4 Commits

Author	SHA1	Message	Date
jasonwitty	bedbe0a2ec	Add flag to override logic and supress terminate option. (--no-kill) ... Flag specifically used to block feature on socktop.io. Will remain undocumented for standard usage.	2026-08-24 07:19:37 -07:00
jason	fe3ef7f25e	fix(ci): pin deb builds to ubuntu-22.04 and enforce the fleet glibc floor (#41) ... * fix(ci): pin deb builds to ubuntu-22.04 and enforce the fleet glibc floor The v1.60.0 debs failed to install on Raspberry Pi OS bookworm: socktop : Depends: libc6 (>= 2.39) but 2.36-9+rpt2+deb12u14 is to be installed Cross-compiled binaries link against the RUNNER's (multiarch) glibc, so the runner picks the minimum glibc the packages demand. ubuntu-latest migrated from 22.04 (glibc 2.35) to 24.04 (glibc 2.39) between the 1.50.x releases and now, silently raising the requirement past the Debian-12 fleet. Pin the build job to ubuntu-22.04 (2.35 — satisfied by bookworm's 2.36) and add a post-build gate that reads each .deb's computed libc6 requirement and fails the run if it exceeds the fleet floor, so the next runner migration turns into a red build instead of a fleet-wide apt error. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ci): extract the libc6 version, not the 6 in 'libc6' The floor gate's second grep matched the trailing digit of the package name before the version ('libc6 (>= 2.34)' -> '6'), failing every target. sed capture group instead; verified against realistic Depends strings. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * 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> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>	2026-08-23 09:42:39 -07:00
jason	4c59716610	Kill a local process from the TUI, and stop the agent reporting dead ones (#40) ... * fix(agent): drop processes that no longer exist `refresh_processes_specifics` was called with remove_dead_processes = false against a long-lived System, so the agent accumulated every process it had ever seen and went on reporting them. Measured on a Pi 5 after a few hours of build churn: 21,648 processes reported, 289 actually running, growing a few every poll. Three consequences, in ascending order of how confusing they are: * unbounded memory growth, and every poll iterating ~75x more entries than it should * process_count — the client's "Top Processes (N total)" — is meaningless * a process you kill keeps its row forever, because the agent keeps sending it. Killing it again reports "no longer exists", since the kernel is telling the truth and the agent is not. The third is how this was found: no amount of client-side reconciliation could fix a list whose producer never forgets anything. Passing `true` is only correct because these two sites use ProcessesToUpdate::All. With `Some(pids)` sysinfo treats every process outside the list as dead and removes it, so the per-PID refresh in collect_process_metrics must keep `false` — noted in a comment there. Verified by driving the TUI against a rebuilt agent: reported count 292 vs 290 real, and a killed row is removed once and never reappears. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * feat(socktop): kill a local process from the TUI btop-style process termination, for local agents only. The signal is sent by socktop itself through a direct sysinfo call — nothing is transmitted to the agent, and the agent and connector have no kill capability at all. Why local-only: the PIDs on screen are reported by the agent, and the signal is sent with socktop's own OS privileges. A PID is therefore only meaningful, and only safe to act on, when the agent lives on this machine; acting on a remote agent's PIDs would signal whatever unrelated local process happened to share that number. local::agent_is_local treats an address as local when it is loopback or when an ephemeral bind succeeds (which only works for an address on one of our own interfaces, so it also covers reaching our own agent by LAN IP), requires every address a hostname resolves to to be local, and fails closed. * `t` on the selected process, and `t` inside Process Details. One key for both: `k` scrolls the thread table in the modal, so it could not be reused there. * The confirmation offers Terminate (focused first, so a reflexive Enter is the safe one), Force kill, and Cancel. Keeping SIGKILL behind a second button rather than a second keybinding means the destructive option has to be chosen deliberately. * The selection hint gained the key, but only for a local agent — advertising a key that deliberately does nothing is worse than no hint. Same for the details modal's help line. The list is reconciled after a signal rather than left to the next poll. A signalled PID goes on a watch list re-checked each metrics tick, because SIGTERM is a request: the process is usually still alive at signal time, and its row should go when it actually exits (or stay, if it ignores the signal). PIDs confirmed gone are remembered briefly, since the agent serves Processes from a 1500ms cache and would otherwise hand back a pre-kill snapshot. A selection whose process has left the list is dropped, and the details view closes for a process that no longer exists — including when it dies on its own, which previously flipped that modal to "Agent Update Required" because the wire cannot distinguish "no such PID" from "endpoint unsupported". Also fixes two pre-existing UI faults found on the way: * The selection hint was sized from its full label including the process name and skipped entirely when that exceeded the pane width — so it vanished exactly when a long-named process was selected. The name is now the elastic part, and widths are measured in columns rather than bytes. * Confirmation and Info dialogs laid their content out over the whole modal rect instead of the block's inner rect, putting the first line of text on the border row, and fell into the catch-all 70%x50% sizing arm, so a one-line question got half the screen. They now size to their content and use the theme's colors like the connection-error modal does. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(kill): close stacked details view, guard PID reuse, scale settle with interval Review fixes for the process-kill feature: 1. Killing from INSIDE the details view left it open forever, frozen on the dead process (reproduced live): the 'Signal sent' Info modal sits on top when the death is confirmed on the next tick, the old top-only close_process_details missed it, gone-PIDs are processed once, and the details-poll fallback was gated on the selection the kill had just cleared. The close now removes the dead PID's view wherever it sits in the stack — which also retires a dead parent's view from under a child in a navigation chain, so backing out lands on the process list rather than a frozen corpse view. Tests updated to the new semantics, plus a regression test for the Info-stacked case. 2. PID-reuse guard: the PID comes from an agent snapshot and the confirmation can sit open indefinitely, so by signal time the kernel may have recycled the number. kill_local_process now takes the name the user confirmed and refuses to signal a PID whose current owner does not match ('PID N now belongs to X, not Y'). 3. A transient request error no longer closes the details view: with process_details_answered set, any Err was read as 'process gone', including socket blips. The view now closes only when the PID is also absent from the agent's own process list. 4. PROC_CACHE_SETTLE scales with the user's processes interval (floor at the old 1.6s default-TTL value), and the tombstone lifetime rides on top of it — users who raise the agent's Processes TTL raise the client interval to match, so the interval is the best client-side signal for how stale an agent snapshot can be. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(kill): resume polling for the details view that resurfaces from a chain Reported: kill still orphaned a window via Enter (child details) -> P (parent details) -> t (terminate parent). The parent's view closed correctly, but the child's view underneath resurfaced with no selection (forget_process_row had cleared it — it pointed at the parent) and wiped data; the selection-gated details poll never refilled it. close_details_for_gone_process now retargets the selection to the uppermost remaining ProcessDetails view (looking through stacked Info/Confirmation modals) and makes the poll due immediately — the same retarget SwitchToParentProcess performs on the way down the chain. Same class of hole in plain navigation, fixed alongside: Esc-ing back from a parent view left the selection on the PARENT, so the resurfaced child-titled view refilled with the parent's data. The dismiss arm now retargets the selection to whatever details view it lands on. Both verified live: child -> P -> kill parent -> child view resurfaces populated and updating; child -> P -> Esc -> child view shows the child. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs: changelog section for the process-kill feature and agent dead-process fix Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-23 08:11:17 -07:00
jason	0322308896	housekeeping-p2: security, correctness, and performance pass before 1.51 (#39) ... * chore: dead-code sweep - Delete socktop_connector/src/connector.rs: orphaned since 08f248c removed 'pub mod connector;' during the modularization refactor. Never compiled (verified under default, wasm, and workspace feature combos) but shipped in the crates.io tarball and contained an outdated copy of the TLS verifier — a trap for anyone patching the pinning bug in the dead copy. - Delete empty socktop/src/ws.rs, tracked editor backup ui/.modal.rs.backup, and stray test_thiserror.rs at the repo root. - Delete the two LEGACY #[allow(dead_code)] process input handlers; the header-click render test now exercises the live _with_selection handler instead (better coverage of the real path). - Drop unused sysinfo dependency from the socktop client. - Replace stale 'temporarily increased for testing' comment on COMPRESSION_THRESHOLD (it already held the production value). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(connector): make certificate pinning real; disable Nagle Security: with --verify-hostname off (the default), the old NoVerify verifier accepted ANY server certificate — the CA loaded from --tls-ca was never consulted, so the documented pinning was a no-op and the connection was trivially MITM-able. Replace it with PinnedCertVerifier: the presented end-entity cert must be byte-identical to a cert in the --tls-ca file (any cert in a multi-cert PEM matches, supporting rotation). Signature validation now uses the ring provider's full algorithm set instead of a hardcoded 3-scheme list. Empty PEM files fail fast instead of failing closed per-handshake. The --verify-hostname path is unchanged (WebPki root-store validation). Also: the third argument of connect_async_tls_with_config is tungstenite's disable_nagle flag, not a verification toggle — we were passing verify_hostname there, leaving Nagle ON for default users. Pass true unconditionally, and disable Nagle on the plain ws:// path too; socktop exchanges small request/response frames where Nagle only adds latency. Client now consumes the connector via a dual path+version dep so these fixes are in local builds and CI before the crates.io publish (cargo strips the path on publish). Connector version -> 1.51.0. Verified E2E: agent A's cert connects to agent A; agent B's cert against agent A fails the handshake (the rpi-worker-1 wrong-PEM scenario); --verify-hostname against a 127.0.0.1 SAN still connects. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(agent): GPU worker thread, async journalctl, correctness + cache fixes Lightweight: - GPU collection moves to a dedicated worker thread that owns the gfxinfo handle for the process lifetime. gfxinfo's active_gpu() runs a full NVML init/teardown (~20ms, blocking) and we were paying it on the async runtime for every collect — measured at ~80% of the agent's entire active CPU on a GPU machine. The handle holds Rc<Nvml> (not Send), so a thread + mpsc/oneshot channel pair confines it; a zero-total-VRAM reply is treated as a dead session (driver reload) and re-probed. - journalctl now runs via tokio::process instead of blocking one of the two runtime workers for the duration of the subprocess. - TtlCell (state.rs) replaces the four hand-rolled static TTL caches; a cached negative result now counts as fresh, so hosts with no matching temp sensor or GPU stop rescanning every request. Single lock+clone on the GPU cache hit path (was two). Correctness: - Process/child CPU times are now microseconds as documented; they were milliseconds, rendering 1000x too small next to (correct) thread times. - Non-Linux per-process CPU%% clamps AFTER dividing by core count; a 4-cores-busy process on an 8-core box reported 12.5% instead of 50%. - Journal timestamps are real RFC 3339 UTC plus an additive timestamp_us field (sorting is now numeric); the old strings were Debug-formatted SystemTime mangled by string replace. - Partition detection uses /sys/block on Linux: whole-disk filesystems on names like nvme0n1 or zram1 are no longer misclassified as partitions. One shared parent_disk_name() replaces two inline copies. - New sampled_at_ms on the metrics payload (additive) records when the snapshot was actually collected, so clients can compute exact rates across the agent's TTL cache. Security/robustness: - key.pem is created 0600 (was umask default 0644, world-readable) and pre-1.51 keys are tightened on startup. - Per-PID detail/journal caches now evict (60s max age, 64 entries max); they previously grew without bound under PID-walking clients. - The two per-PID ws handlers collapse into one generic helper. - /proc/<pid>/stat parsing unified in one comm-safe module. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(tui): responsive input, request timeouts, poisoned-stream reconnect R1 — input latency: the event loop drained input once per iteration, then slept the whole metrics interval; keys and wheel events queued for up to 500ms (or the full interval at slower rates) and applied in bursts. The input block is extracted to drain_input() and the tail sleep replaced by a deadline wait in <=33ms poll slices that handles and repaints input the moment it arrives. Verified: help modal opens <150ms into a 2000ms tick. R2 — freeze-proofing: metrics/processes/disks requests had no timeout; a half-dead connection left ws.next() pending forever and froze the TUI with no way to quit (raw mode eats Ctrl+C as an unread key event). All requests now carry a 5s budget. C3 — desync: replies are matched to requests by order alone, so a timed- out request's late reply would shift every subsequent reply off by one. Any timeout now treats the stream as poisoned and goes through the reconnect flow — a fresh stream is aligned by construction. The modal endpoints additionally mark process details unsupported (flag resets on modal close/selection change) so a detail-less agent doesn't cause a reconnect loop. While disconnected the fetch path idles: recovery belongs to the manual/auto retry paths instead of 5s-timeout hammering. C7 — fit::truncate_middle_cols replaces util::truncate_middle: display- width aware and char-boundary safe; the byte-slicing version panicked the draw loop on non-ASCII device names. Verified live: agent kill -9 mid-session -> error modal in <3s, q exits while disconnected, r reconnects and resumes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * chore: version 1.51.0, path-dep the wasm examples, README notes - socktop, socktop_agent, socktop_connector -> 1.51.0. - socktop_wasm_test and zellij_socktop_plugin consume the in-repo connector via path deps so wasm-feature API drift is caught at PR time instead of after publish. Immediately proved out: the wasm requests module needed the new sampled_at_ms field, invisible to native builds. - zellij plugin gains the standalone [workspace] marker (it could not be cargo-checked in-tree at all before). NOTE: its lib.rs has pre-existing compile errors unrelated to the connector (static mut STATE conflicts with register_plugin!, missing BTreeMap import) — needs its own rework, out of scope here. - README: sampled_at_ms in the example payload. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(review): restore Agent Update Required flow, command field, axis alignment Fixes from Jason's hands-on verification of the branch: 1. Old-agent messaging regression (this branch): a detail-request timeout went through the loud poison/reconnect flow, burying the ProcessDetails modal's 'Agent Update Required' message under a connection-error modal. Old agents IGNORE unknown messages (no late reply, no desync), so the optional per-PID endpoints now use quiet_reconnect(): swap the stream silently (still safe against merely-slow agents) and let the modal show its message. Only a failed reconnect surfaces loudly. Verified against a real v1.40.0 agent: message shows, session stays healthy. 2. Draw starvation (this branch): an agent that never answers get_metrics put the loop in fetch->timeout->poison->restart cycles that never reached the draw call — permanently blank TUI. The iteration now paints before fetching, and a second consecutive metrics timeout trips a circuit breaker: persistent 'Agent is not responding' error, recovery left to the manual/30s retry paths. Verified against a 0.9-era agent. 3. Command & Details pane blank (pre-existing on master): the minimal- refresh optimization dropped cmd/exe/cwd from the detail endpoint's ProcessRefreshKind, so process.cmd() had nothing to return. Restored with UpdateKind::OnlyIfNotSet — immutable values, read once per PID. Regression test added; journal E2E re-verified (100 entries render). 4. Scatter-plot axis misalignment: Y labels used a fixed 4-char field from the era when CPU times were 1000x too small; honest millisecond values (e.g. 136114) blew through it. Labels now right-align to the widest value per frame and X labels/titles share the dynamic padding. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat: journal access notice, 1.60.0, install script, changelog, riscv protoc fallback - Journal pane now distinguishes 'no entries' from 'no journal access': journalctl exits 0 with empty output when the agent's user simply can't see the target's entries (demo mode / user-run agents), explaining itself only on stderr. The agent forwards that hint as an additive JournalResponse.notice and the client renders it with practical advice. Verified E2E via a stub journalctl emulating the unprivileged case. - Version 1.60.0 across all crates (1.51 would read fine, but the repo's scheme is 1.40/1.50/…, and a literal 1.6.0 would sort BELOW 1.50.0 in semver). All user-facing version strings already come from CARGO_PKG_VERSION — a stale binary was the only way to see an old one. - scripts/install.sh: build-from-source install/upgrade for the test fleet (Linux + macOS). Detects in-repo checkouts, installs rustup when missing, replaces a systemd socktop-agent service binary in place and restarts it, requires system protoc on riscv64. - build.rs (agent + connector): fall back to $PROTOC / PATH when protoc-bin-vendored has no binary for the host (riscv64) — native SBC builds previously panicked in the build script. - CHANGELOG.md covering v1.50.0 -> 1.60.0. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test: add notice field to cache test initializer Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: untrack zellij plugin build dir; installer updates all PATH copies - Remove zellij_socktop_plugin/target from git (3,577 files committed by accident in bf6ac87): the root .gitignore anchors /target to the repo root, so the standalone plugin's own build dir wasn't covered. Ignore target/ at any depth (also fixes the pre-existing '/socktop-wasm-test/target' entry, which pointed at a hyphenated path that doesn't exist). - install.sh now updates EVERY copy of socktop/socktop_agent on PATH, not just $PREFIX: a stale 'cargo install' in ~/.cargo/bin shadows /usr/local/bin on most PATHs, so an install could 'succeed' while 'socktop --version' kept reporting the old release. Extra copies that can't be written are warned about, not fatal, and the script now prints which binary is actually active on PATH at the end. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(installer): manage the socktop-agent systemd service Upgrade path (unit already present): NEVER touch the unit file — it is the operator's config (SSL, tokens, ports live there as Environment= lines). Only the binary at the unit's own ExecStart path is replaced, then the service restarts. Flags/args preserved by construction. Fresh path (no unit): full first-time setup mirroring the deb postinst and the agent-service docs — create the socktop system user/group and /var/lib/socktop, install docs/socktop-agent.service (ExecStart rewritten to wherever this run installed the agent; embedded fallback for old refs), daemon-reload, enable --now, and print how to turn on TLS/token. Also: system-level operations get their own sudo decision (SYS_SUDO) — previously they inherited the PREFIX sudo flag, so a writable --prefix made the service section run groupadd/systemctl unprivileged and die. No sudo at all now skips service management with a warning instead of failing the install. Both branches dry-run verified with stubbed systemctl/sudo. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(installer): don't bind fresh agent services onto occupied ports The Orange Pi install put the new service straight into a crash-restart loop: the unit's default --port 3000 collided with a Docker service already publishing 3000 (Umami; Gitea and friends default there too). Fresh installs now scan 3000/3001/3010/3231/3232 via ss and configure the unit on the first free port, warning loudly when 3000 was taken and printing the resulting ws:// URL. Upgrades still never touch the unit. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(agent): detect NVIDIA GPUs on distros without the unversioned NVML soname On Debian and derivatives the NVIDIA driver ships only libnvidia-ml.so.1 (the unversioned symlink belongs to the dev package), and nvml-wrapper's default init dlopens the unversioned name — so gfxinfo reported 'No GPU found' on a fully functional RTX A2000 host while nvidia-smi worked fine. Arch-family distros ship the symlink, which is why the desktop never showed this. The GPU worker now falls back to initializing NVML directly with the versioned soname when gfxinfo's probe fails, collecting name/util/vram through the same handle-caching path. nvml-wrapper was already in the tree via gfxinfo — same version, no new build cost. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(agent): box the NVML handle variant (clippy large_enum_variant) CI clippy runs with -D warnings; Nvml is a large struct next to the 16-byte Box<dyn Gpu> variant. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(installer): survive self-modification mid-run; sturdier unit detection Root cause of the mixed-up second install on the A2000 host: when run from the clone it manages, the script's own git checkout/merge REPLACES scripts/install.sh while bash is still executing it. Bash reads scripts lazily by byte offset, so it resumed parsing the NEW file at the OLD offset and executed an arbitrary tail of it — observed as the fresh- service path running on a host whose unit already existed: the port scan saw the still-running old service on 3000 and silently wrote a new unit on 3001, while enable --now on the already-active service changed nothing until a manual daemon-reload. Fix: the whole script now runs inside main(), invoked as 'main "$@"; exit $?' so bash parses everything up front and never reads the file again after main returns (the exit lives in the same parse unit — demonstrated necessary: with a bare 'main "$@"' ending, bash still executed the swapped file's trailing content after main returned). Also: unit existence is now checked with 'systemctl cat' instead of grepping the full list-unit-files output. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs(installer): use a durable ref in the usage example Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>	2026-08-21 12:45:10 -07:00