Overlord/porthole
1
1
Fork 0
Code Releases 1 Activity
Files
0c9f0585fa7dd5be287a26973e8df0534a0991f2
porthole/spec/porthole-spec.md
Overlord bc8f1cc5d2 Fix --via/-J grammar and --force process-group kill in spec 
Found both while translating the spec into code:

- `ssh -J a,b,c` alone isn't valid - ssh -J only carries jump hosts
  before the final hop; it still needs a positional destination. --via's
  last entry has to double as that target, which also makes --via
  mandatory (there's no other field naming "the host ssh connects to"),
  reversing the earlier "no" in the required column.
- A plain single-pid SIGKILL on `close --force` would leave the `ssh`
  child orphaned, since a killed process can't forward anything to it.
  Documents sending SIGKILL to the whole process group instead, relying
  on the supervisor's own setsid() call making it the group leader.
2026-08-13 15:43:36 +02:00

462 lines
23 KiB
Markdown
Raw Blame History

# `porthole` — Spec v0.2
Named, managed SSH port forwards. Wraps `ssh -L/-R/-D` so forwards are
addressable by name instead of by PID, terminal tab, or shell history.
---
## 1. Overview
**Problem:** SSH forwards are anonymous and ephemeral. They die when a
terminal closes, when a laptop sleeps, or when a network blips — silently,
with no notification. There's no built-in way to list what's currently
forwarded, and multi-hop / reverse forwards have enough flag surface that
people end up hand-rolling shell aliases per-tunnel.
**Solution:** Persist forward definitions as named profiles. Run forwards as
a supervised background process (not tied to a shell session), with
auto-reconnect, health status, and a single command to see everything that's
open.
**Non-goals:** Not a replacement for a VPN or a full SOCKS/proxy manager.
Not a secrets manager — SSH auth still comes from your existing SSH config,
agent, or identity files. No GUI. Surviving a full reboot/logout is also
out of scope for v0.1 — see §8.
---
## 2. Data model
### 2.1 Profile
A saved definition. Does not imply anything is running.
| Field | Type | Notes |
|------------------|-----------|-----------------------------------------------------------------------|
| `name` | string | Unique key. `[a-z0-9_-]+`, 1-64 chars (matches vmic's naming rule). |
| `kind` | enum | `local` \| `remote` \| `dynamic` |
| `mapping` | string | Raw `-L/-R/-D` payload, see §5.1 |
| `via` | string[] | Ordered hop list; each entry `[user@]host[:port]` (see §3.1) |
| `user` | string? | Defaults to current user / `ssh_config` |
| `identity` | path? | Identity file override |
| `ssh_port` | int | Default `22`; final target only, see §5.1 |
| `reconnect` | bool | Default `true` |
| `retry_interval` | int (sec) | Default `5`; base reconnect delay (§4.1) |
| `backoff_max` | int (sec) | Default `60`; cap on the doubling reconnect delay (§4.1) |
| `keepalive` | int (sec) | `ServerAliveInterval`, default `15` |
| `created_at` | timestamp | |
| `updated_at` | timestamp | |
### 2.2 Instance (runtime state)
Exists only while a profile is open. Tracked separately from the profile so
`list`/`status` can report live data without touching the saved definition.
**The absence of an instance file is what "closed" means** — there is no
separate `down` state; see §3 for exactly when the file is created/removed.
| Field | Type | Notes |
|---------------------|------------|----------------------------------------------------------------|
| `name` | string | FK to profile |
| `pid` | int | Supervisor process PID, not raw `ssh` PID |
| `state` | enum | `up` \| `reconnecting` \| `error` |
| `opened_at` | timestamp | Anchor for "session uptime" (§5.5) — set once, at `open` |
| `connected_at` | timestamp? | Start of the *current* unbroken connection; resets each reconnect (§5.5) |
| `last_error` | string? | Most recent failure message, if any |
| `reconnect_count` | int | Since last manual `open` |
| `last_reconnect_at` | timestamp? | |
### 2.3 Storage
- Profiles: `~/.config/porthole/profiles/<name>.toml`
- Runtime state: `~/.local/state/porthole/<name>.json` (written by the
supervisor, not hand-edited; absence means the profile is closed)
- Lock: `~/.local/state/porthole/<name>.lock` (advisory `flock`, held for
the supervisor's entire lifetime — see §3)
- Logs: `~/.local/state/porthole/<name>.log`, rotated to a single
`<name>.log.1` backup once it exceeds 10 MiB (checked on each reconnect
attempt, not per line — these are meant to run for months, unlike vmic's
short-lived CLI invocations)
---
## 3. Supervisor architecture
`open` has to hand off to a process that keeps running after the invoking
shell/terminal exits — the same one-shot-CLI-can't-host-a-daemon problem
`vmic` solves by spawning detached `pw-loopback` subprocesses tracked by
pid. porthole has no external long-running helper to shell out to (there's
no `ssh-loopback` equivalent), so it supervises `ssh` itself via a hidden
re-exec of its own binary:
1. `porthole open <name>` validates the profile, tries to acquire
`<name>.lock` (if already held by a live pid: no-op, print status, exit
0 — see §5.2), then spawns *itself* via `std::env::current_exe()` with a
hidden subcommand: `porthole __supervise <name>`.
2. The spawned process detaches before doing anything else: `stdin` from
`/dev/null`, `stdout`/`stderr` appended to `<name>.log`, and
`libc::setsid()` called via `CommandExt::pre_exec` so it leaves the
parent's process group/session — it survives the terminal closing and
doesn't receive the shell's Ctrl-C/SIGHUP.
3. The foregrounding `open` call blocks briefly (bounded, a few seconds)
waiting for the supervisor to write its pid + initial `state` into the
instance file, so callers get an accurate exit code for immediate
failures (bad auth, port conflict) instead of racing a background
process. `-f/--foreground` skips the detach step entirely and runs the
supervisor loop inline, attached to the current session.
4. The supervisor's loop: spawn `ssh` with the flags in §3.1, wait on it,
classify the exit per §4, sleep/backoff or give up accordingly, and
rewrite the instance file after every state change. A signal handler
installed on the supervisor (not exit-code inference on `ssh` itself,
which is unreliable) is what distinguishes an intentional `close` from
a dropped connection: `close` sends SIGTERM to the *supervisor* pid,
whose handler kills its `ssh` child, waits briefly, deletes its lock and
instance file, and exits — any *other* way the `ssh` child ends (any
exit status) is treated as a failure to reconnect from, per §4. Because
`step 2`'s `setsid()` makes the supervisor its own process group leader
and `ssh` inherits that group, `--force` sends SIGKILL to the whole
group (`kill(-pid, SIGKILL)`) rather than just the supervisor pid — a
plain single-pid SIGKILL would leave `ssh` running, orphaned and
untracked, since a killed process can't forward anything to its child.
A supervisor that exits on its own due to a fatal failure (§4.2) leaves
the instance file in place with `state: error` rather than deleting it,
so the failure stays visible to `status`/`list` until the user acts.
### 3.1 Forced `ssh` flags
Every `ssh` invocation porthole spawns gets these, non-configurable,
regardless of the user's own `~/.ssh/config`:
| Flag | Why |
|--------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------|
| `-o BatchMode=yes` | A headless supervised process must never block on a password/passphrase/host-key TTY prompt — without this, a first connection to an unknown host or a locked key just hangs forever, indistinguishable from "reconnecting." |
| `-o ExitOnForwardFailure=yes` | Makes `ssh` exit non-zero immediately if the requested forward can't be bound, instead of staying up as a plain (forward-less) session that *looks* healthy. |
| `-o ConnectTimeout=10` | Bounds how long one connection attempt can hang before porthole's own backoff logic (§4.1) gets a turn. |
| `-N` | No remote command — porthole only ever wants the forward, never a shell. |
| `-J <hops>` + positional target | See below — `--via`'s *last* hop is the actual connection target, not another jump. |
`ServerAliveInterval` comes from the profile's `keepalive` field (not
hardcoded), so it stays user-tunable.
**`--via` → `ssh` argument translation:** `ssh -J a,b,c` is not itself a
valid invocation — `-J` only ever carries jump hosts *before* the final
hop; `ssh` still needs a positional `destination` to actually connect (and
run the forward from). So porthole splits `--via`'s comma list at the last
entry: everything before it becomes `-J`'s value (omitted entirely if
`--via` has only one hop), and the last entry becomes `ssh`'s positional
target argument. `--via jumpbox` → `ssh ... jumpbox` (no `-J`). `--via
bastion1,bastion2` → `ssh -J bastion1 ... bastion2` (connect through
bastion1, forward runs from bastion2). This is also why `--via` is
**required**, not optional (see §5.1) — there is no other field
representing "the host `ssh` actually connects to"; `--via`'s last hop
*is* that field.
---
## 4. Reconnect & failure handling
### 4.1 Backoff
- Base delay is the profile's `retry_interval` (default 5s).
- Each consecutive failed attempt doubles the delay, capped at
`backoff_max` (default 60s).
- The backoff counter resets to the base delay once a connection has
stayed up continuously for 60s — an in-memory supervisor detail, not
persisted to the instance file — so one good connection after a flaky
patch doesn't leave the *next* reconnect waiting a full capped delay.
- If `reconnect` is `false`, there is no retry loop at all: a single
failed attempt goes straight to `state: error` and the supervisor exits.
### 4.2 Failure classification
`ssh`'s own exit code doesn't reliably distinguish "retry me" from "stop
retrying," so porthole classifies by matching `ssh`'s stderr:
| Pattern (stderr substring) | Class | Behavior |
|-----------------------------------------------------------------|------------|--------------------------------------------------|
| `Permission denied` | fatal | `state: error`, stop, supervisor exits |
| `Host key verification failed` | fatal | `state: error`, stop, supervisor exits |
| `bind: Address already in use` | fatal | `state: error`, stop, supervisor exits |
| `Connection refused` / `No route to host` / connect timeout | transient | backoff + retry |
| `Could not resolve hostname` | transient | backoff + retry (bounded by `backoff_max`) |
| anything else / unrecognized | transient* | backoff + retry, but see escalation below |
\* After 10 consecutive unrecognized failures in a row, porthole treats the
profile as effectively broken (`state: error`, stop) rather than retrying
under an unrecognized failure forever. `last_error` always holds the raw
`ssh` message either way, for `status` to show.
A profile that lands in `state: error` stays that way — including its
instance file — until the user runs `open` again (fresh attempt, fresh
backoff) or `close` (clears it). It is deliberately *not* self-healing past
a fatal classification.
---
## 5. Commands
### 5.1 `porthole add <name> [flags]`
Aliases: `create`, `new`
Saves a new profile. Does **not** open it.
| Flag | Arg | Required | Default | Description |
|--------------------|-------------------------------|-------------------|---------------------------|----------------------------------------|
| `-l, --local` | `[bind:]port:host:hostport` | one of `-l/-r/-d` | — | Local forward: your machine → remote |
| `-r, --remote` | `[bind:]port:host:hostport` | one of `-l/-r/-d` | — | Remote forward: remote → your machine |
| `-d, --dynamic` | `[bind:]port` | one of `-l/-r/-d` | — | Dynamic forward (SOCKS proxy) |
| `--via` | `[user@]host[:port][,...]` | **yes** | — | Comma-separated hop chain; the *last* hop is the actual `ssh` connection target, any before it are `-J` jumps (§3.1) |
| `-u, --user` | `user` | no | current user / ssh_config | Default user for the final target and any `--via` hop that doesn't specify its own |
| `-i, --identity` | `path` | no | ssh_config default | |
| `-p, --port` | `port` | no | `22` | SSH port on the final target only — a `--via` hop needs its own inline `:port` if it isn't 22 |
| `--reconnect` | `bool` | no | `true` | Auto-reconnect on drop (§4) |
| `--retry-interval` | `seconds` | no | `5` | Base reconnect delay (§4.1) |
| `--backoff-max` | `seconds` | no | `60` | Cap on the doubling reconnect delay (§4.1) |
| `--keepalive` | `seconds` | no | `15` | `ServerAliveInterval` |
Exactly one of `-l`/`-r`/`-d` is required. Providing more than one is an
error.
**Validation:**
- `name` must not already exist (use `edit` to modify); 1-64 chars,
`[a-z0-9_-]+`.
- `mapping` port syntax validated against the same grammar `ssh` accepts;
stored verbatim as whichever `-l/-r/-d` payload was given (without the
flag itself) — `kind` records which one it was, so re-deriving the right
`-L`/`-R`/`-D` flag at `open` time is a lookup, not a re-parse.
- `--via` requires at least one hop (its last entry is the mandatory
connection target, see §3.1); hosts are resolved/checked against
`~/.ssh/config` if present, but not required to exist there.
**Examples:**
```
porthole add db --local 5432:db.internal:5432 --via jumpbox
porthole add admin-ui --local 8080:localhost:8080 --via bastion1,bastion2:2222 --user ops
porthole add webhook --remote 9000:localhost:3000 --via public-vps
porthole add proxy --dynamic 1080 --via edge-host
```
---
### 5.2 `porthole open <name> [flags]`
Alias: `start`
Starts a saved forward as a background-supervised process (§3).
| Flag | Description |
|--------------------|--------------------------------------------------------------------------------------------------------------------|
| `-f, --foreground` | Run attached in current shell instead of detaching. Ctrl-C closes it cleanly (removes the instance file, same as `close`). |
| `--once` | Open without auto-reconnect, regardless of profile setting (§4.1). |
| `--all` | Ignore `<name>`; open every profile with `reconnect: true` that isn't already running. Per-profile failures are warnings, not a whole-batch failure — this exists specifically as the hook for external autostart mechanisms, see §8. |
**Behavior:**
- If already open (a live supervisor pid holds `<name>.lock`): no-op,
print current status, exit 0.
- If the instance file exists but its pid is dead (crash, or the machine
rebooted): treated as not-running, proceeds to spawn a fresh supervisor.
- If port bind fails (already in use): exit non-zero with the conflicting
process info if discoverable (`lsof`-style lookup — best-effort, degrades
to a plain "port in use" message if `lsof`/`ss` isn't on `PATH`), don't
silently retry.
- `open` blocks briefly (bounded, a few seconds) waiting for the detached
supervisor to confirm it's alive, so an immediate failure (bad auth,
bind conflict) reports a non-zero exit rather than appearing to succeed.
Anything that fails *after* that point (a later reconnect) only shows up
via `status`/`list`, not `open`'s own exit code.
**Examples:**
```
porthole open db
porthole open db --foreground
porthole open proxy --once
porthole open --all
```
---
### 5.3 `porthole close <name> [flags]`
Alias: `stop`
Stops a running forward and removes its instance file. Profile definition
is untouched.
| Flag | Description |
|-----------|----------------------------------------------------------------------------------|
| `--force` | SIGKILL the supervisor (and its `ssh` child) immediately instead of graceful SIGTERM + wait |
`close` on a profile that's already stopped (no live pid) is a no-op, exit
0 — it still clears a stale instance file left over from a crash, same as
the crash-recovery path in `open`.
**Examples:**
```
porthole close db
porthole close db --force
```
---
### 5.4 `porthole edit <name> [flags]`
Updates a saved profile. Accepts the same flags as `add` (all optional —
only provided flags are changed).
**Decision:** `edit` never restarts a running instance, and there is no
`--restart` flag. If `<name>` is currently running, `edit` prints a warning
that the change won't take effect until the next `open`/`close` cycle and
exits 0 — consistent with vmic's `edit`, which never auto-migrates a live
topology without telling the user exactly what to run instead. Keeping
this explicit avoids a footgun where editing a profile silently bounces a
tunnel someone else might be relying on.
**Examples:**
```
porthole edit db --retry-interval 10
porthole edit db --local 5433:db.internal:5432
```
---
### 5.5 `porthole status <name>`
Deep-dive health for one forward.
| Flag | Description |
|----------|-----------------------------------------------------|
| `--json` | Machine-readable output (matches `list --json`) |
**Output includes:**
- Profile summary (kind, mapping, via, user)
- Current state (`up` / `reconnecting` / `error`, or `closed` if no
instance file exists at all — §2.2)
- **Session uptime**: elapsed time since `opened_at` (the original `open`
call), regardless of intervening reconnects
- **Connection uptime**: elapsed time since `connected_at` (the current
unbroken connection) — resets on every reconnect, absent while
`reconnecting`/`error`
- Reconnect count and timestamp of last reconnect
- Last error message, if any
- Path to log file
**Example:**
```
porthole status db
```
---
### 5.6 `porthole list`
Alias: `ls`
All saved profiles with live status. Fast, scannable — no deep diagnostics
(use `status` for that).
**Columns:** `NAME KIND MAPPING VIA STATE UPTIME`
A profile with no instance file shows `STATE: closed` and an empty
`UPTIME`. `UPTIME` otherwise shows connection uptime (§5.5).
**Flags:**
| Flag | Description |
|-------------|-------------------------------------------------------------------|
| `--running` | Show only currently-open forwards (`STATE` in `up`/`reconnecting`) |
| `--json` | Machine-readable output |
**Example:**
```
porthole list
porthole list --running
```
---
### 5.7 `porthole remove <name>`
Aliases: `rm`, `delete`
Deletes a saved profile. Closes it first if running.
| Flag | Description |
|------------------|--------------------------------------------------------------------|
| `--keep-running` | Delete the profile but leave an active instance running untracked |
An instance left running via `--keep-running` is no longer visible to
`list`/`status` (its profile is gone), but is still caught by `wipe`
(§5.8), which matches by supervisor process signature rather than tracked
state — same as vmic's `wipe`.
---
### 5.8 `porthole wipe`
Alias: `reset`
Closes and deletes **every** forward, including any supervisor/`ssh`
processes matching porthole's signature that aren't in the current profile
store (e.g. orphaned after a crash). Confirmation prompt unless `--yes` —
unlike vmic's `wipe` (no prompt), porthole's tears down active network
tunnels rather than just audio routing, so the extra confirmation is a
deliberate, not accidental, difference.
| Flag | Description |
|-------------|----------------------------|
| `-y, --yes` | Skip confirmation prompt |
---
### 5.9 `porthole completions <shell>`
Generates a shell completion script. `<shell>` ∈ `bash`, `zsh`, `fish`.
---
## 6. Global options
| Flag | Description |
|-----------------|----------------|
| `-h, --help` | Print help |
| `-V, --version` | Print version |
---
## 7. Exit codes
| Code | Meaning |
|------|--------------------------------------------------------------------|
| `0` | Success |
| `1` | Error — see the printed message |
| `2` | CLI usage error (bad/missing arguments — clap's own exit code) |
Kept deliberately flat, matching vmic: granular per-failure codes (not
found vs. already-exists vs. bind conflict, etc.) only pay for themselves
once something is actually scripting against them, and nobody's asked for
that yet. Every error still gets a specific, greppable message on stderr.
---
## 8. Surviving reboots
**Explicitly out of scope for v0.1**, and worth calling out since it's
adjacent to the overview's own motivating problem: the supervisor is a
plain process, not a system service, so a full reboot or logout kills it
along with everything else — reconnect (§4) covers network blips and
sleep/wake, not "the machine came back up."
The only piece porthole commits to now is `open --all` (§5.2), which exists
specifically so an external mechanism can drive it — a systemd `--user`
unit, a login item, a cron `@reboot` line. porthole does not register,
manage, or template any of those itself; that's a v0.2+ decision (§9) once
it's clear which one people actually want.
---
## 9. Open questions for v0.2
- Autostart: if/when to commit to a specific mechanism (systemd user unit
template? login item?) now that `open --all` exists as the hook.
- Multi-hop `--via`: confirmed as a straight passthrough to `ssh -J`
(§3.1/§5.1) for v0.1; revisit only if per-hop supervised status ever
becomes a real ask.
- Templating (`dbtun`-style): saved "kind" templates (e.g. `--template
postgres` implies port 5432) — worth adding as sugar over `add`, or
scope creep?
- Config export/import for moving profiles between machines.
View Git Blame Copy Permalink