwrap_or_else(|_| "NVIDIA GPU".into()),
+                    utilization_gpu_pct: device
+                        .utilization_rates()
+                        .map(|u| u.gpu.clamp(0, 100))
+                        .unwrap_or(0),
+                    mem_used_bytes: mem.used,
+                    mem_total_bytes: mem.total,
+                }])
+            }
+        }
+    }
+
+    fn run(rx: mpsc::Receiver<Reply>) {
+        let mut handle: Option<Handle> = None;
+        // Probing failed: remember and answer None without re-initing the GPU
+        // stack per request. The agent's negative cache stops asking anyway.
+        let mut probe_failed = false;
+        while let Ok(reply) = rx.recv() {
+            if handle.is_none() && !probe_failed {
+                handle = probe();
+                probe_failed = handle.is_none();
+            }
+            let out = handle.as_ref().and_then(collect_from);
+            // A live GPU cannot report 0 total VRAM; zeros mean the session
+            // died (e.g. driver reload). Drop the handle so the next request
+            // re-probes.
+            if let Some(v) = &out
+                && !v.is_empty()
+                && v.iter().all(|g| g.mem_total_bytes == 0)
+            {
+                handle = None;
+            }
+            let _ = reply.send(out.filter(|v| !v.is_empty()));
+        }
+    }
 }

socktop_agent/src/metrics.rs

@@ -13,49 +13,60 @@ use std::collections::HashMap;
 use std::fs;
 #[cfg(target_os = "linux")]
 use std::io;
-use std::process::Command;
 use std::sync::Mutex;
 use std::time::Duration as StdDuration;
-use std::time::{Duration, Instant, SystemTime, UNIX_EPOCH};
+use std::time::{Duration, SystemTime, UNIX_EPOCH};
 use sysinfo::{ProcessRefreshKind, ProcessesToUpdate};
 #[cfg(feature = "logging")]
 use tracing::warn;
 
 // NOTE: CPU normalization env removed; non-Linux now always reports per-process share (0..100) as given by sysinfo.
 
-// Read (utime, stime) in milliseconds from /proc/{pid}/stat in one go.
-// Returns (0, 0) if the file can't be read.
-//
-// We use `rfind(')')` to step past the `comm` field, which can contain
-// arbitrary characters (including spaces and parens), then index the
-// post-comm fields by position. This is the same trick `read_proc_jiffies`
-// uses below — `split_whitespace().collect::<Vec<_>>()` from the start of
-// the file would mis-parse process names with spaces, and also wastes an
-// allocation per call. Two callers used to read this file twice (once for
-// user, once for system); now it's one syscall per detailed-process record.
+/// Shared parsing for `/proc/<pid>/stat` (and per-thread `task/<tid>/stat`).
+///
+/// The second field, `comm`, can contain arbitrary bytes including spaces and
+/// parentheses, so naive whitespace splitting mis-parses such names. All
+/// callers step past the LAST `')'` and index the remaining space-separated
+/// fields from there: 0 = state, 1 = ppid, 11 = utime, 12 = stime,
+/// 19 = starttime.
 #[cfg(target_os = "linux")]
-fn get_cpu_times_ms(pid: u32) -> (u64, u64) {
+mod procstat {
+    /// Everything after `") "` — the post-comm fields.
+    pub fn after_comm(stat: &str) -> Option<&str> {
+        stat.get(stat.rfind(')')? + 2..)
+    }
+    pub fn field(stat: &str, n: usize) -> Option<&str> {
+        after_comm(stat)?.split_whitespace().nth(n)
+    }
+    /// (utime, stime) in clock ticks.
+    pub fn utime_stime(stat: &str) -> Option<(u64, u64)> {
+        let mut it = after_comm(stat)?.split_whitespace();
+        let utime = it.nth(11)?.parse().ok()?;
+        let stime = it.next()?.parse().ok()?;
+        Some((utime, stime))
+    }
+    /// One clock tick at USER_HZ=100 (universal on Linux) in microseconds.
+    pub const TICK_US: u64 = 10_000;
+}
+
+// Read (utime, stime) in MICROSECONDS from /proc/{pid}/stat in one syscall.
+// Returns (0, 0) if the file can't be read. Units match the wire contract
+// (`DetailedProcessInfo.cpu_time_user` is documented as µs) and the thread
+// records — this used to return ms, making process/child CPU times render
+// 1000x too small next to thread times.
+#[cfg(target_os = "linux")]
+fn get_cpu_times_us(pid: u32) -> (u64, u64) {
     let Ok(s) = fs::read_to_string(format!("/proc/{pid}/stat")) else {
         return (0, 0);
     };
-    let Some(rpar) = s.rfind(')') else {
+    let Some((utime, stime)) = procstat::utime_stime(&s) else {
         return (0, 0);
     };
-    let Some(after) = s.get(rpar + 2..) else {
-        return (0, 0);
-    };
-    let mut it = after.split_whitespace();
-    // Post-comm field offsets: state, ppid, pgrp, session, tty_nr, tpgid,
-    // flags, minflt, cminflt, majflt, cmajflt, utime, stime, ...
-    // utime is offset 11; stime follows.
-    let utime = it.nth(11).and_then(|s| s.parse::<u64>().ok()).unwrap_or(0);
-    let stime = it.next().and_then(|s| s.parse::<u64>().ok()).unwrap_or(0);
-    // 1 tick = 10ms at 100 Hz (USER_HZ).
-    (utime * 10, stime * 10)
+    (utime * procstat::TICK_US, stime * procstat::TICK_US)
 }
 
 #[cfg(not(target_os = "linux"))]
-fn get_cpu_times_ms(_pid: u32) -> (u64, u64) {
+fn get_cpu_times_us(_pid: u32) -> (u64, u64) {
     (0, 0)
 }
 // Runtime toggles (read once)
@@ -117,46 +128,32 @@ fn name_cache_cleanup_threshold() -> usize {
     })
 }
 
-// Tiny TTL caches to avoid rescanning sensors every 500ms
+// Tiny TTL caches to avoid rescanning sensors every 500ms.
+//
+// The cached type is Option<...>: a fresh `None` means "we looked recently
+// and found nothing" — machines with no matching sensor/GPU no longer rescan
+// on every request, only once per TTL.
 const TTL: Duration = Duration::from_millis(1500);
-struct TempCache {
-    at: Option<Instant>,
-    v: Option<f32>,
-}
-static TEMP: OnceCell<Mutex<TempCache>> = OnceCell::new();
+static TEMP: crate::state::TtlCell<Option<f32>> = crate::state::TtlCell::new();
+static GPUS: crate::state::TtlCell<Option<Vec<crate::gpu::GpuMetrics>>> =
+    crate::state::TtlCell::new();
 
-// Last time `state.components` was refreshed (by any caller). Both
+// Gate on `state.components` refreshes (hwmon scans). Both
 // collect_fast_metrics and collect_disks need fresh sensor values; without
-// this gate they were each doing their own `Components::refresh` on their
-// own cadence, paying the hwmon syscall cost twice per polling cycle.
-// 1s is short enough that disk temps stay accurate (they change slowly) and
-// long enough to suppress back-to-back refreshes from concurrent endpoints.
+// this they each paid the hwmon syscall cost on their own cadence. 1s keeps
+// disk temps accurate (they change slowly) while suppressing back-to-back
+// refreshes from concurrent endpoints.
 const COMPONENTS_REFRESH_TTL: Duration = Duration::from_millis(1000);
-static COMPONENTS_LAST_REFRESH: OnceCell<Mutex<Option<Instant>>> = OnceCell::new();
+static COMPONENTS_STAMP: crate::state::TtlCell<()> = crate::state::TtlCell::new();
 
-/// Refresh `state.components` only if the cached refresh timestamp is older
-/// than `COMPONENTS_REFRESH_TTL`. Caller must already hold the components
-/// lock.
+/// Refresh `state.components` at most once per `COMPONENTS_REFRESH_TTL`.
+/// Caller must already hold the components lock.
 fn refresh_components_if_stale(components: &mut sysinfo::Components) {
-    let lock = COMPONENTS_LAST_REFRESH.get_or_init(|| Mutex::new(None));
-    let mut last = match lock.lock() {
-        Ok(g) => g,
-        Err(_) => return, // Poisoned — skip; values stay as-is until next call
-    };
-    let now = Instant::now();
-    let stale = last.is_none_or(|t| now.duration_since(t) >= COMPONENTS_REFRESH_TTL);
-    if stale {
+    if COMPONENTS_STAMP.claim_stale(COMPONENTS_REFRESH_TTL) {
         components.refresh(false);
-        *last = Some(now);
     }
 }
 
-struct GpuCache {
-    at: Option<Instant>,
-    v: Option<Vec<crate::gpu::GpuMetrics>>,
-}
-static GPUC: OnceCell<Mutex<GpuCache>> = OnceCell::new();
-
 // Static caches for unchanging data
 static HOSTNAME: OnceCell<String> = OnceCell::new();
 struct NetworkNameCache {
@@ -166,54 +163,6 @@ struct NetworkNameCache {
 static NETWORK_CACHE: OnceCell<Mutex<NetworkNameCache>> = OnceCell::new();
 static CPU_VEC: OnceCell<Mutex<Vec<f32>>> = OnceCell::new();
 
-fn cached_temp() -> Option<f32> {
-    if !temp_enabled() {
-        return None;
-    }
-    let now = Instant::now();
-    let lock = TEMP.get_or_init(|| Mutex::new(TempCache { at: None, v: None }));
-    let mut c = lock.lock().ok()?;
-    if c.at.is_none_or(|t| now.duration_since(t) >= TTL) {
-        c.at = Some(now);
-        // caller will fill this; we just hold a slot
-        c.v = None;
-    }
-    c.v
-}
-
-fn set_temp(v: Option<f32>) {
-    if let Some(lock) = TEMP.get()
-        && let Ok(mut c) = lock.lock()
-    {
-        c.v = v;
-        c.at = Some(Instant::now());
-    }
-}
-
-fn cached_gpus() -> Option<Vec<crate::gpu::GpuMetrics>> {
-    if !gpu_enabled() {
-        return None;
-    }
-    let now = Instant::now();
-    let lock = GPUC.get_or_init(|| Mutex::new(GpuCache { at: None, v: None }));
-    let mut c = lock.lock().ok()?;
-    if c.at.is_none_or(|t| now.duration_since(t) >= TTL) {
-        // mark stale; caller will refresh
-        c.at = Some(now);
-        c.v = None;
-    }
-    c.v.clone()
-}
-
-fn set_gpus(v: Option<Vec<crate::gpu::GpuMetrics>>) {
-    if let Some(lock) = GPUC.get()
-        && let Ok(mut c) = lock.lock()
-    {
-        c.v = v.clone();
-        c.at = Some(Instant::now());
-    }
-}
-
 // Collect only fast-changing metrics (CPU/mem/net + optional temps/gpus).
 pub async fn collect_fast_metrics(state: &AppState) -> Metrics {
     let ttl = StdDuration::from_millis(metrics_ttl_ms());
@@ -253,10 +202,13 @@ pub async fn collect_fast_metrics(state: &AppState) -> Metrics {
     let swap_used = sys.used_swap();
     drop(sys);
 
-    // CPU temperature: only refresh sensors if cache is stale
-    let cpu_temp_c = if cached_temp().is_some() {
-        cached_temp()
-    } else if temp_enabled() {
+    // CPU temperature: only rescan sensors when the cached result (even a
+    // cached "no sensor found") goes stale.
+    let cpu_temp_c = if !temp_enabled() {
+        None
+    } else if let Some(cached) = TEMP.get_fresh(TTL) {
+        cached
+    } else {
         let val = {
             let mut components = state.components.lock().await;
             refresh_components_if_stale(&mut components);
@@ -273,10 +225,8 @@ pub async fn collect_fast_metrics(state: &AppState) -> Metrics {
                 }
             })
         };
-        set_temp(val);
+        TEMP.set(val);
         val
-    } else {
-        None
     };
 
     // Networks with reusable name cache
@@ -320,47 +270,37 @@ pub async fn collect_fast_metrics(state: &AppState) -> Metrics {
         cache.infos.clone()
     };
 
-    // GPUs: if we already determined none exist, short-circuit (no repeated probing)
-    let gpus = if gpu_enabled() {
-        if state.gpu_checked.load(std::sync::atomic::Ordering::Acquire)
-            && !state.gpu_present.load(std::sync::atomic::Ordering::Relaxed)
-        {
-            None
-        } else if cached_gpus().is_some() {
-            cached_gpus()
-        } else {
-            let v = match collect_all_gpus() {
-                Ok(v) if !v.is_empty() => Some(v),
-                Ok(_) => None,
-                Err(_e) => {
-                    #[cfg(feature = "logging")]
-                    warn!("gpu collection failed: {_e}");
-                    None
-                }
-            };
-            // First probe records presence; subsequent calls rely on cache flags.
-            if !state
-                .gpu_checked
-                .swap(true, std::sync::atomic::Ordering::AcqRel)
-            {
-                if v.is_some() {
-                    state
-                        .gpu_present
-                        .store(true, std::sync::atomic::Ordering::Release);
-                } else {
-                    state
-                        .gpu_present
-                        .store(false, std::sync::atomic::Ordering::Release);
-                }
-            }
-            set_gpus(v.clone());
-            v
-        }
-    } else {
+    // GPUs: negative-probe cache short-circuits GPU-less hosts; otherwise the
+    // TTL cache answers, and only a stale miss reaches the worker thread.
+    let gpus = if !gpu_enabled()
+        || (state.gpu_checked.load(std::sync::atomic::Ordering::Acquire)
+            && !state.gpu_present.load(std::sync::atomic::Ordering::Relaxed))
+    {
         None
+    } else if let Some(cached) = GPUS.get_fresh(TTL) {
+        cached
+    } else {
+        let v = collect_all_gpus().await;
+        // First probe records presence; subsequent calls rely on the flags.
+        if !state
+            .gpu_checked
+            .swap(true, std::sync::atomic::Ordering::AcqRel)
+        {
+            state
+                .gpu_present
+                .store(v.is_some(), std::sync::atomic::Ordering::Release);
+        }
+        GPUS.set(v.clone());
+        v
     };
 
+    let sampled_at_ms = SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|d| d.as_millis() as u64)
+        .unwrap_or(0);
+
     let metrics = Metrics {
+        sampled_at_ms,
         cpu_total,
         cpu_per_core,
         mem_total,
@@ -381,6 +321,48 @@ pub async fn collect_fast_metrics(state: &AppState) -> Metrics {
     metrics
 }
 
+/// Best-effort parent-disk name for a partition device name:
+/// "nvme0n1p1" -> "nvme0n1", "mmcblk0p2" -> "mmcblk0", "sda1" -> "sda".
+/// Works with or without a "/dev/" prefix.
+fn parent_disk_name(name: &str) -> &str {
+    if let Some(pos) = name.rfind('p') {
+        let suffix = &name[pos + 1..];
+        if !suffix.is_empty() && suffix.chars().all(|c| c.is_ascii_digit()) {
+            return &name[..pos];
+        }
+    }
+    name.trim_end_matches(|c: char| c.is_ascii_digit())
+}
+
+/// Whether a device name refers to a partition rather than a whole disk.
+///
+/// On Linux, whole-disk devices are directories under /sys/block and
+/// partitions are not, so "is a partition" = "not in /sys/block, but the
+/// derived parent is". This gets right the cases the old name heuristic got
+/// wrong: a whole-disk filesystem on nvme0n1 (ends in a digit but IS in
+/// /sys/block) and zram1 (a whole device). Non-Linux keeps the heuristic.
+fn is_partition_name(name: &str) -> bool {
+    let bare = name.strip_prefix("/dev/").unwrap_or(name);
+    #[cfg(target_os = "linux")]
+    {
+        let sys_block = std::path::Path::new("/sys/block");
+        if sys_block.is_dir() {
+            return !sys_block.join(bare).is_dir()
+                && sys_block.join(parent_disk_name(bare)).is_dir();
+        }
+    }
+    is_partition_heuristic(bare)
+}
+
+/// Name-based fallback for platforms without /sys/block: a p<digits> marker
+/// or a trailing non-zero digit.
+fn is_partition_heuristic(bare: &str) -> bool {
+    bare.contains("p1")
+        || bare.contains("p2")
+        || bare.contains("p3")
+        || bare.ends_with(|c: char| c.is_ascii_digit() && c != '0')
+}
+
 // Cached disks
 pub async fn collect_disks(state: &AppState) -> Vec<DiskInfo> {
     let ttl = StdDuration::from_millis(disks_ttl_ms());
@@ -445,19 +427,7 @@ pub async fn collect_disks(state: &AppState) -> Vec<DiskInfo> {
                 return None;
             }
 
-            // Determine if this is a partition
-            let is_partition = name.contains("p1")
-                || name.contains("p2")
-                || name.contains("p3")
-                || name.ends_with('1')
-                || name.ends_with('2')
-                || name.ends_with('3')
-                || name.ends_with('4')
-                || name.ends_with('5')
-                || name.ends_with('6')
-                || name.ends_with('7')
-                || name.ends_with('8')
-                || name.ends_with('9');
+            let is_partition = is_partition_name(&name);
 
             // Try to find temperature for this disk
             let temperature = disk_temps.iter().find_map(|(key, &temp)| {
@@ -491,25 +461,7 @@ pub async fn collect_disks(state: &AppState) -> Vec<DiskInfo> {
 
     for partition in &partitions {
         if partition.is_partition {
-            // Extract parent disk name
-            // nvme0n1p1 -> nvme0n1, sda1 -> sda, mmcblk0p1 -> mmcblk0
-            let parent_name = if let Some(pos) = partition.name.rfind('p') {
-                // Check if character after 'p' is a digit
-                if partition
-                    .name
-                    .chars()
-                    .nth(pos + 1)
-                    .is_some_and(|c| c.is_ascii_digit())
-                {
-                    &partition.name[..pos]
-                } else {
-                    // Handle sda1, sdb2, etc (just trim trailing digit)
-                    partition.name.trim_end_matches(char::is_numeric)
-                }
-            } else {
-                // Handle sda1, sdb2, etc (just trim trailing digit)
-                partition.name.trim_end_matches(char::is_numeric)
-            };
+            let parent_name = parent_disk_name(&partition.name);
 
             // Look up temperature for the PARENT disk, not the partition
             // Strip /dev/ prefix if present for matching
@@ -553,21 +505,7 @@ pub async fn collect_disks(state: &AppState) -> Vec<DiskInfo> {
     // Add partitions after their parent disk
     for partition in partitions {
         if partition.is_partition {
-            // Find parent disk index
-            let parent_name = if let Some(pos) = partition.name.rfind('p') {
-                if partition
-                    .name
-                    .chars()
-                    .nth(pos + 1)
-                    .is_some_and(|c| c.is_ascii_digit())
-                {
-                    &partition.name[..pos]
-                } else {
-                    partition.name.trim_end_matches(char::is_numeric)
-                }
-            } else {
-                partition.name.trim_end_matches(char::is_numeric)
-            };
+            let parent_name = parent_disk_name(&partition.name);
 
             // Find where to insert this partition (after its parent)
             if let Some(parent_idx) = disks.iter().position(|d| d.name == parent_name) {
@@ -619,15 +557,8 @@ fn read_total_jiffies() -> io::Result<u64> {
 #[cfg(target_os = "linux")]
 #[inline]
 fn read_proc_jiffies(pid: u32) -> Option<u64> {
-    let path = format!("/proc/{pid}/stat");
-    let s = fs::read_to_string(path).ok()?;
-    // Find the right parenthesis that terminates comm; everything after is space-separated fields starting at "state"
-    let rpar = s.rfind(')')?;
-    let after = s.get(rpar + 2..)?; // skip ") "
-    let mut it = after.split_whitespace();
-    // utime (14th field) is offset 11 from "state", stime (15th) is next
-    let utime = it.nth(11)?.parse::<u64>().ok()?;
-    let stime = it.next()?.parse::<u64>().ok()?;
+    let s = fs::read_to_string(format!("/proc/{pid}/stat")).ok()?;
+    let (utime, stime) = procstat::utime_stime(&s)?;
     Some(utime.saturating_add(stime))
 }
 
@@ -657,9 +588,17 @@ pub async fn collect_processes_all(state: &AppState) -> ProcessesPayload {
     // filter when downgrading to a minimal refresh spec.
     let mut sys_guard = state.sys.lock().await;
     let sys = &mut *sys_guard;
+    // `true` = remove processes that no longer exist. With `false`, this
+    // long-lived System kept every process it had ever seen: the list grew
+    // without bound (21,648 entries on a machine with 289 processes after a
+    // few hours of build churn), process_count was meaningless, and — the
+    // reason this was found — a process you killed kept its row forever,
+    // because the agent went on reporting it. Safe here only because this is
+    // `ProcessesToUpdate::All`; with `Some(pids)` it would treat every process
+    // outside that list as dead and drop it.
     sys.refresh_processes_specifics(
         ProcessesToUpdate::All,
-        false,
+        true,
         ProcessRefreshKind::nothing().with_memory().without_tasks(),
     );
 
@@ -788,8 +727,11 @@ pub async fn collect_processes_all(state: &AppState) -> ProcessesPayload {
 
         //JW too complicated. simplify to remove strange behavior
 
-        // For active systems, get accurate CPU metrics
-        sys.refresh_processes_specifics(ProcessesToUpdate::All, false, kind.with_cpu());
+        // For active systems, get accurate CPU metrics.
+        // `true` = drop processes that have exited; see the Linux path above for
+        // what `false` cost us (an ever-growing list that kept reporting dead
+        // processes). Correct only because this is `ProcessesToUpdate::All`.
+        sys.refresh_processes_specifics(ProcessesToUpdate::All, true, kind.with_cpu());
 
         // } else {
         //     // For idle systems, just get basic process info
@@ -818,9 +760,12 @@ pub async fn collect_processes_all(state: &AppState) -> ProcessesPayload {
             };
 
             // Convert to percentage of total CPU capacity
-            // e.g., 100% on 2 cores of 8 core system = 25% total CPU
-            let raw = p.cpu_usage(); // This is per-core percentage
-            let total_cpu = raw.clamp(0.0, 100.0) / cpu_count;
+            // e.g., 100% on 2 cores of 8 core system = 25% total CPU.
+            // sysinfo reports per-core percentage which EXCEEDS 100 for
+            // multi-threaded processes, so clamp AFTER dividing — clamping
+            // first truncated e.g. 400%-on-8-cores to 12.5% instead of 50%.
+            let raw = p.cpu_usage();
+            let total_cpu = (raw / cpu_count.max(1.0)).clamp(0.0, 100.0);
 
             proc_cache.reusable_vec.push(ProcessInfo {
                 pid,
@@ -984,15 +929,8 @@ fn proc_state_label(c: char) -> &'static str {
 #[cfg(target_os = "linux")]
 fn read_parent_pid_from_proc(pid: u32) -> Option<u32> {
     let stat = fs::read_to_string(format!("/proc/{pid}/stat")).ok()?;
-    // Format: pid (comm) state ppid ... — comm can contain spaces/parens,
-    // so we step past the closing paren first.
-    let ppid_start = stat.rfind(')')?;
-    // After ") ": state, ppid, ... — ppid is the second field.
-    stat[ppid_start + 1..]
-        .split_whitespace()
-        .nth(1)?
-        .parse::<u32>()
-        .ok()
+    // Post-comm field 1 is ppid.
+    procstat::field(&stat, 1)?.parse::<u32>().ok()
 }
 
 /// Collect process information from /proc files
@@ -1035,16 +973,9 @@ fn collect_process_info_from_proc(
     let thread_count = st.threads;
     let status = proc_state_label(st.state_ch).to_string();
 
-    // Read start time from stat — comm-safe via rfind(')').
+    // starttime is post-comm field 19.
     let start_time = if let Ok(stat) = fs::read_to_string(format!("/proc/{pid}/stat")) {
-        let stat_end = stat.rfind(')')?;
-        // After ") ": state, ppid, ..., starttime — starttime is the 20th
-        // post-comm field (index 19).
-        stat[stat_end + 1..]
-            .split_whitespace()
-            .nth(19)?
-            .parse::<u64>()
-            .ok()?
+        procstat::field(&stat, 19)?.parse::<u64>().ok()?
     } else {
         0
     };
@@ -1079,7 +1010,7 @@ fn collect_process_info_from_proc(
         .map(|p| p.to_string_lossy().to_string());
 
     // One read of /proc/{pid}/stat covers both user + system CPU times.
-    let (cpu_time_user, cpu_time_system) = get_cpu_times_ms(pid);
+    let (cpu_time_user, cpu_time_system) = get_cpu_times_us(pid);
 
     Some(DetailedProcessInfo {
         pid,
@@ -1192,18 +1123,7 @@ fn collect_thread_info(pid: u32) -> Vec<crate::types::ThreadInfo> {
             continue;
         };
 
-        // Thread/comm names can contain spaces or parens, so step past the
-        // last ')' before parsing post-comm fields. Post-comm offsets:
-        //   0: state, 1: ppid, 2: pgrp, ..., 11: utime, 12: stime
-        let Some(rpar) = stat_content.rfind(')') else {
-            continue;
-        };
-        let Some(after) = stat_content.get(rpar + 1..) else {
-            continue;
-        };
-        let mut it = after.split_whitespace();
-        let status = it
-            .next()
+        let status = procstat::field(&stat_content, 0)
             .and_then(|s| s.chars().next())
             .map(|c| match c {
                 'R' => "Running",
@@ -1218,14 +1138,9 @@ fn collect_thread_info(pid: u32) -> Vec<crate::types::ThreadInfo> {
             .unwrap_or("Unknown")
             .to_string();
 
-        // 10 fields between state and utime (ppid..cmajflt).
-        let utime = it.nth(10).and_then(|s| s.parse::<u64>().ok()).unwrap_or(0);
-        let stime = it.next().and_then(|s| s.parse::<u64>().ok()).unwrap_or(0);
-
-        // Convert clock ticks to microseconds (assuming 100 Hz)
-        // 1 tick = 10ms = 10,000 microseconds
-        let cpu_time_user = utime * 10_000;
-        let cpu_time_system = stime * 10_000;
+        let (utime, stime) = procstat::utime_stime(&stat_content).unwrap_or((0, 0));
+        let cpu_time_user = utime * procstat::TICK_US;
+        let cpu_time_system = stime * procstat::TICK_US;
 
         threads.push(crate::types::ThreadInfo {
             tid,
@@ -1257,10 +1172,17 @@ pub async fn collect_process_metrics(
     system.refresh_processes_specifics(
         ProcessesToUpdate::Some(&[sysinfo::Pid::from_u32(pid)]),
         false,
+        // cmd/exe/cwd feed the modal's Command & Details pane. They're
+        // immutable per process, so OnlyIfNotSet reads them once per PID and
+        // serves the cache afterwards — the "minimal refresh" optimization
+        // had dropped them entirely, leaving the pane blank.
         ProcessRefreshKind::nothing()
             .with_memory()
             .with_cpu()
-            .with_disk_usage(),
+            .with_disk_usage()
+            .with_cmd(sysinfo::UpdateKind::OnlyIfNotSet)
+            .with_exe(sysinfo::UpdateKind::OnlyIfNotSet)
+            .with_cwd(sysinfo::UpdateKind::OnlyIfNotSet),
     );
 
     let process = system
@@ -1350,7 +1272,7 @@ pub async fn collect_process_metrics(
     let threads = collect_thread_info(pid);
 
     // One read of /proc/{pid}/stat covers both user + system CPU times.
-    let (cpu_time_user, cpu_time_system) = get_cpu_times_ms(pid);
+    let (cpu_time_user, cpu_time_system) = get_cpu_times_us(pid);
 
     // Now construct the detailed info without holding the lock
     let detailed_info = DetailedProcessInfo {
@@ -1384,9 +1306,26 @@ pub async fn collect_process_metrics(
     })
 }
 
-/// Collect journal entries for a specific process
-pub fn collect_journal_entries(pid: u32) -> Result<JournalResponse, String> {
-    let output = Command::new("journalctl")
+/// Epoch microseconds -> RFC 3339 UTC for display. The old code
+/// Debug-formatted a SystemTime and string-replaced it into a raw epoch
+/// string that was neither ISO 8601 nor what the field documented.
+fn format_journal_timestamp(timestamp_us: u64) -> String {
+    time::OffsetDateTime::from_unix_timestamp_nanos(timestamp_us as i128 * 1000)
+        .ok()
+        .and_then(|t| {
+            t.format(&time::format_description::well_known::Rfc3339)
+                .ok()
+        })
+        .unwrap_or_else(|| timestamp_us.to_string())
+}
+
+/// Collect journal entries for a specific process.
+///
+/// Async via tokio::process — journalctl can take hundreds of ms on slow
+/// storage, and the old std::process call blocked one of the runtime's two
+/// worker threads for the duration.
+pub async fn collect_journal_entries(pid: u32) -> Result<JournalResponse, String> {
+    let output = tokio::process::Command::new("journalctl")
         .args([
             &format!("_PID={pid}"),
             "--output=json",
@@ -1394,6 +1333,7 @@ pub fn collect_journal_entries(pid: u32) -> Result<JournalResponse, String> {
             "--no-pager",
         ])
         .output()
+        .await
         .map_err(|e| format!("Failed to execute journalctl: {e}"))?;
 
     if !output.status.success() {
@@ -1415,27 +1355,14 @@ pub fn collect_journal_entries(pid: u32) -> Result<JournalResponse, String> {
         let json: serde_json::Value =
             serde_json::from_str(line).map_err(|e| format!("Failed to parse journal JSON: {e}"))?;
 
-        // Extract relevant fields
-        let timestamp_str = json
+        // __REALTIME_TIMESTAMP is epoch microseconds as a string.
+        let timestamp_us = json
             .get("__REALTIME_TIMESTAMP")
             .and_then(|v| v.as_str())
-            .unwrap_or("0");
+            .and_then(|v| v.parse::<u64>().ok())
+            .unwrap_or(0);
 
-        // Convert timestamp to ISO 8601 format
-        let timestamp = if let Ok(ts_micros) = timestamp_str.parse::<u64>() {
-            let ts_secs = ts_micros / 1_000_000;
-            let ts_nanos = (ts_micros % 1_000_000) * 1000;
-            let time = SystemTime::UNIX_EPOCH
-                + Duration::from_secs(ts_secs)
-                + Duration::from_nanos(ts_nanos);
-            // Simple ISO 8601 format - we can improve this if needed
-            format!("{time:?}")
-                .replace("SystemTime { tv_sec: ", "")
-                .replace(", tv_nsec: ", ".")
-                .replace(" }", "")
-        } else {
-            timestamp_str.to_string()
-        };
+        let timestamp = format_journal_timestamp(timestamp_us);
 
         let priority = match json.get("PRIORITY").and_then(|v| v.as_str()) {
             Some("0") => LogLevel::Emergency,
@@ -1482,6 +1409,7 @@ pub fn collect_journal_entries(pid: u32) -> Result<JournalResponse, String> {
 
         entries.push(JournalEntry {
             timestamp,
+            timestamp_us,
             priority,
             message,
             unit,
@@ -1493,7 +1421,26 @@ pub fn collect_journal_entries(pid: u32) -> Result<JournalResponse, String> {
     }
 
     // Sort by timestamp (newest first)
-    entries.sort_by(|a, b| b.timestamp.cmp(&a.timestamp));
+    entries.sort_by_key(|e| std::cmp::Reverse(e.timestamp_us));
+
+    // journalctl exits 0 with no output when the invoking user simply cannot
+    // SEE the process's entries (e.g. a user-run agent asking about a system
+    // service) — but it explains itself on stderr ("You are currently not
+    // seeing messages from other users and the system…"). Pass that hint
+    // along so the client can distinguish "no logs" from "no access".
+    let notice = if entries.is_empty() {
+        let err = String::from_utf8_lossy(&output.stderr);
+        let hint: String = err
+            .lines()
+            .map(str::trim)
+            .filter(|l| !l.is_empty())
+            .take(2)
+            .collect::<Vec<_>>()
+            .join(" ");
+        if hint.is_empty() { None } else { Some(hint) }
+    } else {
+        None
+    };
 
     let response_timestamp = SystemTime::now()
         .duration_since(UNIX_EPOCH)
@@ -1507,6 +1454,70 @@ pub fn collect_journal_entries(pid: u32) -> Result<JournalResponse, String> {
         entries,
         total_count,
         truncated,
+        notice,
         cached_at: response_timestamp,
     })
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// comm can contain spaces and parens; parsing must key off the LAST ')'.
+    #[cfg(target_os = "linux")]
+    #[test]
+    fn procstat_handles_hostile_comm_names() {
+        let stat = "1234 (weird name) (2)) R 1 2 3 4 5 6 7 8 9 10 700 800 0 0 20";
+        assert_eq!(procstat::field(stat, 0), Some("R"));
+        assert_eq!(procstat::field(stat, 1), Some("1"));
+        assert_eq!(procstat::utime_stime(stat), Some((700, 800)));
+    }
+
+    /// USER_HZ ticks convert to MICROSECONDS — the wire contract. This used
+    /// to be *10 (ms), rendering process CPU times 1000x too small next to
+    /// thread times.
+    #[cfg(target_os = "linux")]
+    #[test]
+    fn cpu_times_are_microseconds() {
+        assert_eq!(procstat::TICK_US, 10_000);
+    }
+
+    #[test]
+    fn parent_disk_name_strips_partition_suffixes() {
+        assert_eq!(parent_disk_name("nvme0n1p1"), "nvme0n1");
+        assert_eq!(parent_disk_name("nvme1n1p12"), "nvme1n1");
+        assert_eq!(parent_disk_name("mmcblk0p2"), "mmcblk0");
+        assert_eq!(parent_disk_name("sda1"), "sda");
+        assert_eq!(parent_disk_name("/dev/nvme0n1p1"), "/dev/nvme0n1");
+        // 'p' inside a word is not a partition marker.
+        assert_eq!(parent_disk_name("mapper/vg-lv"), "mapper/vg-lv");
+    }
+
+    /// The old heuristic flagged whole-disk names ending in a digit
+    /// (nvme0n1, zram1) as partitions. On Linux /sys/block decides; this
+    /// pins the real-machine behavior for devices every Linux box has.
+    #[cfg(target_os = "linux")]
+    #[test]
+    fn sys_block_devices_are_not_partitions() {
+        let sys_block = std::path::Path::new("/sys/block");
+        if !sys_block.is_dir() {
+            return; // exotic environment; nothing to assert
+        }
+        for entry in std::fs::read_dir(sys_block).unwrap().flatten() {
+            let name = entry.file_name().to_string_lossy().into_owned();
+            assert!(
+                !is_partition_name(&name),
+                "{name} is a whole disk but was flagged as a partition"
+            );
+        }
+    }
+
+    #[test]
+    fn journal_timestamps_are_rfc3339() {
+        let s = format_journal_timestamp(1_786_752_000_000_000);
+        assert_eq!(s, "2026-08-15T00:00:00Z");
+        // Sub-second precision survives.
+        let s = format_journal_timestamp(1_786_752_000_123_456);
+        assert!(s.starts_with("2026-08-15T00:00:00.123456"), "{s}");
+    }
+}

socktop_agent/src/state.rs

@@ -74,6 +74,55 @@ pub struct AppState {
     pub cache_journal_entries: Arc<Mutex<HashMap<u32, CacheEntry<crate::types::JournalResponse>>>>,
 }
 
+/// TTL-gated value behind a std Mutex, for `static` caches on hot paths.
+/// Replaces the hand-rolled TempCache/GpuCache/refresh-timestamp statics
+/// that each reimplemented the same at/value pair.
+pub struct TtlCell<T> {
+    inner: std::sync::Mutex<CacheEntry<T>>,
+}
+
+impl<T: Clone> Default for TtlCell<T> {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl<T: Clone> TtlCell<T> {
+    pub const fn new() -> Self {
+        Self {
+            inner: std::sync::Mutex::new(CacheEntry::new()),
+        }
+    }
+    /// The stored value, only while fresh. Poisoned lock reads as a miss.
+    pub fn get_fresh(&self, ttl: Duration) -> Option<T> {
+        let g = self.inner.lock().ok()?;
+        if g.is_fresh(ttl) {
+            g.value.clone()
+        } else {
+            None
+        }
+    }
+    pub fn set(&self, v: T) {
+        if let Ok(mut g) = self.inner.lock() {
+            g.set(v);
+        }
+    }
+    /// True exactly once per TTL window: restamps and tells the caller to do
+    /// the refresh. Atomic check-and-stamp so concurrent callers don't both
+    /// refresh.
+    pub fn claim_stale(&self, ttl: Duration) -> bool {
+        let Ok(mut g) = self.inner.lock() else {
+            return false;
+        };
+        if g.at.is_none_or(|t| t.elapsed() >= ttl) {
+            g.at = Some(Instant::now());
+            true
+        } else {
+            false
+        }
+    }
+}
+
 #[derive(Clone, Debug)]
 pub struct CacheEntry<T> {
     pub at: Option<Instant>,
@@ -87,7 +136,7 @@ impl<T> Default for CacheEntry<T> {
 }
 
 impl<T> CacheEntry<T> {
-    pub fn new() -> Self {
+    pub const fn new() -> Self {
         Self {
             at: None,
             value: None,

socktop_agent/src/tls.rs

@@ -24,6 +24,17 @@ pub fn cert_paths() -> (PathBuf, PathBuf) {
 pub fn ensure_self_signed_cert() -> anyhow::Result<(PathBuf, PathBuf)> {
     let (cert_path, key_path) = cert_paths();
     if cert_path.exists() && key_path.exists() {
+        // Keys generated by agents older than 1.60 were written with the
+        // default umask (typically 0644): tighten them on startup.
+        #[cfg(unix)]
+        {
+            use std::os::unix::fs::PermissionsExt;
+            if let Ok(meta) = fs::metadata(&key_path)
+                && meta.permissions().mode() & 0o077 != 0
+            {
+                let _ = fs::set_permissions(&key_path, fs::Permissions::from_mode(0o600));
+            }
+        }
         return Ok((cert_path, key_path));
     }
     fs::create_dir_all(cert_path.parent().unwrap())?;
@@ -79,7 +90,16 @@ pub fn ensure_self_signed_cert() -> anyhow::Result<(PathBuf, PathBuf)> {
 
     let mut f = fs::File::create(&cert_path)?;
     f.write_all(cert_pem.as_bytes())?;
-    let mut k = fs::File::create(&key_path)?;
+    // The private key must not be world-readable (File::create honors the
+    // umask, which typically yields 0644).
+    let mut key_opts = fs::OpenOptions::new();
+    key_opts.write(true).create(true).truncate(true);
+    #[cfg(unix)]
+    {
+        use std::os::unix::fs::OpenOptionsExt;
+        key_opts.mode(0o600);
+    }
+    let mut k = key_opts.open(&key_path)?;
     k.write_all(key_pem.as_bytes())?;
 
     println!(

socktop_agent/src/types.rs

@@ -30,6 +30,11 @@ pub struct ProcessInfo {
 
 #[derive(Debug, Clone, Serialize)]
 pub struct Metrics {
+    /// Epoch ms when this snapshot was actually collected. The agent serves
+    /// TTL-cached snapshots, so the client needs the AGENT's sample time to
+    /// compute rates — measuring against client receive time turned cache
+    /// hits into a 0-then-2x sawtooth in the network graphs.
+    pub sampled_at_ms: u64,
     pub cpu_total: f32,
     pub cpu_per_core: Vec<f32>,
     pub mem_total: u64,
@@ -93,7 +98,8 @@ pub struct ProcessMetricsResponse {
 
 #[derive(Debug, Clone, Serialize)]
 pub struct JournalEntry {
-    pub timestamp: String, // ISO 8601 formatted timestamp
+    pub timestamp: String, // RFC 3339 UTC, for display
+    pub timestamp_us: u64, // epoch microseconds, for sorting/formatting
     pub priority: LogLevel,
     pub message: String,
     pub unit: Option<String>, // systemd unit name
@@ -120,5 +126,9 @@ pub struct JournalResponse {
     pub entries: Vec<JournalEntry>,
     pub total_count: u32,
     pub truncated: bool,
+    /// journalctl's own explanation when the result is empty because of
+    /// journal ACCESS (not absence of logs) — e.g. a user-run agent asking
+    /// about a system service. None when entries exist or nothing to say.
+    pub notice: Option<String>,
     pub cached_at: u64, // Unix timestamp when this data was cached
 }

socktop_agent/src/ws.rs

@@ -16,9 +16,7 @@ use crate::metrics::{collect_disks, collect_fast_metrics, collect_processes_all}
 use crate::proto::pb;
 use crate::state::AppState;
 
-// Compression threshold based on typical payload size
-// Temporarily increased for testing - revert to 768 for production
-//const COMPRESSION_THRESHOLD: usize = 50_000;
+// Payloads at or below this many bytes are sent as-is; larger ones are gzipped.
 const COMPRESSION_THRESHOLD: usize = 768;
 
 // Reusable buffer for compression to avoid allocations
@@ -52,6 +50,66 @@ pub async fn ws_handler(
     ws.on_upgrade(move |socket| handle_socket(socket, state))
 }
 
+/// Per-PID cache limits: entries older than MAX_AGE are swept on every
+/// insert and the map is capped at MAX_ENTRIES (oldest evicted first), so a
+/// client walking PIDs cannot grow agent memory without bound.
+const PER_PID_CACHE_MAX_AGE: std::time::Duration = std::time::Duration::from_secs(60);
+const PER_PID_CACHE_MAX_ENTRIES: usize = 64;
+
+/// Serve a per-PID request from a TTL cache, collecting on miss. One home
+/// for the logic that get_process_metrics and get_journal_entries used to
+/// duplicate ~50 lines apiece.
+async fn respond_per_pid_cached<T, Fut>(
+    socket: &mut WebSocket,
+    cache: &Mutex<HashMap<u32, crate::state::CacheEntry<T>>>,
+    pid: u32,
+    ttl: std::time::Duration,
+    request_name: &str,
+    collect: impl FnOnce() -> Fut,
+) where
+    T: serde::Serialize + Clone,
+    Fut: std::future::Future<Output = Result<T, String>>,
+{
+    {
+        let cache = cache.lock().await;
+        if let Some(entry) = cache.get(&pid)
+            && entry.is_fresh(ttl)
+            && let Some(v) = entry.get()
+        {
+            let _ = send_json(socket, v).await;
+            return;
+        }
+    }
+    match collect().await {
+        Ok(resp) => {
+            {
+                let mut cache = cache.lock().await;
+                cache.retain(|_, e| e.at.is_some_and(|t| t.elapsed() < PER_PID_CACHE_MAX_AGE));
+                while cache.len() >= PER_PID_CACHE_MAX_ENTRIES {
+                    let oldest = cache.iter().min_by_key(|(_, e)| e.at).map(|(k, _)| *k);
+                    match oldest {
+                        Some(k) => cache.remove(&k),
+                        None => break,
+                    };
+                }
+                cache
+                    .entry(pid)
+                    .or_insert_with(crate::state::CacheEntry::new)
+                    .set(resp.clone());
+            }
+            let _ = send_json(socket, &resp).await;
+        }
+        Err(err) => {
+            let error_response = serde_json::json!({
+                "error": err,
+                "request": request_name,
+                "pid": pid
+            });
+            let _ = send_json(socket, &error_response).await;
+        }
+    }
+}
+
 async fn handle_socket(mut socket: WebSocket, state: AppState) {
     state
         .client_count
@@ -126,84 +184,30 @@ async fn handle_socket(mut socket: WebSocket, state: AppState) {
                 if let Some(pid_str) = text.strip_prefix("get_process_metrics:")
                     && let Ok(pid) = pid_str.parse::<u32>()
                 {
-                    let ttl = std::time::Duration::from_millis(250); // 250ms TTL
-
-                    // Check cache first
-                    {
-                        let cache = state.cache_process_metrics.lock().await;
-                        if let Some(entry) = cache.get(&pid)
-                            && entry.is_fresh(ttl)
-                            && let Some(cached_response) = entry.get()
-                        {
-                            let _ = send_json(&mut socket, cached_response).await;
-                            continue;
-                        }
-                    }
-
-                    // Collect fresh data
-                    match crate::metrics::collect_process_metrics(pid, &state).await {
-                        Ok(response) => {
-                            // Cache the response
-                            {
-                                let mut cache = state.cache_process_metrics.lock().await;
-                                cache
-                                    .entry(pid)
-                                    .or_insert_with(crate::state::CacheEntry::new)
-                                    .set(response.clone());
-                            }
-                            let _ = send_json(&mut socket, &response).await;
-                        }
-                        Err(err) => {
-                            let error_response = serde_json::json!({
-                                "error": err,
-                                "request": "get_process_metrics",
-                                "pid": pid
-                            });
-                            let _ = send_json(&mut socket, &error_response).await;
-                        }
-                    }
+                    respond_per_pid_cached(
+                        &mut socket,
+                        &state.cache_process_metrics,
+                        pid,
+                        std::time::Duration::from_millis(250),
+                        "get_process_metrics",
+                        || crate::metrics::collect_process_metrics(pid, &state),
+                    )
+                    .await;
                 }
             }
             Message::Text(ref text) if text.starts_with("get_journal_entries:") => {
                 if let Some(pid_str) = text.strip_prefix("get_journal_entries:")
                     && let Ok(pid) = pid_str.parse::<u32>()
                 {
-                    let ttl = std::time::Duration::from_secs(1); // 1s TTL
-
-                    // Check cache first
-                    {
-                        let cache = state.cache_journal_entries.lock().await;
-                        if let Some(entry) = cache.get(&pid)
-                            && entry.is_fresh(ttl)
-                            && let Some(cached_response) = entry.get()
-                        {
-                            let _ = send_json(&mut socket, cached_response).await;
-                            continue;
-                        }
-                    }
-
-                    // Collect fresh data
-                    match crate::metrics::collect_journal_entries(pid) {
-                        Ok(response) => {
-                            // Cache the response
-                            {
-                                let mut cache = state.cache_journal_entries.lock().await;
-                                cache
-                                    .entry(pid)
-                                    .or_insert_with(crate::state::CacheEntry::new)
-                                    .set(response.clone());
-                            }
-                            let _ = send_json(&mut socket, &response).await;
-                        }
-                        Err(err) => {
-                            let error_response = serde_json::json!({
-                                "error": err,
-                                "request": "get_journal_entries",
-                                "pid": pid
-                            });
-                            let _ = send_json(&mut socket, &error_response).await;
-                        }
-                    }
+                    respond_per_pid_cached(
+                        &mut socket,
+                        &state.cache_journal_entries,
+                        pid,
+                        std::time::Duration::from_secs(1),
+                        "get_journal_entries",
+                        || crate::metrics::collect_journal_entries(pid),
+                    )
+                    .await;
                 }
             }
             Message::Close(_) => break,

socktop_agent/tests/cache_tests.rs

@@ -42,6 +42,7 @@ async fn test_process_cache_ttl() {
     };
 
     let journal_response = JournalResponse {
+        notice: None,
         entries: vec![],
         total_count: 0,
         truncated: false,

socktop_agent/tests/process_details.rs

@@ -33,7 +33,7 @@ async fn test_collect_journal_entries_self() {
     // Test collecting journal entries for our own process
     let pid = process::id();
 
-    match collect_journal_entries(pid) {
+    match collect_journal_entries(pid).await {
         Ok(response) => {
             assert!(response.cached_at > 0);
             println!(
@@ -74,7 +74,7 @@ async fn test_collect_journal_entries_invalid_pid() {
     // Test with an invalid PID - journalctl might still return empty results
     let invalid_pid = 999999;
 
-    match collect_journal_entries(invalid_pid) {
+    match collect_journal_entries(invalid_pid).await {
         Ok(response) => {
             println!(
                 "✓ Journal query completed for invalid PID {} (empty result expected): {} entries",
@@ -87,3 +87,19 @@ async fn test_collect_journal_entries_invalid_pid() {
         }
     }
 }
+
+/// The Command & Details pane went blank when the minimal-refresh
+/// optimization dropped cmd from the detail endpoint's refresh kind.
+#[tokio::test]
+async fn test_process_metrics_include_command() {
+    let state = AppState::new();
+    let pid = std::process::id();
+    let resp = collect_process_metrics(pid, &state)
+        .await
+        .expect("collect self");
+    assert!(
+        !resp.process.command.is_empty(),
+        "command should not be empty for self (cmdline is always readable)"
+    );
+    println!("command = {}", resp.process.command);
+}

socktop_connector/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "socktop_connector"
-version = "1.50.0"
+version = "1.60.1"
 edition = "2024"
 license = "MIT"
 description = "WebSocket connector library for socktop agent communication"

socktop_connector/build.rs

@@ -1,8 +1,12 @@
 fn main() -> Result<(), Box<dyn std::error::Error>> {
-    // Set the protoc binary path to use the vendored version for CI compatibility
-    // SAFETY: We're only setting PROTOC in a build script environment, which is safe
-    unsafe {
-        std::env::set_var("PROTOC", protoc_bin_vendored::protoc_bin_path()?);
+    // Vendored protoc for reproducible builds where available. It ships no
+    // riscv64 binary, so on such hosts leave $PROTOC / PATH lookup to
+    // prost-build (apt: protobuf-compiler).
+    // SAFETY: We're only setting PROTOC in a build script environment.
+    if let Ok(protoc) = protoc_bin_vendored::protoc_bin_path() {
+        unsafe {
+            std::env::set_var("PROTOC", protoc);
+        }
     }
 
     prost_build::compile_protos(&["processes.proto"], &["."])?;

socktop_connector/src/networking/connection.rs

@@ -6,7 +6,7 @@ use crate::error::{ConnectorError, Result};
 use std::io::BufReader;
 use std::sync::Arc;
 use tokio_tungstenite::tungstenite::client::IntoClientRequest;
-use tokio_tungstenite::{MaybeTlsStream, WebSocketStream, connect_async};
+use tokio_tungstenite::{MaybeTlsStream, WebSocketStream};
 use url::Url;
 
 #[cfg(feature = "tls")]
@@ -15,7 +15,7 @@ use {
     rustls::{
         DigitallySignedStruct, RootCertStore, SignatureScheme,
         client::danger::{HandshakeSignatureValid, ServerCertVerified, ServerCertVerifier},
-        crypto::ring,
+        crypto::{WebPkiSupportedAlgorithms, ring},
         pki_types::{CertificateDer, ServerName, UnixTime},
     },
     rustls_pemfile::Item,
@@ -64,7 +64,8 @@ async fn connect_without_ca_and_config(url: &str, config: &ConnectorConfig) -> R
         );
     }
 
-    let (ws, _) = connect_async(req).await?;
+    // `true` disables Nagle: small request/response frames, latency matters.
+    let (ws, _) = tokio_tungstenite::connect_async_with_config(req, None, true).await?;
     Ok(ws)
 }
 
@@ -85,7 +86,12 @@ async fn connect_with_ca_and_config(
             der_certs.push(der);
         }
     }
-    root.add_parsable_certificates(der_certs);
+    if der_certs.is_empty() {
+        return Err(ConnectorError::protocol_error(format!(
+            "no certificates found in --tls-ca file: {ca_path}"
+        )));
+    }
+    root.add_parsable_certificates(der_certs.iter().cloned());
 
     let mut cfg = ClientConfig::builder()
         .with_root_certificates(root)
@@ -114,57 +120,89 @@ async fn connect_with_ca_and_config(
     }
 
     if !config.verify_hostname {
-        #[derive(Debug)]
-        struct NoVerify;
-        impl ServerCertVerifier for NoVerify {
-            fn verify_server_cert(
-                &self,
-                _end_entity: &CertificateDer<'_>,
-                _intermediates: &[CertificateDer<'_>],
-                _server_name: &ServerName,
-                _ocsp_response: &[u8],
-                _now: UnixTime,
-            ) -> std::result::Result<ServerCertVerified, rustls::Error> {
-                Ok(ServerCertVerified::assertion())
-            }
-            fn verify_tls12_signature(
-                &self,
-                _message: &[u8],
-                _cert: &CertificateDer<'_>,
-                _dss: &DigitallySignedStruct,
-            ) -> std::result::Result<HandshakeSignatureValid, rustls::Error> {
-                Ok(HandshakeSignatureValid::assertion())
-            }
-            fn verify_tls13_signature(
-                &self,
-                _message: &[u8],
-                _cert: &CertificateDer<'_>,
-                _dss: &DigitallySignedStruct,
-            ) -> std::result::Result<HandshakeSignatureValid, rustls::Error> {
-                Ok(HandshakeSignatureValid::assertion())
-            }
-            fn supported_verify_schemes(&self) -> Vec<SignatureScheme> {
-                vec![
-                    SignatureScheme::ECDSA_NISTP256_SHA256,
-                    SignatureScheme::ED25519,
-                    SignatureScheme::RSA_PSS_SHA256,
-                ]
-            }
-        }
-        cfg.dangerous().set_certificate_verifier(Arc::new(NoVerify));
-        // Note: hostname verification disabled (default). Set SOCKTOP_VERIFY_NAME=1 to enable strict SAN checking.
+        // Default mode: certificate PINNING without hostname verification.
+        // The server must present a certificate byte-identical to one in the
+        // --tls-ca file. This intentionally ignores expiry and chain building
+        // (the operator pinned this exact cert), but unlike a blanket accept
+        // it makes MITM certs fail the handshake.
+        cfg.dangerous()
+            .set_certificate_verifier(Arc::new(PinnedCertVerifier::new(der_certs)));
     }
     let cfg = Arc::new(cfg);
+    // Third argument is tungstenite's `disable_nagle`: always true — socktop
+    // exchanges small request/response frames where Nagle only adds latency.
     let (ws, _) = tokio_tungstenite::connect_async_tls_with_config(
         req,
         None,
-        config.verify_hostname,
+        true,
         Some(Connector::Rustls(cfg)),
     )
     .await?;
     Ok(ws)
 }
 
+/// Accepts exactly the certificates the user pinned via `--tls-ca`, nothing else.
+///
+/// Used when hostname verification is off (the default for self-signed
+/// home-lab certs). Signature validation still runs with the ring provider's
+/// full algorithm set; only the certificate identity check is replaced —
+/// by an exact DER comparison against the pinned certificate(s).
+#[cfg(feature = "tls")]
+#[derive(Debug)]
+struct PinnedCertVerifier {
+    pinned: Vec<CertificateDer<'static>>,
+    algorithms: WebPkiSupportedAlgorithms,
+}
+
+#[cfg(feature = "tls")]
+impl PinnedCertVerifier {
+    fn new(pinned: Vec<CertificateDer<'static>>) -> Self {
+        Self {
+            pinned,
+            algorithms: ring::default_provider().signature_verification_algorithms,
+        }
+    }
+}
+
+#[cfg(feature = "tls")]
+impl ServerCertVerifier for PinnedCertVerifier {
+    fn verify_server_cert(
+        &self,
+        end_entity: &CertificateDer<'_>,
+        _intermediates: &[CertificateDer<'_>],
+        _server_name: &ServerName,
+        _ocsp_response: &[u8],
+        _now: UnixTime,
+    ) -> std::result::Result<ServerCertVerified, rustls::Error> {
+        if self.pinned.iter().any(|p| p == end_entity) {
+            Ok(ServerCertVerified::assertion())
+        } else {
+            Err(rustls::Error::InvalidCertificate(
+                rustls::CertificateError::ApplicationVerificationFailure,
+            ))
+        }
+    }
+    fn verify_tls12_signature(
+        &self,
+        message: &[u8],
+        cert: &CertificateDer<'_>,
+        dss: &DigitallySignedStruct,
+    ) -> std::result::Result<HandshakeSignatureValid, rustls::Error> {
+        rustls::crypto::verify_tls12_signature(message, cert, dss, &self.algorithms)
+    }
+    fn verify_tls13_signature(
+        &self,
+        message: &[u8],
+        cert: &CertificateDer<'_>,
+        dss: &DigitallySignedStruct,
+    ) -> std::result::Result<HandshakeSignatureValid, rustls::Error> {
+        rustls::crypto::verify_tls13_signature(message, cert, dss, &self.algorithms)
+    }
+    fn supported_verify_schemes(&self) -> Vec<SignatureScheme> {
+        self.algorithms.supported_schemes()
+    }
+}
+
 #[cfg(not(feature = "tls"))]
 async fn connect_with_ca_and_config(
     _url: &str,
@@ -181,3 +219,73 @@ async fn connect_with_ca_and_config(
 fn ensure_crypto_provider() {
     let _ = ring::default_provider().install_default();
 }
+
+#[cfg(all(test, feature = "tls"))]
+mod tests {
+    use super::*;
+
+    fn verifier(pinned: &[&[u8]]) -> PinnedCertVerifier {
+        let _ = ring::default_provider().install_default();
+        PinnedCertVerifier::new(
+            pinned
+                .iter()
+                .map(|b| CertificateDer::from(b.to_vec()))
+                .collect(),
+        )
+    }
+
+    fn verify(v: &PinnedCertVerifier, presented: &[u8]) -> bool {
+        v.verify_server_cert(
+            &CertificateDer::from(presented.to_vec()),
+            &[],
+            &ServerName::try_from("agent.test").unwrap(),
+            &[],
+            UnixTime::now(),
+        )
+        .is_ok()
+    }
+
+    /// The regression this verifier exists to prevent: the old NoVerify
+    /// accepted ANY certificate when hostname verification was off, so the
+    /// documented pinning was a no-op. The pinned cert must be accepted and
+    /// every other cert rejected.
+    #[test]
+    fn only_the_pinned_certificate_is_accepted() {
+        let v = verifier(&[b"pinned-cert-der"]);
+        assert!(verify(&v, b"pinned-cert-der"));
+        assert!(!verify(&v, b"some-mitm-cert"), "unpinned cert accepted");
+        assert!(!verify(&v, b""), "empty cert accepted");
+    }
+
+    /// A --tls-ca file may hold several certs (e.g. during rotation); any of
+    /// them must satisfy the pin.
+    #[test]
+    fn any_cert_in_a_multi_cert_pem_satisfies_the_pin() {
+        let v = verifier(&[b"old-cert", b"new-cert"]);
+        assert!(verify(&v, b"old-cert"));
+        assert!(verify(&v, b"new-cert"));
+        assert!(!verify(&v, b"third-party-cert"));
+    }
+
+    /// Fail closed: an empty pin set must reject everything rather than
+    /// falling back to accept-all.
+    #[test]
+    fn an_empty_pin_set_rejects_all_certificates() {
+        let v = verifier(&[]);
+        assert!(!verify(&v, b"anything"));
+    }
+
+    /// Signature schemes come from the real provider, not a hardcoded list —
+    /// an agent using e.g. RSA-PKCS1 must still be able to handshake.
+    #[test]
+    fn signature_schemes_come_from_the_provider() {
+        let v = verifier(&[b"x"]);
+        let schemes = v.supported_verify_schemes();
+        assert!(
+            schemes.len() > 3,
+            "suspiciously short scheme list: {schemes:?}"
+        );
+        assert!(schemes.contains(&SignatureScheme::RSA_PKCS1_SHA256));
+        assert!(schemes.contains(&SignatureScheme::ECDSA_NISTP256_SHA256));
+    }
+}

socktop_connector/src/types.rs

@@ -54,6 +54,10 @@ pub struct GpuInfo {
 
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct Metrics {
+    /// Epoch ms when the agent actually collected this snapshot (agents may
+    /// serve TTL-cached data). Absent on agents older than 1.60.
+    #[serde(default)]
+    pub sampled_at_ms: Option<u64>,
     pub cpu_total: f32,
     pub cpu_per_core: Vec<f32>,
     pub mem_total: u64,
@@ -147,6 +151,10 @@ pub struct JournalResponse {
     pub entries: Vec<JournalEntry>,
     pub total_count: u32,
     pub truncated: bool,
+    /// Agent-side explanation for an empty result (journal access limits).
+    /// Absent on agents older than 1.60.
+    #[serde(default)]
+    pub notice: Option<String>,
     pub cached_at: u64, // Unix timestamp when this data was cached
 }
 

socktop_connector/src/wasm/requests.rs

@@ -46,6 +46,7 @@ pub async fn send_request_and_wait(
                 // For now, return a placeholder metrics response indicating binary data received
                 // TODO: Implement proper protobuf decoding for binary data
                 let placeholder_metrics = Metrics {
+                    sampled_at_ms: None,
                     cpu_total: 0.0,
                     cpu_per_core: vec![0.0],
                     mem_total: 0,

socktop_wasm_test/Cargo.lock

@@ -475,9 +475,7 @@ dependencies = [
 
 [[package]]
 name = "socktop_connector"
-version = "0.1.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3a63dadaa5105df11b0684759a829012257d48e72a469cc554c0cf4394605f5a"
+version = "1.51.0"
 dependencies = [
  "flate2",
  "js-sys",

socktop_wasm_test/Cargo.toml

@@ -10,8 +10,8 @@ edition = "2021"
 crate-type = ["cdylib"]
 
 [dependencies]
-# Use WASM features for WebSocket connectivity (published version)
-socktop_connector = { version = "0.1.5", default-features = false, features = ["wasm"] }
+# Use WASM features for WebSocket connectivity (in-repo connector via path)
+socktop_connector = { path = "../socktop_connector", default-features = false, features = ["wasm"] }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
 wasm-bindgen = "0.2"

zellij_socktop_plugin/Cargo.toml

@@ -3,6 +3,9 @@ name = "zellij_socktop_plugin"
 version = "0.1.0"
 edition = "2021"
 
+# Standalone package, not part of the parent workspace (same as socktop_wasm_test)
+[workspace]
+
 [lib]
 crate-type = ["cdylib"]
 
@@ -10,7 +13,7 @@ crate-type = ["cdylib"]
 zellij-tile = "0.40.0"
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
-socktop_connector = { version = "0.1.5", default-features = false, features = ["wasm"] }
+socktop_connector = { path = "../socktop_connector", default-features = false, features = ["wasm"] }
 futures = "0.3"
 
 [dependencies.chrono]