tractor/ai/tpt-backends/03_wg_tunnel_bindspace.md

# Plan 03 — WireGuard (and other tunnels) as a *nested bindspace* via `pyroute2`

Tracks gh [#482] + the tunnelled-maddr item of [#443].
Prereq reading:
[`00_shared_backend_contract.md`](./00_shared_backend_contract.md).

**Thesis**: WireGuard is **not** a `MsgTransport`. It is an
interface-layer tunnel that is transparent to `socket(2)`, so
the correct abstraction is a *bindspace* — a scoped,
`@acm`-managed network context that an existing L4 transport
(`tcp`, and later `quic`/`tipc`-over-UDP-bearer) binds *inside*.
This plan implements `Address.namespace` (spec'd but unused
since day one) and the composed/tunnelled maddr grammar, with
`pyroute2` as the netlink codec and as much of the I/O moved
onto `trio` as the library's sans-io layer allows.

[#482]: https://github.com/goodboy/tractor/issues/482
[#443]: https://github.com/goodboy/tractor/issues/443

---

## 1. What exists today (verified, per #482)

- `wrap_address()` accepts maddr `str`s (leading-`/` dispatch,
  `_addr.py:262`). `parse_maddr()` and `mk_maddr()` support plain
  TCP/UDS addresses plus nested, canonical bearer-first `/wg/`
  stacks represented locally as `TunnelledAddress` wrappers.
- there is no `wg` proto in the multiaddr *spec* yet, but
  multiformats/py-multiaddr#108 (key form `u<base64url>`) is
  **merged** as of 2026-07-28 (`f86519da`) — and unreleased, the
  latest `0.2.0` predating it. Spec registration is still tracked
  by multiformats/py-multiaddr#107 and gh #483.
- **today's deployable story remains declarative**: run `wg-quick`
  out-of-band, parse the maddr, strip its wrapper to the overlay
  `(host, port)`, verify the pubkey against the live tunnel,
  hand the overlay addr to `registry_addrs=`/`tpt_bind_addrs=`.
  #482 already contains working example code for exactly this.
- `Address.namespace` exists in the Protocol
  (`_addr.py:94-101`, "the if-available OS-specific network
  namespace key"). `TunnelledAddress` implements it from its spec;
  no concrete transport backend implements it yet.

## 2. Three layers, three PRs

| layer | what | dep | ships |
| --- | --- | --- | --- |
| **A. declarative** | commit #482's examples; `parse_maddr()` learns `/wg/u<key>` → overlay `Address` + verified pubkey | `multiaddr` (already), `wg(8)` CLI | first |
| **B. `pyroute2` read/verify** | replace the `subprocess.run(['sudo','wg','show'])` shelling with netlink queries | `pyroute2` extra | second |
| **C. `@acm` lifecycle** | create/configure/tear down wg ifaces + netns *from the runtime*, as nested bindspaces; implement `Address.namespace` | `pyroute2` + `CAP_NET_ADMIN` | third |

Each is independently valuable and independently reviewable.
**Do not attempt C first** — the interesting design (nested
bindspace `@acm`s) is only well-posed once A has pinned the
address grammar and B has proven the netlink path under trio.

---

## 3. Layer A — declarative `wg` maddrs

### 3.1 the address shape

The decision: **a wg segment annotates an existing address, it
does not create a new address type.** Two candidate encodings;
**pick (a)**:

- **(a) `TunnelledAddress` wrapper** (recommended):
  ```python
  class TunnelledAddress(
      msgspec.Struct,
      frozen=True,
  ):
      overlay: Address         # e.g. TCPAddress
      tunnel: WGTunnelSpec     # proto-specific, frozen
  ```
  with `.proto_key` **delegating to `overlay.proto_key`** so every
  existing table lookup (`_addr_to_transport`,
  `enable_transports` guard at `_root.py:391`,
  `transport_from_addr()`) keeps working untouched, and
  `.unwrap()` delegating to `overlay.unwrap()` so **nothing new
  crosses the wire**. `.namespace` and `.bindspace` come from
  the tunnel spec. The wrapper is stripped (`→ .overlay`) at the
  moment of bind/connect.
  - ⚠️ `is_wrapped_addr()` (`_addr.py:194`) tests
    `type(addr) in _address_types.values()` — a `bidict` of
    proto_key→type. `TunnelledAddress` isn't in it and must not
    be (it's not 1:1 with a proto). So either add an explicit
    `isinstance(addr, TunnelledAddress)` clause there, or give
    the wrapper a marker and test structurally. Do the former;
    it's two lines and honest.
  - the reflection in `Endpoint.start_listener()`
    (`inspect.getmodule(self.addr)`) would resolve to the
    *wrapper's* module, not the transport's. **So the wrapper
    must be unwrapped before it reaches `Endpoint`** — i.e. by
    the bindspace `@acm` (layer C) or by `parse_maddr()`
    (layer A). State this loudly in the docstring; it's the #1
    way to get this wrong.
- (b) add fields to each existing `Address` type. Rejected:
  duplicates tunnel logic per-backend and pollutes `.unwrap()`.

```python
class WGTunnelSpec(
    msgspec.Struct,
    frozen=True,
):
    peer_pubkey: str          # std-base64 `wg(8)` form
    iface: str = 'wg0'
    netns: str|None = None
    # layer-C-only fields, unset in layer A
    maybe_endpoint: tuple[str, int]|None = None
    maybe_allowed_ips: tuple[str, ...] = ()
```

### 3.2 `parse_maddr()`/`mk_maddr()`

Grammar — **verified** against py-multiaddr#108, first on the
`baudco/py-multiaddr@wg_support` branch and re-verified after it
merged upstream (`multiformats/py-multiaddr@f86519da`); all three
forms below parse *and* round-trip. Note the codec also validates
that the key decodes to exactly 32 bytes, so a truncated key is a
`StringParseError`, not a silently-mangled parse:

```
/ip4/192.168.1.50/udp/51820/wg/u<A_pub>/ip4/10.0.11.1/tcp/1616
\_______ bearer __________/\__ key __/\______ overlay ______/
 underlay, wg `ListenPort`             the `MsgTransport` bind
```

The `/wg/` segment is **infix, not suffix** — the segments
*before* it are the wg **bearer** (the underlay `(ip, udp-port)`
that `wg(8)` itself listens on, per the codec docstring's own
`/ip4/1.2.3.4/udp/51820/wg/{key}` example), and the segments
*after* are the **overlay** endpoint that `tractor` binds.

⚠️ **CORRECTION** — an earlier revision of this plan (and the
examples in gh #482) used a *suffix* form
`/ip4/10.0.11.1/tcp/1616/wg/u<key>`. That parses, but it is
semantically inverted: it puts the overlay addr where the bearer
belongs, `tcp` where wg's `udp` `ListenPort` goes, and declares
no overlay endpoint at all. `parse_wg_maddr()` in
`examples/multihost/wg_lan/` now rejects it with an actionable
error.
Observed protocol-name lists, for writing the `match`:

| maddr | `[p.name for p in m.protocols()]` |
| --- | --- |
| `/ip4/1.2.3.4/udp/51820/wg/u<k>` | `['ip4','udp','wg']` |
| `/ip4/../udp/../wg/u<k>/ip4/../tcp/..` | `['ip4','udp','wg','ip4','tcp']` |

- so the three parts have **three different owners**, and only the
  third is an `Endpoint`:

  | part | socket owner / provisioner | runtime role |
  | --- | --- | --- |
  | bearer | kernel-owned; externally provisioned in layer A, tractor bindspace-provisioned in layer C | control-plane metadata, never an `Endpoint` |
  | `/wg/u<key>` | nothing — it's an identity | parsed and explicitly verified |
  | overlay | `tractor`'s `IPCServer` | application `MsgTransport`, as `.overlay` |

  This owner-split is the real axis of the design, *not* whether
  the maddr stack is "composed" (it is).
- ⚠️ **CORRECTION**, an earlier draft of this section specced a
  hand-rolled `_peel_tunnel_segs(proto_names) -> (bearer_names,
  tunnel_specs, overlay_names)`. **Do not write it.**
  `py-multiaddr` already ships the whole tunnel compose/peel API
  and it was simply missed here — see its README "En/decapsulate"
  and "Tunneling" sections, and gh #443's 2nd bullet which links
  them. Verified against the pinned rev:

  | need | API |
  | --- | --- |
  | isolate the bearer | `ma.decapsulate_code(P_WG)` |
  | drop the overlay, keep bearer+key | `ma.decapsulate(overlay_ma)` |
  | per-seg maddrs | `ma.split()` |
  | rejoin a seg tail | `Multiaddr.join(*segs)` |
  | read the key | `ma.value_for_protocol('wg')` |
  | recompose | `bearer.encapsulate(key).encapsulate(overlay)` |

  `.decapsulate_code()` handles the infix `/wg/` seg cleanly
  *because* it cuts on proto-code and never tries to match an
  addr value — the key seg has no addr of its own. This is the
  same NIH trap gh #429 existed to close, one layer up.

- ⚠️ `value_for_protocol('ip4')` on a *full* tunnelled maddr
  silently returns the **first** match, i.e. the bearer's host.
  Always call it on a peeled sub-maddr, never the whole stack.

- `parse_maddr()` gains a case on
  `[('ip4'|'ip6'), 'udp', 'wg', ('ip4'|'ip6'), <overlay-l4>]` →
  peel w/ the API above, decode the multibase key to std-base64,
  and return `TunnelledAddress(overlay=..., tunnel=WGTunnelSpec(
  ...))` w/ the bearer recorded in the spec.
- keep the existing 2-proto cases byte-identical; add the new
  case *after* them.
- nesting (wg-in-wg) falls out of `.decapsulate_code()` cutting
  at the *last* occurrence — peel repeatedly rather than
  recursing through a bespoke splitter.
- `mk_maddr()` inverse for `TunnelledAddress` is just
  `.encapsulate()` composition; don't rebuild `str`s by hand.
- **pending an upstream release**: py-multiaddr#108 is merged, so
  `Multiaddr('/…/wg/u…')` parses off a PEP 621 direct-revision pin,
  since no release carries the codec. Gate parser entry on
  `_wg_proto_code()`, implemented as
  `protocols.protocol_with_name('wg')` under
  `except ProtocolNotFoundError`. Do **not** probe by parsing a
  dummy like `Multiaddr('/wg/uAAAA')` — the codec enforces a
  32-byte key, so that raises even when the proto *is* known. Do
  **not** hand-roll a `wg` parser in `tractor` — the whole point
  of #429 was dropping the NIH parser.

### 3.3 pure codecs + explicit verification

Port #482 §2's pure helpers into
`tractor/discovery/_tunnel.py`, keeping the impure probe cleanly
separated until layer B:

```python
def parse_wg_maddr(maddr: str) -> TunnelledAddress: ...   # pure
def wg8_pubkey(multibase_key: str) -> str: ...            # pure
async def verify_wg_peer(spec: WGTunnelSpec) -> bool: ... # layer B
```

Layer A's example-local `verify_wg_peer()` may shell out (`wg show
<if> peers`), but layer B replaces that probe with one explicit async
function backed by pyroute2. Never call it implicitly from
`wrap_address()`/`parse_maddr()` — parsing must stay pure and
side-effect-free; verification is the *caller's* explicit step
(and later, the bindspace `@acm`'s).

### 3.4 deliverables

- `examples/` scripts distilled from #482 §§3-5 (this is the
  unchecked "commit examples from ^" bullet in #443). They live
  under `examples/multihost/` — `test_docs_examples.py` walks
  `examples/` recursively and runs every collected file as a
  subproc asserting `rc == 0` (it doesn't even filter by
  extension, so a stray `README.md` would be `python`-run too),
  and `'multihost' not in p[0]` is already in its exclusion
  list. Anything needing a real second host or a live tunnel
  belongs there.
- a `docs/` page: tunnel setup, the maddr form, the two-host
  run. Keep prose in the docs; keep the examples runnable and
  minimal.
- tests: maddr round-trip, `TunnelledAddress` delegation
  (`proto_key`/`unwrap` identical to overlay), `wrap_address()`
  regression (a tunnelled maddr `str` → `TunnelledAddress`; a
  plain one → unchanged), and **a real end-to-end over a
  locally-created wg pair** gated on `CAP_NET_ADMIN` (see §5.4).

---

## 4. Layer B — `pyroute2` under `trio`

### 4.1 the library situation (verify at implementation time)

`pyroute2` ≥0.9 rewrote its core onto **asyncio**
(`AsyncIPRoute`; the sync `IPRoute` wraps it with its own loop).
It also ships a `WireGuard` netlink (generic-netlink) module
supporting `.set(iface, private_key=..., peer={...})` and
`.info(iface)`, plus `pyroute2.netns` / `NetNS` for namespaces,
and `IPRoute.link('add', kind='wireguard', ifname=...)`.

Three integration options, in increasing trio-nativeness:

- **(1) `trio.to_thread.run_sync()` around the sync API.**
  Netlink ops here are one-shot, sub-millisecond, and happen at
  bind/teardown time only — *not* in the msg hot path. This is
  the **correct default**: it's ~10 lines, uses a battle-tested
  API, and costs nothing where it's used.
- **(2) sans-io: `trio.socket` + pyroute2's message codecs.**
  `pyroute2`'s message classes
  (`pyroute2.netlink.rtnl.*`, `pyroute2.netlink.generic.wireguard.wgmsg`)
  encode/decode independently of its I/O core. So a
  `tractor/ipc/_netlink.py` with a small trio `NetlinkSocket`
  (`trio.socket.socket(AF_NETLINK, SOCK_RAW|SOCK_DGRAM, proto)`,
  `sendto`/`recv`, seq/pid matching, `NLMSG_DONE`/`NLMSG_ERROR`
  handling) + pyroute2 codecs is very achievable and is the
  honest reading of "as much trio wrapping as possible where any
  other async support can be replaced".
  **Do this for the paths we actually need** (link add/del,
  addr add, wg get/set, netns bind) and *only* those — a
  general netlink client is out of scope.
- (3) reimplement the codecs. Never.

**Recommended split**: ship (1) first so layer B is a small,
reviewable, behaviour-preserving swap of `verify_wg_peer()`'s
body; then land (2) as a follow-up commit for the read path
(`wg get`, `link get`) where the sans-io surface is smallest,
and keep (1) for the privileged mutating ops. Measure before
converting anything else — there is no perf argument here, only
a "no foreign event loop in a trio actor" argument, which (1)
already satisfies (a thread is not an event loop).

Explicitly **do not** pull in `trio-asyncio` for pyroute2 or infect
every wg-using actor merely to service one-shot netlink calls. A
dedicated `wgman` actor (§5.1) is the one plausible asyncio-hosted
shape: it can use tractor's own `.to_asyncio` task linkage while
keeping the foreign loop and provisioning authority out of ordinary
actor processes.

### 4.2 API shape

Pure-ish, functional, `@acm` for anything with teardown:

```python
async def read_wg_peers(
    iface: str = 'wg0',
    netns: str|None = None,
) -> tuple[str, ...]: ...          # base64 pubkeys

async def read_wg_pubkey(iface: str = 'wg0', ...) -> str: ...
```

and `verify_wg_peer()` becomes a thin composition over one shared key
snapshot.
Note the pure-getter rule: no `read_wg_peers(..., create=True)`.

---

## 5. Layer C — nested bindspace `@acm`s + `Address.namespace`

This is the part #443 and `multiaddr_declare_eps.md` actually
ask for: *"for any tunneled maddr-`str`-entry we deliver a
data-structure which can easily be passed to nested `@acm`s
which consecutively setup nested net bindspaces for binding the
endpoint addrs"*.

Layer C is where tractor takes ownership of bindspace orchestration.
For a fully bootstrapped deployment it may create the netns and wg
iface, configure peers/routes, and ask the kernel to establish the
bearer's UDP `ListenPort` through netlink/`pyroute2`. "Kernel-owned"
describes the data-plane socket, not who provisions it: tractor owns
the lifecycle while `Endpoint`/`MsgTransport` remain responsible only
for the overlay application socket.

### 5.1 candidate default: first-child `wgman`

For a WG-enabled deployment profile, consider eagerly spawning one
private **WireGuard manager** (`wgman`) as the root actor's logical
first child. It is a narrow network-control-plane service, not a
general worker and not an application-visible transport endpoint.
The naive profile gets one manager for the actor tree; advanced
deployments may disable it for pre-provisioned networking or place
one manager in each capability/bindspace security domain.

"First child" describes supervision and teardown ordering, not a
serial startup barrier. Submit the `wgman` spawn in the same startup
wave as ordinary children, start its pyroute2 import, generic-netlink
discovery and declared-tunnel reconciliation immediately, and publish
a readiness signal separately. Sibling processes can boot in parallel;
only their first WG-dependent bind/dial waits for manager readiness.
This overlaps setup with actor-tree startup and avoids every sibling
paying its own pyroute2/loop/socket initialization latency.

The initial manager can still call the sync helpers from §4.1. A
natural follow-up is to spawn it with `infect_asyncio=True` and keep
`AsyncWireGuard` clients alive on asyncio's host loop through
`tractor.to_asyncio.run_task()`. Tractor then owns cross-loop task
linkage, cancellation and error propagation, while normal siblings
remain plain Trio actors. Keep one client per realized namespace or
other kernel control domain; do not share a pyroute2 socket across
domains merely to reduce object count.

Keep the authority surface deliberately small:

- accept structured inspect/verify/ensure/release requests derived
  from `WGTunnelSpec`, `BindspaceSpec` and explicit `role`; never
  expose arbitrary pyroute2 calls, shell commands or `setns()` RPC;
- let the root/supervisor mediate access initially, or hand siblings
  a scoped manager capability; do not register a privileged `wgman`
  endpoint for unrestricted cluster-wide discovery;
- never return private keys or namespace FDs to application actors;
  pass secrets and live capabilities into the manager through the
  supervisor-owned bootstrap path;
- grant only the capabilities required for the manager's assigned
  domain. Prefer a manager already placed in that user/net namespace
  over one process holding ambient authority across every namespace;
- make ensure/release idempotent and reference-count ownership so one
  sibling cannot tear down a tunnel still borrowed by another.

The root owns the manager's lifetime. `wgman` must outlive all
siblings borrowing its tunnels and exit before the root drops the
underlying namespace/capability handles. A manager crash fails closed:
dependent operations receive an explicit service error; restart, if
enabled, reconciles declared state idempotently before advertising
readiness again. Do not silently let siblings fall back to privileged
local provisioning, since that defeats both the security boundary and
the single warm control-plane benefit.

Treat eager `wgman` as a measured deployment-profile choice. Compare
root startup with no WG declarations, pre-provisioned read-only WG,
and runtime-managed tunnels before making it unconditional whenever
the `wg` extra is installed. The intended invariant is "one warm
manager per simple WG actor tree", not "every tractor program spawns
a privileged child".

### 5.2 the composition

The maddr describes the composed network path and can be used as
either a source/listen or destination/dial handle. It does **not**
select the local instance of that network stack. A netns, VRF,
interface, user namespace, or equivalent platform resource is
orthogonal augmentation carried alongside/below the maddr.

Keep two bindspace representations with deliberately different
lifetimes:

```python
class BindspaceSpec(msgspec.Struct, frozen=True):
    '''Serializable spawn/config declaration.'''
    kind: str                 # `netns`, later `vrf`, ...
    key: str|None             # requested name/key, if any
    lifecycle: Literal['attach', 'open']


class BindspaceIdentity(msgspec.Struct, frozen=True):
    '''Stable identity of the realized platform resource.'''
    kind: str
    key: str|None             # mutable name, absent after unlink
    inode: int                # stable Linux namespace identity


class BindspaceHandle(ProcessLocal):
    '''Scoped, non-serializable capability for one live bindspace.'''
    spec: BindspaceSpec
    identity: BindspaceIdentity
    namespace_fd: int|None
    ownership: Literal['owned', 'borrowed']


@acm
async def open_bindspace(
    spec: BindspaceSpec,
) -> AsyncGenerator[BindspaceHandle, None]:
    '''
    Provision/borrow one bindspace and yield its live capability.

    '''
```

The initial model limits `BindspaceKind` to `netns` while preserving
the required lifetime split. `BindspaceSpec` and
`BindspaceIdentity` are frozen msgspec structs which cross
config/spawn serialization. `BindspaceHandle` also uses msgspec's
generic struct storage by inheriting the global
`tractor.msg.ProcessLocal` marker. Its hidden unsupported sentinel
blocks direct and nested default msgspec encoding without a recursive
IPC hot-path scan. The handle validates any supplied FD against
`BindspaceIdentity.inode`; explicit FD transfer belongs to the
supervisor bootstrap path. An FD avoids name-resolution TOCTOU,
survives rename/unlink, and identifies the exact namespace the parent
provisioned. Extend the kind/field union only when a second platform
resource is implemented.

`BindspaceSpec.lifecycle` is explicit serialized policy:
`'attach'` borrows an existing resource and `'open'` creates/owns one.
`open_bindspace()` dispatches that policy by bindspace kind. Never
infer it from a listen/dial role: either role may use pre-provisioned
or locally owned networking.

The first lifecycle implementation is deliberately borrow-only:
`attach_netns()` opens either `/proc/self/ns/net` when
`BindspaceSpec.key = CURRENT_NETNS`, or a named entry beneath
`/var/run/netns`. It derives identity from the opened FD, yields
`ownership='borrowed'`, and closes only that FD on exit. "Attach" does
not call `setns()`; it never creates, enters or removes a namespace.
Future `open_netns()` creation and owned teardown remain a separate
privileged supervisor change.

`open_netns()` is that owned counterpart: it requires a named spec,
creates through pyroute2 in a shielded worker call, attaches the live
FD, and yields `ownership='owned'`. FD closure precedes another
shielded pyroute2 removal call on every post-creation exit, including
cancellation. It still never calls `setns()`; process entry remains a
spawn/bootstrap operation.

`open_bindspace()` is **not** an address factory and does not return a
`TunnelledAddress`. At the declaration layer, listener allocation can
use the handle to replace an overlay while preserving every tunnel:

```python
async with open_bindspace(
    bindspace_spec,
) as bindspace:
    listen_decl = declared_addr.get_random(
        bindspace=bindspace,
    )
    transport_addr = strip_tunnels(listen_decl)
```

That sketch intentionally leaves the `.get_random()`/bindspace value
contract open. A concrete transport call returns a concrete overlay;
a declaration-level call may replace the overlay and return a new
`TunnelledAddress`. In either case wrappers remain until the final
transport bind/dial boundary, where `strip_tunnels()` is mandatory.
At the listener boundary, keep the split explicit:
`Endpoint.addr` is the peeled concrete address used for transport
reflection, while `Endpoint.declared_addr` retains the original
wrapper for namespace diagnostics and later bindspace orchestration.

Per-platform provisioning still composes one resource context per
tunnel/bindspace layer:

```python
@acm
async def open_netns(
    spec: BindspaceSpec,
) -> AsyncGenerator[BindspaceHandle, None]: ...

@acm
async def open_wg_iface(
    spec: WGTunnelSpec,
    config: WGInterfaceConfig,
    bindspace: BindspaceHandle,
    role: Literal['listen', 'dial'],
) -> AsyncGenerator[WGTunnelSpec, None]: ...
```

`WGInterfaceConfig` and each `WGPeerConfig` are process-local and
rejected by the global `ProcessLocal` wire guard. The interface config
owns its private key, local addresses and listen port; each peer owns
its public key, allowed CIDRs, optional endpoint, preshared key and
keepalive. Reprs redact private/preshared keys. `WGTunnelSpec` remains
serializable public maddr-derived identity/endpoint data. This split
supports multi-peer listeners without overloading the tunnel maddr.

and a driver that folds a list of specs into nested contexts
(`contextlib.AsyncExitStack` for the N-deep case). The
`parse_endpoints()` API (`_multiaddr.py:153`) is the front door:
it already returns
`dict[name, list[Address|TunnelledAddress]]` and the
`multiaddr_declare_eps.md` sketch anticipates the recursive
`dict[str, list[Address]]|dict[...]` return for tunnelled
entries. Extend it to carry the tunnel stack, not to *enter* it.

The caller supplies `role` to tunnel-resource contexts such as
`open_wg_iface()`; do not infer it from maddr shape. Bindspace
lifecycle remains the independent explicit policy above. The same
composed maddr can name a server source or client destination (§5.4).

### 5.3 `Address.namespace`, at last

- `TunnelledAddress.namespace` → `(kind, id)` e.g.
  `('netns', 'tractor-wg0')`.
- existing plain backends implement it explicitly as `None`, so the
  Protocol does not lie and tunnel delegation needs no `getattr()`
  fallback.
- `Endpoint.namespace` reads the retained declaration rather than its
  peeled transport addr; both `Endpoint.pformat()` and
  `Server.pformat()` expose that value as the cheapest proof the layer
  is wired.

Use `github/ns_aware@e4688cad` as prototype evidence, not code to
cherry-pick unchanged. Its `/proc/<pid>/ns/<type>` inode reader and
`ip netns identify` probe establish the useful `(key, inode)` identity
pair. Layer C should move that shape into `BindspaceIdentity`, avoid a
subprocess where netlink/procfs suffices, and hold the namespace FD in
`BindspaceHandle` to pin the identity.

### 5.4 the netns/process reality — read this before designing

**The headline consequence, stated up front**: netns is a
**runtime-level config API, not an actor-app-code API.** It is
declared as part of how an actor process is *brought up* — a
spawn-time/boot-time input alongside `enable_transports` and
`tpt_bind_addrs` — and it is **not** dynamically re-enterable by
app code once the actor is live. There is deliberately no
`await actor.enter_netns(...)`. Two hard reasons, both below:
`setns(2)` doesn't retroactively move existing sockets, and it's
per-thread rather than per-process. Anything that *looks* like a
mid-life API here would be a footgun that silently leaves the IPC
server bound in the old namespace.

- `setns(2)` with `CLONE_NEWNET` affects **the calling thread
  only**, and sockets already created keep their original netns.
  A trio actor is effectively single-threaded for our purposes,
  so "enter the netns, *then* bind" works — but any
  `to_thread` worker (§4.1 option 1!) is in the **original**
  netns unless it also `setns`. Concretely: a wg query issued
  via `trio.to_thread` will hit the wrong namespace. Either
  pass `netns=` down to `pyroute2` (which does the
  fork/setns dance itself) or pin a dedicated worker. **This is
  the single subtlest bug in this plan — write the test first.**
- entering a netns is *process-global-ish and irreversible-ish*
  in practice. Therefore: **netns membership belongs to the
  actor process, decided before the runtime binds**, not to a
  mid-life actor API. Design:
  - the root/parent decides the `BindspaceSpec`, provisions or
    borrows it, and passes the spec plus an inherited/transferred
    namespace-FD capability through the spawn backend (there's already
    `enable_transports`/`accept_addrs` plumbing at
    `_runtime.py:1595-1615` — the netns rides alongside).
  - the child spawn/bootstrap trampoline calls `setns()` **before**
    `_runtime.async_main()`, `IPCServer.listen_on()`, parent-channel
    connection, or creation of any worker thread/socket.
  - only after successful entry does the child drop namespace-entry
    privileges and initialize the actor runtime.
  - a root/single-actor process follows the same ordering: enter during
    root bootstrap, never after actor runtime startup.
  - iface/route/WG provisioning is genuinely scoped and remains under
    the parent/supervisor's `BindspaceHandle` context.
  - document the constraint rather than hiding it; a
    `RuntimeError` if namespace entry is attempted after bootstrap.
- capabilities: iface/netns creation/config needs `CAP_NET_ADMIN`;
  entering an existing Linux namespace normally requires
  `CAP_SYS_ADMIN` in the owning user namespace. Never `sudo` from
  inside the runtime. A privileged parent/helper should provision the
  stack and open the namespace FD; the child receives only the scoped
  capability and temporary authority needed to enter it, then drops
  that authority before actor code runs. This separates create/config
  authority from enter/use authority and fits user-namespace/capability
  deployments without granting every actor broad ambient caps.
  Two supported modes remain:
  (i) pre-provisioned out-of-band (layers A/B — the default,
  and what #482 documents), (ii) runtime-managed when the supervising
  process/helper holds the required caps. Probe exact required caps and
  *fail loudly with an actionable message* otherwise.
- role semantics are explicit:
  - `listen`: may create/own the local bindspace, iface, routes, WG
    peer/listener state, and random local overlay; lifetime normally
    extends through all listeners and the actor process.
  - `dial`: may borrow an actor-wide bindspace or ensure local routing
    and tunnel state reaches the remote stack; it does not own the
    remote maddr and may need no new local resource at all.
  - source/destination use is an operation property, never permanently
    encoded into the maddr or inferred from segment ordering.
- teardown follows capability ownership, not just address type:
  - owned listener bindspaces tear down after endpoints/channels and
    the actor process have exited;
  - borrowed dial/actor-wide bindspaces only release their handle;
  - nested resources exit inside-out, but shared resources remain until
    their owning supervisor drops the final capability.
- teardown must be idempotent and tolerant: an iface/netns
  already gone must not strand the rest of the teardown — the
  exact lesson `_uds.close_listener()`'s `FileNotFoundError`
  tolerance and `_serve_ipc_eps()`'s per-ep `try/except`
  encode. Mirror both.

### 5.5 tests for layer C

- unit: fold-N-tunnel-specs-into-nested-`@acm`s, with fakes; assert
  enter/exit ordering (outermost-last-out) via a trace list.
- integration, gated on `CAP_NET_ADMIN` (skip otherwise, and in
  CI run it in a `--cap-add NET_ADMIN` container job): create two
  netns + a wg pair entirely in-process, boot a `tractor` root in
  one and a subactor in the other, `find_actor()` across the
  tunnel. This is a *fantastic* test to have and is fully
  self-contained — no second host, no `sudo` in the test body.
- the `to_thread`-netns-mismatch regression from §5.4, written
  **first** (red), then the fix (green), per project convention.
- bootstrap ordering: assert the child reports the expected namespace
  inode before parent-channel connect and listener creation.
- FD capability: rename/unlink the namespace name after opening its FD
  and prove child entry still selects the pinned inode.
- privilege drop: prove actor code lacks provisioning caps after entry.
- role/ownership: fake listen/dial resources and assert owned listener
  teardown versus borrowed dial-handle release.
- `wgman` bootstrap: prove sibling process startup overlaps manager
  reconciliation while the first WG operation still waits for its
  readiness signal.
- `wgman` authority: reject arbitrary callers/operations and prove an
  unprivileged sibling cannot receive secrets, FDs or provisioning
  authority through the manager API.
- `wgman` lifetime: prove it outlives tunnel borrowers, fails pending
  requests explicitly on crash and reconciles before restart-ready.

---

## 6. "Other shuttle-able tpts"

The generalization the #482 follow-up gestures at: once
`TunnelledAddress` + `open_bindspace()` exist, the same
machinery covers any iface-layer tunnel `pyroute2` can drive —
`ipip`/`gre`/`sit`/`vxlan`/`geneve`/`bridge`/`veth`. Keep
`WGTunnelSpec` as *one* frozen struct among a
`TunnelSpec = WGTunnelSpec|VxlanTunnelSpec|...` union with a
`kind: ClassVar[str]`, and dispatch `open_*` by `match` on it.
Design for it now (union + `match`), implement only `wg` +
`netns`. `veth`-pairs-in-netns is the natural second one because
it makes the §5.5 integration test possible without wg at all —
consider doing it *first* for exactly that reason.

## 7. Non-goals

- no wg userspace implementation, no key exchange, no
  `wg-quick` reimplementation (config-file parsing is
  out of scope; take structured input).
- no persistence of private keys beyond what layer C's iface
  creation needs (and that stays in `get_rt_dir()`, 0600).
- macOS/Windows: layers B/C are Linux-only. Layer A (declarative)
  works anywhere `wg` does. Gate accordingly and say so in the
  docs — do not silently no-op.

## 8. Risks

| risk | mitigation |
| --- | --- |
| `to_thread` worker runs in the wrong netns | §5.4; pass `netns=` to pyroute2 or pin a worker; test-first |
| namespace name is renamed/replaced between provision and spawn | pass an open namespace FD; verify `(key, inode)` after child entry |
| child starts sockets/threads before `setns()` | enter in the spawn bootstrap trampoline before `_runtime.async_main()`; assert inode ordering |
| ambient capabilities leak into actor app code | split provision/enter authority and drop caps before runtime initialization |
| dial path tears down a shared actor bindspace | encode ownership in `BindspaceHandle`; borrowed handles never remove resources |
| py-multiaddr#108 merged but unreleased | PEP 621 direct-revision pin + `_wg_proto_code()` gate; replace with a release floor once published |
| `TunnelledAddress` leaks into transport reflection/type dispatch | keep wrappers through declaration/bindspace handling, call `strip_tunnels()` at channel/endpoint boundaries, and retain the boundary regressions |
| privileged ops in a library | never `sudo`; explicit cap probe + actionable error; pre-provisioned is the default |
| pyroute2 0.9 asyncio core drags a loop into every actor | use a worker for one-shots; confine persistent asyncio to an infected `wgman` (§4.1, §5.1) |
| eager `wgman` serializes or slows root bootstrap | spawn it in parallel; gate only WG-dependent operations on readiness; measure before making the profile unconditional |
| `wgman` becomes a cluster-wide privilege oracle | keep it private/scoped, expose structured verbs only and split managers by capability domain |
| netns teardown strands actor teardown | idempotent/tolerant teardown mirroring `_uds.close_listener()` |

## 9. Follow-up issue seeds

- `veth`-in-netns bindspace (unblocks capless-ish integration
  testing, and is a great local multi-"host" test rig)
- composed/tunnelled maddr grammar shared with plan 02's
  `/…/quic-v1/…` stacks (gh #443)
- `wg` proto into the multiaddr **spec** (gh #483), then flip
  `MsgTransport.maddr` to always return `Multiaddr` (the third
  #443 bullet)
- first-child `wgman` prototype: concurrent bootstrap, scoped sibling
  access, infected-asyncio pyroute2 ownership and restart reconciliation
- runtime-managed wg key rotation / peer add-remove through `wgman` —
  the natural "actor that owns the network" demo