Per contract §0 ("if this doc disagrees with the code, the code
wins; fix it in the same PR"), fold the step-0 probe results and
the as-landed impl back into `01_tipc_backend.md`.
Settled the two claims §9 flagged as unverified,
- `SO_ACCEPTCONN` on `AF_TIPC` **works** (answers `1`); we never
needed trio's `except OSError` carve-out.
- dup-name bind → **silent crosstalk is real**: both binds
succeed and dials alternate strictly, so a `.get_random()`
clash is never `EADDRINUSE`.
Corrections where the plan was wrong,
- §5.2's `tipc_event` is **48B not 40B** (`4+4+4+8+28`), and
python exposes `TIPC_WAIT_FOREVER` as `-1` so it needs masking
before packing as `'I'`.
- §7.2's pytest mark goes in `_testing/pytest.py::
pytest_configure()`, NOT `pyproject.toml` — the repo has no
`markers` ini table.
- §7.4's "10k → 10k distinct" is a ~1.2% flaky assert by
birthday bound on a 32b instance space; use `>= n-2` w/ the
arithmetic documented.
- §2.2's `unwrapped_type` and §3.2's `from_addr()` sketch still
showed the 2-tuple + the `'tipc:<stype>:<scope>'` prefix hack
that §2.2 itself had already withdrawn.
Two hazards the plan never anticipated, now recorded in §9,
- an unpublished-name dial answers `EHOSTUNREACH` which python
maps to a **bare `OSError`**, NOT a `ConnectionError` subtype,
so the `_reraise_as_connerr()` wrap is contract-§4 mandatory.
- a connect-then-drop peer answers `ENOTCONN` from
`getpeername()`, which — since `.get_stream_addrs()` runs
BEFORE the handshake — used to kill the whole actor.
Also withdraw §9's "fold a 6-byte digest into `(stype_low,
instance)`" escalation: varying `_stype` per-actor would need
65536 topology subscriptions and kills layer B outright.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
`mk_maddr()`/`parse_maddr()` learn,
/tipc/<stype>/<instance>/<scope>
mirroring how `uds` maps onto the spec-legal `/unix`.
XXX `str`-ONLY for now: there is no registered `/tipc` proto
in the multiaddr table (upstream track gh #483 +
multiformats/py-multiaddr#107) and `Multiaddr()` rejects an
unregistered name outright. `MsgTransport.maddr`s return type
is already `Multiaddr|str` (and `MsgpackUDSStream` already
exercises the `str` branch), so this fits — but it IS why gh
`parse_maddr()` therefore special-cases the `/tipc/` prefix
BEFORE handing anything to `Multiaddr()`.
Also drive the maddr mapping-table tests off `_address_types`
instead of a hardcoded len/dict so the next backend can't
fail them for the wrong reason.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
Wire the backend through every registration site (contract §2)
so `--tpt-proto tipc` is a first-class suite mode,
- `_state.TransportProtocolKey` gains the key
- `_addr._address_types` + `._default_lo_addrs`
- `_addr.wrap_address()` gets a `case ('tipc', *_)`; being a
4-elem seq it can't collide w/ `tcp`s or `uds`s 2-tuple
cases, so NO ordering hazard (and a bare seq-pattern matches
the `list` form `msgpack` decodes to).
- `_types`: the `Address` union, `_msg_transports`,
`_key_to_transport`, `_addr_to_transport` and the
`transport_from_stream()` family match. That last one keys
off `._tipc.AF_TIPC` (which carries the uapi fallback) NOT
`socket.AF_TIPC` which is linux-only.
Test-harness side,
- `get_rando_addr()` gains a `tipc` branch; `.get_random()`
already salts w/ `uuid4`+pid so both within- and cross-proc
isolation come for free.
- the `tpt_protos` fixture calls an addr-type's optional
`.is_available()` and `pytest.fail()`s w/ its reason. Keeps
a module-less box from turning `--tpt-proto tipc` into a few
hundred confusing connect-timeouts. Generic on purpose —
plans 02/03 need the same hook.
- the discovery `daemon` fixture's readiness probe learns to
dial a TIPC service name (it previously assumed tcp-or-uds
and blew up on the 4-tuple).
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
`SpawnSpec.reg_addrs`/`.bind_addrs` pinned the wire shape to
a 2-tuple, so a `tipc` addr (`('tipc', stype, inst, scope)`)
died at the child w/ `msgspec.ValidationError: Expected array
of length 2, got 4` -> `invalid SpawnSpec IPC msg`.
Point those fields at `UnwrappedAddress` (which `SpawnSpec`s
own TODO already asked for) and widen the alias.
XXX VARIADIC (`tuple[str|int, ...]`) rather than a union of
the two concrete shapes, bc `msgspec` refuses a union holding
more than one array-like type.
?TODO, the real fix is the full proto-key migration (contract
§1.1) after which this becomes a tagged union keyed off elem
0 and per-proto validation comes back.
Note the alias is declared TWICE — `.msg.types` re-declares it
to dodge a circular import (`._addr` -> `.ipc._tcp` -> `.msg`)
and *that* copy is what actually validates the wire msg.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
Wire `.connect_to()` (dial by service name), `.connected()`
and `.get_stream_addrs()` on top of `MsgpackTransport` so
`trio.SocketStream` + the existing `<I`-prefix framing carry
`msgpack` msgs over TIPC unchanged.
XXX both ends of a connected TIPC sock answer `TIPC_ADDR_ID`
port-ids and a port-id carries NO service name, so,
- the *dialling* side re-asserts the name it actually dialled
over `._raddr` (same move as `MsgpackUDSStream`s peer-pid
re-assign),
- the *accepting* side keeps a `TIPC_NAME_UNKNOWN` sentinel
plus the observed `(node, ref)`. It doesn't need more — the
`Aid` from `._do_handshake()` already carries the peer's
logical identity.
Also normalize dial failures: TIPC answers an unpublished-name
lookup with `EHOSTUNREACH`, which python maps to a **bare**
`OSError` and NOT a `ConnectionError` subtype the way
`ECONNREFUSED` maps to `ConnectionRefusedError`. The
discovery-ping path needs the `ConnectionError` shape, so the
`_reraise_as_connerr()` wrap is load-bearing, not polish.
XXX ALSO tolerate a dead peer in `.get_stream_addrs()`!
Unlike tcp/uds — where the kernel keeps answering the peer
addr until *we* close — TIPC answers `ENOTCONN` once the peer
is gone. Since `MsgpackTransport.__init__()` calls
`.get_stream_addrs()` (via `Channel.from_stream()`) BEFORE the
handshake, an unguarded `OSError` there escapes
`handle_stream_from_peer()`s handshake tolerance (contract §4)
and tears down the WHOLE actor. Any connect-then-drop peer — a
port scan, a liveness probe, a cancelled dial — was a remote
actor-kill. A dead peer must cost us an addr, not the runtime.
Deats,
- `TIPC_IMPORTANCE` exposed as a `.connect_to()` kwarg — TIPC
can rank a conn's traffic under congestion, which no other
backend can do. Defaulted to the kernel default for now;
wiring the parent<->child chan to `HIGH` is a follow-up.
- `TIPC_DEST_DROPPABLE = 0` so undeliverable msgs surface as
errors instead of being silently dropped.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
First slice of the `AF_TIPC` tpt backend: the addr type, the
`is_tipc_available()` capability predicate and the
name-publishing listener. No `MsgTransport` yet.
An actor's TIPC addr is a *service name* `(stype, instance)`:
`.bind()`ing the singleton `TIPC_ADDR_NAMESEQ` range IS the
service registration (it shows up in `tipc nametable show`)
and a peer's `.connect()`-by-name IS the lookup — so the
kernel does discovery for us, no registrar hop.
Deats,
- `.unwrap()` is proto-keyed as `('tipc', stype, inst, scope)`
using the `multiaddr` proto spelling so `wrap_address()`
can't confuse it with `tcp`s or `uds`s 2-tuples.
- `.rebind_from_sockname = False` bc `getsockname()` answers
a port-id; `.from_addr()` raises on a bare `TIPC_ADDR_ID`
rather than fabricate an un-dialable addr.
- `.bindspace` is the TIPC *scope*, i.e. literally the set of
hosts a published name is reachable from. `ZONE` scope is
deprecated/aliased so fold it to `CLUSTER` on input.
- mod stays importable on non-linux (uapi-value fallbacks,
the `_uds.SO_PASSCRED` precedent) bc `._addr` builds its
registration tables at import time.
XXX a `.get_random()` clash does NOT raise `EADDRINUSE` —
TIPC accepts multiple publishers of one name and round-robins
connects between them (verified against a live kernel), so a
collision is *silent crosstalk*. Hence the `blake2b` digest
and its (birthday-bounded) collision test.
Also,
- a generic `.is_available() -> (ok, why_not)` classmethod;
deliberately spelled generically (NOT `is_tipc_*`) so the
sibling env-dependent backends — `quic`/`iroh` (gh #353)
and the `wg` netns bindspace (gh #482) — get the same gate
for free. Its consumer lands w/ the reg tables.
- register a `tipc` pytest mark; the kernel-touching cases
self-skip unless `sudo modprobe tipc` has been run.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
Gate `Endpoint.start_listener()`s `getsockname()`-vs-`.addr`
reconciliation on a new per-addr-type `ClassVar[bool]`, set
`True` on both `TCPAddress` and `UDSAddress` so existing
behaviour is bit-for-bit unchanged.
That reconciliation exists ONLY to learn a kernel-assigned
port from a `port=0` tcp bind (its own comment says so). The
incoming `tipc` backend (gh #378) has no late-binding
analogue AND its `getsockname()` answers a `TIPC_ADDR_ID`
port-id rather than the name-seq it published — rebinding
from that would swap a dialable service name for an
un-dialable, un-reconstructable port id.
So opting out is semantically right rather than a hack.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
Guard test for `.start_listener()`s post-bind
`getsockname()`-vs-`.addr` round-trip, landed *before* that
reconciliation gets gated on an opt-out `ClassVar`.
- tcp: a `port=0` bind MUST still learn the kernel-picked
port, since the reconciliation is the only path that ever
does.
- uds: the sock-file path must survive the `.from_addr()`
round-trip unchanged.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
Just drop it — `pformat_boxed_tb()` spells its knobs
`tb_box_indent`/`tb_body_indent`, and that fn's default
(1-space box indent) is what the caller wanted anyway.
Regressed-by: 888af602 (`pformat_cs()` mv into `.devx.pformat`)
Found-via: `/run-tests` test_pformat_caller_frame_renders
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
`pformat_boxed_tb()` has never accepted an `indent` kwarg but
`pformat_caller_frame(box_tb=True)` has been passing one since
`888af602`. Nothing in the suite covered the branch, so the
`TypeError` only ever surfaced from `_mk_send_mte()` — i.e.
EVERY send-side `MsgTypeError` blew up while formatting itself
and masked the real msg-spec violation behind a bogus
`TypeError`.
Red on purpose per the test-first convention; the 1-line fix
lands next.
Also pin `pformat_boxed_tb()`s signature so a future typo'd
kwarg fails loudly at the call site instead of only when some
rare error path runs.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
it lands" framing in plan-03 and the example README was stale in
both directions: the branch pin is obsolete, yet you still can't
just `pip install multiaddr`.
Deats,
- §3.2's grammar table is now re-verified against the upstream
merge (`f86519da`) rather than only `baudco@wg_support` in a
throwaway venv. Also notes the codec enforces a 32-byte key,
so a truncated one is a `StringParseError` and not a silently
mangled parse.
- §1 says merged-but-unreleased; the still-open work is spec
registration (py-multiaddr#107 + gh #483).
- §3.4 swaps "pin the branch" for the `[tool.uv.sources]` `rev`
pin, and fixes the `_have_wg_maddr_proto()` recipe it
suggested — probing w/ `Multiaddr('/wg/uAAAA')` now ALWAYS
raises bc the codec wants 32B, i.e. that feature-detect would
report `False` even w/ the proto perfectly well known.
- risk table row goes "#108 not merged" -> "merged but
unreleased".
- example README: `uv sync` alone now suffices bc of the pin;
documents the 32B check and points at
`_have_wg_maddr_proto()` as the gate.
The one surviving `baudco` mention is deliberate, it records
where the grammar was *first* verified.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
`_segments()` called `Multiaddr(maddr)` purely to validate, then
swallowed every failure under `except Exception: pass`. That was
harmless pre-#108 — w/o a `wg` codec there was nothing to
validate — but now that the codec is pinned in, the swallow is
load-bearing and disabled: a malformed key sails past validation
into `wg8_pubkey()`, which happily emits a corrupt b64 str, and
the returned struct then fails its own `.maddr` round-trip. No
raise, just quietly wrong output.
Deats,
- add `_have_wg_maddr_proto()`, the gate plan-03 already
referenced but which never actually existed. Impl'd as
`protocols.protocol_with_name('wg')` under
`except ProtocolNotFoundError` and cached in a mod global,
same shape as the TIPC plan's `is_tipc_available()`.
- only validate when that gate is `True`, and let
`StringParseError` propagate — a maddr which doesn't parse
must NOT reach `wg8_pubkey()`.
- keep the degraded split for a pre-#108 install, now w/ an
explicit `XXX` naming the validation you give up.
So parsing stays pure but becomes total-or-raises. Our own
`ValueError`s (missing `/wg/` seg, bare tunnel w/o an overlay
ep) are unaffected, as is the `wg(8)` b64 round-trip.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
py-multiaddr#108 (the `/wg/u<key>` maddr proto) merged upstream
on 2026-07-28 as `f86519da`, but ships in no release yet — the
latest `0.2.0` predates it by ~4 months and carries no `wg`
codec at all. So `examples/multihost/wg_lan/` can't parse its
own maddrs off PyPI.
Pinned by `rev` and not `branch` so CI stays reproducible. Note
the lock now records the git source *instead of* the `>=0.2.0`
specifier, i.e. the dep floor above is fully overridden for as
long as this pin lives.
TODO, drop the pin (and bump that floor) the moment a release
carries the codec; the only consumer is the `wg_lan` example
set.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
One record covering all 9 commits on this branch, per the NLNet
generative-AI policy and the existing `ai/prompt-io/claude/`
convention.
Uses diff-ref mode for both the plan docs and the example code
(`git diff main..ng_tpts_planning -- <path>`) rather than
duplicating content already in `git log -p`. Kept verbatim in
the `.raw.md`: the four verified findings (trio's
family-agnostic `SocketStream`/`SocketListener`, the round-trip
table proving `/wg/` is infix, the proto-key `UnwrappedAddress`
rationale, and `setns(2)`'s per-thread reality), since those are
reasoning rather than diffable output.
`## Human edits` records that the steering here was substantial
and mid-session rather than post-hoc: two model claims about wg
maddr semantics were challenged and retracted (incl. in an
already-posted issue comment), and the proto-key +
netns-as-runtime-config framings were human-directed. Also notes
the one model-initiated correction — a pre-publication
self-review that downgraded the `uniffi`/asyncio thesis and the
TIPC duplicate-binder claim to explicitly-flagged assumptions.
Prompt-IO: ai/prompt-io/claude/20260813T001102Z_27c34aeb_prompt_io.md
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
`tests/test_docs_examples.py` walks `examples/` **recursively**
and subproc-runs every collected file asserting `rc == 0`. Ran
its exact filter against the tree: all 4 of our files were being
collected — including `README.md`, since the filter never checks
the extension, so CI would have literally tried `python
README.md`. These need a real second host + a live `wg` tunnel,
so they can't ever satisfy that gate.
`'multihost' not in p[0]` is already in the test's exclusion
list w/ no dir yet using it, so this is a pure `git mv` — zero
test changes — and it's what the exclusion was plainly there
for. Collection drops 24 -> 20 files, 0 of them ours.
Also records *why* in the two places someone would look before
adding the next one: a callout at the top of the example README
and a note on plan 03's §3.4 deliverables. Anything needing a
second host or live tunnel goes under `examples/multihost/`.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
Re-renders gh #482's examples w/ the corrected (infix) maddr
grammar, as the "layer A" slice of the wg plan: declarative
maddrs only, tunnel pre-provisioned out-of-band, zero runtime
changes.
- `wg_maddr.py`: a `frozen=True` `msgspec.Struct` addr carrying
`bearer`/`peer_pubkey`/`inner` (+ `inner_proto`), a `.maddr`
property that re-renders the canonical form, and pure
`mb_pubkey()`/`wg8_pubkey()`/`parse_wg_maddr()`. The parser
rejects #482's inverted suffix form w/ an actionable error and
stays **side-effect free** — `verify_wg_peer()` is a separate,
explicitly impure step the caller composes, never something a
parse path shells out to.
- `host_a_srv.py`/`host_b_client.py`: the two-host runs, passing
only `addr.inner` into `open_nursery()`/`open_root_actor()`,
which is the whole point — the bearer + key layers are already
established before any bind happens.
- `README.md`: the grammar + the 3-owners table, the `#108`
branch install line, tunnel setup, and a "what changed vs
#482" section enumerating the corrections.
Runnable-shaped but **not yet run against a live tunnel**; that's
next, and the reason these sit on the planning branch rather than
in `examples/` proper. `_segments()` marks its stopgap for when
the `wg` codec isn't installed.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
The prior revision (and gh #482's examples) had it as a suffix,
`/ip4/10.0.11.1/tcp/1616/wg/u<key>`. Wrong: verified against
`baudco/py-multiaddr@wg_support` (py-multiaddr#108) installed in
a throwaway venv, the canonical form is
/ip4/192.168.1.50/udp/51820/wg/u<A_pub>/ip4/10.0.11.1/tcp/1616
where segs *before* `/wg/` are the **bearer** — the underlay
`(ip, udp-port)` `wg(8)` itself listens on (`ListenPort`), per
the codec docstring's own example — and segs *after* are the
**overlay** ep, the only part we ever bind. The suffix form does
parse, which is why it slipped through, but it's semantically
inverted: overlay addr where the bearer belongs, `tcp` where
wg's `udp` goes, and no overlay ep declared at all.
Records the observed `[p.name for p in m.protocols()]` lists so
the `match` can be written against fact, and replaces the
"composed vs not" framing w/ what's actually the design axis:
three parts, three **owners** — bearer bound by the kernel via
`wg-quick`/`pyroute2`, `/wg/u<key>` bound by nothing (it's an
identity, verified out-of-band), overlay bound by our
`IPCServer` as `.inner`. `_peel_tunnel_segs()` correspondingly
grows a 3rd return, splitting *at* the tunnel seg so nested
tunnels fall out for free.
Also hoists the netns conclusion to the top of §5.3 where it
can't be missed: netns is a **runtime-level config API, not an
actor-app-code one**. It's a spawn/boot-time input alongside
`enable_transports`/`tpt_bind_addrs`, deliberately w/ no
`await actor.enter_netns(...)`, because `setns(2)` neither moves
already-created sockets nor applies beyond the calling thread —
so a mid-life API would silently leave the IPC server bound in
the old ns.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
Shape-matching in `wrap_address()` doesn't survive 4 backends and
the plans were papering over it: TIPC's natural unwrapped form is
a `(str, int)`, indistinguishable from `TCPAddress`, and iroh's
is a `(str, str)`, which the *existing* UDS case
(`case (_, filename) if type(filename) is str`) already swallows.
So the contract doc (§1.1) now carries the conclusion as a
**recommended prerequisite for all three backends**: make the
unwrapped form carry an explicit proto-key spelled with the
`multiaddr` protocol name — `('tcp', host, port)`,
`('unix', path)`, `('tipc', stype, inst, scope)`. `wrap_address()`
then collapses from an order-sensitive `match` to
`_address_types[addr[0]]` and the whole collision class stops
existing, while the on-wire form finally agrees w/
`mk_maddr()`/`parse_maddr()` instead of being an independent
invention.
Two consequences spelled out: it's a wire-format change
(`SpawnSpec`, `_root_mailbox`, `_registry_addrs`) + every fixture
+ downstream config, so it wants its own migration commit landed
*before* any new backend; and it's the moment to stop handing raw
tuples to users at all — `Address` becomes the public currency
and `UnwrappedAddress` an internal serialization detail, the same
discipline `ipaddress` uses (you pass `IPv4Address`, never a
4-tuple).
Plan 01 §2.2 is rewritten to match and to explicitly **retract**
its own earlier `('tipc:<stype>:<scope>', instance)` self-tagging
prefix hack — it keeps `wrap_address()` order-sensitive and does
nothing for the iroh/UDS collision, so the doc says don't
resurrect it. Registration checklist item 4 likewise becomes "do
the migration first, then this is a one-line `_address_types`
entry".
Also seeds a `/tipc` multiaddr-spec submission as a follow-up,
mirroring the `wg` track (multiformats/py-multiaddr#107/#108 + gh
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
Landing page for `ai/tpt-backends/`: points at the contract spec
as required first reading, tables the 3 plans against their
issues/deps/size, and states the landing order + why.
Deats,
- TIPC first as the cheap proof the table-registration story
generalizes to a genuinely new proto (stdlib-only, and
`trio`'s sock wrappers are family-agnostic).
- `wg` layer-A next since it's deployable-today doc/example work.
- QUIC last, gated on its own prep PR.
- notes that plans 01 and 02 both want the same
`Address.rebind_from_sockname` gate, so whichever lands first
ships it.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
Plan doc for gh #482 + the tunnelled-maddr item of #443. Pushes
back on the framing that `wg` is a tpt: it's transparent to
`socket(2)`, so it belongs as a *bindspace* — a scoped
`@acm`-managed net ctx that an existing L4 tpt binds *inside* —
and it's what finally implements the long-spec'd (never
implemented) `Address.namespace`.
Deats, 3 independently-shippable layers,
- A) declarative: commit #482's examples, teach `parse_maddr()`
the `/…/wg/u<key>` suffix -> a `TunnelledAddress` wrapper whose
`.proto_key`/`.unwrap()` delegate to `.inner` so nothing new
crosses the wire and every existing table lookup keeps working.
- B) swap the `subprocess.run(['sudo', 'wg', 'show'])` shelling
for `pyroute2`. Default to `trio.to_thread` around the sync API
(these are one-shot ops at bind/teardown, never hot-path), w/
sans-io codecs + a trio `AF_NETLINK` sock as the follow-up for
the read paths. Explicitly forbids dragging `trio-asyncio` in.
- C) `open_bindspace()`/`open_netns()`/`open_wg_iface()` `@acm`s
folded w/ an `AsyncExitStack`, + filling in the
`# !TODO, always be ns aware!` placeholder already sitting in
`Endpoint.pformat()`.
Also flags the subtlest bug in the whole thing: `setns(2)` is
*per-thread*, so a `pyroute2` query issued via `trio.to_thread`
lands in the *original* netns. Test-first, per usual.
Further, designs for the generalization (`TunnelSpec` union +
`match` dispatch) while only implementing `wg`+netns, and calls
out `veth`-in-netns as the better *first* one bc it makes a
fully self-contained two-"host" integration test possible w/o
`wg` at all.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
Plan doc for gh #353. Picks `iroh` (the `uniffi` FFI pkg) over
`aioquic`/`quiche` bc node-id addressing + hole-punching + relay
fallback is the whole point; `aioquic` stays documented as the
fallback since ~90% of the adapters here are reusable against a
sans-io core.
Deats,
- the layering: iroh `Endpoint` per actor, `Connection` per peer
(pooled via `trionics.maybe_open_context()`, not a hand-rolled
cache), one bi-stream per `Channel`. 4-byte prefix framing
stays so `MsgpackTransport` is untouched.
- `_uniffi_trio.py`: uniffi only uses `asyncio` as the executor
for its rust-future poll loop, so a ~40-line
`TrioToken.run_sync_soon()` bridge replaces it. Spells out the
real hazards — strong ref on the `ctypes` trampoline, poll-code
propagation, and a *bounded* shielded cancel-drain so a wedged
rust future can't make an actor un-cancellable.
- `IrohAddress` w/ ALPN as the `.bindspace`, the `(str, str)`
unwrapped form's collision w/ the UDS match-case, and why
`get_root()` needs a persisted secret key -> a lazy
`default_lo_addrs()` + a pure-getter/explicit-setter split.
- `QuicMsgStream(trio.abc.HalfCloseableStream)` +
`QuicListener(trio.abc.Listener)`, incl. the exact
EOF/reset/use-after-close semantics `_transport.py` already
match-cases on, and hanging the acceptor tasks off the
existing `Endpoint.listen_tn`.
- a prep-PR boundary: annotation widening, the shared
`rebind_from_sockname` gate and a `tpt_key`-based
`transport_from_stream()` dispatch, all landable w/ tcp/uds as
the only backends.
Further, notes this is our first tpt w/ real transport security
+ peer auth, so an inbound node-id allowlist hook belongs here —
and that it says nothing about the other backends.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
Plan doc for gh #378, the cheapest new backend we can add: it's
stdlib-only (CPython ships `AF_TIPC` + 23 `TIPC_*` consts) and
per the contract doc `trio`'s stream/listener wrappers don't care
about the addr family, so `MsgpackTransport` framing and
`trio.serve_listeners()` are reused verbatim.
Deats,
- `TIPCAddress` as a *service name* `(type, instance)` w/ scope
as the `.bindspace`; `bind()` publishes the singleton
name-range, peers `connect()` by name and the kernel resolves
+ load-balances. I.e. registration/lookup for free, no
registrar in the loop.
- the self-tagging `('tipc:<stype>:<scope>', instance)` unwrapped
form + why it must be match-ordered before `TCPAddress`'s.
- `get_random()` via a blake2b digest of the actor id (there's no
`port=0` analogue) and the silent-crosstalk risk that follows:
TIPC *allows* dup binders and round-robins, so a collision
doesn't `EADDRINUSE`, it cross-talks.
- an `Address.rebind_from_sockname` ClassVar to opt out of
`Endpoint.start_listener()`'s `getsockname()` reconcile, which
for TIPC always returns a port-id, never the bound name.
- the `TIPC_TOP_SRV` topology-service subscription as an `@acm`
yielding a chan of typed name-table events — push-based
register/dereg, the real "end game cluster proto" bit.
- commit sequencing, hard capability gating (`modprobe tipc`;
bare `AF_TIPC` is `EAFNOSUPPORT` on a stock box), CI matrix
notes, risks + follow-up seeds.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
First doc of a new `ai/tpt-backends/` set: the normative
description of what a `tractor` tpt backend *is* as of `main`,
written so the 3 sibling plans (TIPC, QUIC, `wg`) can be worked
independently (by another model/provider) w/o design drift.
Deats,
- the backend duck-type as empirically derived from
`_tcp.py`/`_uds.py`: the `Address` protocol surface, the
mod-level `start_listener()`/`close_listener()` pair and
`Msgpack<Proto>Stream(MsgpackTransport)`.
- the ONE reflection you can't break:
`Endpoint.start_listener()` resolves the tpt mod via
`inspect.getmodule(self.addr)`, so an `Address` type and its
listener fns MUST live in the same mod.
- a 10-item registration checklist (`_address_types`,
`_key_to_transport`, `_addr_to_transport`, `wrap_address()`
match-cases, `TransportProtocolKey`, maddr tables, ..) incl.
the import-time `_default_lo_addrs` trap.
- where the `trio.SocketListener` assumption is *actually*
load-bearing (just the `getsockname()` reconcile) vs. merely
annotated.
- the handshake/discovery invariants a new backend inherits,
dep policy (extras + import-laziness per the #470 boot-latency
budget), `--tpt-proto` harness plumbing and code style.
Also, records a verified finding the plans lean on hard:
`trio.SocketStream`/`SocketListener` are addr-*family* agnostic
— the only ctor checks are "is a trio sock" + `SOCK_STREAM` (+
an `OSError`-suppressed `SO_ACCEPTCONN`) — so any `SOCK_STREAM`
family CPython can make drops into the existing
`trio.serve_listeners()` path unmodified.
(this patch was generated in some part by `claude-code` using `claude-opus-5` (`anthropic`))
Replace the pre-start wall-clock sleep with an indefinite
checkpoint so cancellation ordering cannot race a timer in slow CI.
Record that `Cancelled` escapes each `start_or_cancel()` await
before the enclosing nursery or cancel scope handles it.
Review: PR #479 (goodboy)
https://github.com/goodboy/tractor/pull/479
(this patch was generated in some part by `opencode` using
`gpt-5.6-sol` (`openai`))
Compare the complete canonical `Nursery.start()` protocol error
before re-surfacing ambient cancellation. Preserve child-owned
`RuntimeError` objects whose messages only resemble Trio's wording.
Cover the colliding prefix and assert the original error remains the
exception group's sole leaf.
Review: PR #479 (goodboy)
https://github.com/goodboy/tractor/pull/479
(this patch was generated in some part by `opencode` using
`gpt-5.6-sol` (`openai`))
Resolve#474 with a new `tests/trionics/test_taskc.py` (9
tests) covering the `trio.Nursery.start()` wrapper landed in
PR #464, incl. the `modden.runtime.progman.open_wks()` use
case dug out as a minimal repro.
Deats,
- the lossy `RuntimeError('child exited without calling
task_status.started()')` only fires when the child absorbs
its ambient cancel pre-`.started()` (graceful-teardown
pattern); a well-behaved child surfaces `Cancelled` direct
from `.start()` on `trio` 0.29 - verified empirically 1st.
- `test_sibling_err_not_masked_by_startup_rte`: the `modden`
case; ONLY the root-cause sibling `ValueError` escapes the
nursery with the wrapper vs. bare-`.start()`'s lossy
riding-along startup-RTE noise.
- `test_pure_oob_cancel_not_morphed_to_rte`: plain ancestor
`cs.cancel()` exits clean vs. bare's eg-wrapped RTE.
- `test_genuine_startup_rte_still_raised`: sans cancellation
the protocol-bug RTE re-raises same as bare.
- `test_childs_own_rte_never_demoted_to_cancel`: exact-msg +
`isinstance`-guard regression cover; a child's own
`RuntimeError('never got started!')`/`RuntimeError(1234)`
never demotes to a `Cancelled`.
- `test_started_value_and_args_passthru`: positional args,
`name=` and `.started()`-value forwarding.
Also,
- `use_start_or_cancel=False` params pin upstream `trio`'s
current lossy behaviour as wart-documentation: a break on
a `trio` upgrade likely means new upstream porcelain and
the wrapper deserves a re-audit.
- verified 0 flakes over 50 hammer runs + 2 impl mutations
each caught by exactly the targeted tests.
Prompt-IO: ai/prompt-io/claude/20260702T161624Z_65bf9df5_prompt_io.md
(this patch was generated in some part by [`claude-code`][claude-code-gh])
[claude-code-gh]: https://github.com/anthropics/claude-code
Skip raw packet, decoded message, peer, and channel formatting when
transport logging is disabled. Keep message processing and wire
reads outside the guards so logging controls never affect IPC flow.
Also narrow the remaining pretty-struct TODO to require a
non-raising formatter with native-repr fallback.
(this patch was generated in some part by `opencode` using
`gpt-5.6-sol` (`openai`))
Remove the stale `Channel.from_addr()` design note now that its
`at_least_level()` guard avoids inactive pretty-rendering work.
(this patch was generated in some part by `opencode` using
`gpt-5.6-sol` (`openai`))
Use `Logger.isEnabledFor()` in `at_least_level()` so logger-local
and global disable controls short-circuit payload rendering.
Add `Channel.send()` coverage for effective-level, per-logger, and
global suppression while ensuring transport remains unchanged.
Caught-during: review remediation
Found-via: `/run-tests` test_log_guard_skips_payload_formatting
Review: PR #458 (goodboy)
https://github.com/goodboy/tractor/pull/458#issuecomment-5258207470
(this patch was generated in some part by `opencode` using
`gpt-5.6-sol` (`openai`))
Wrap `log.transport()` in `Channel.send()` and `log.runtime()` in
`PldRx.decode_pld()` with `log.at_least_level()` checks so that
expensive `pformat(payload)` / `repr(msg)` / `repr(pld)` calls are
skipped entirely when the respective log level is not active.
Previously the f-string arguments were eagerly evaluated before being
passed to the log method, even though `StackLevelAdapter.log()` would
then discard the message internally via its own `isEnabledFor()` check.
On high-frequency IPC paths this caused `pformat` to dominate CPU
usage (~60-70 %) as reported in #455.
This also restores the full diagnostic output (msg type, decoded
payload) that was temporarily commented out in 0373164 as a stopgap.
Resolves#455
Drop the stale sentinel experiment and fix the cancellation-path
comment. Document that cached regular `__aexit__()` failures are
always re-raised at the final consumer boundary.
Review: PR #488 (goodboy)
https://github.com/goodboy/tractor/pull/488
(this patch was generated in some part by `opencode` using
`gpt-5.6-sol` (`openai`))
Remove the unused `value` assignment after cache eviction and skip
unpacking stale resource state before raising its invariant error.
Review: PR #488 (Copilot)
https://github.com/goodboy/tractor/pull/488#pullrequestreview-4850557500
(this patch was generated in some part by `opencode` using
`gpt-5.6-sol` (`openai`))
Block the final user on its cached resource's `__aexit__()` and
raise regular cleanup errors at that user's ctx boundary.
Deats,
- serialize user registration and teardown under each cache-key lock
- keep queued entrants on the same lock through resource replacement
- preserve `Cancelled`, `KeyboardInterrupt`, and `SystemExit` flow
- cover successful, failing, cancelled, and re-entry teardown paths
Prompt-IO: ai/prompt-io/opencode/20260804T030309Z_65bf9df5_prompt_io.md
(this patch was generated in some part by `opencode` using `gpt-5.6-sol` (`openai`))
Copilot's 2nd-pass review (PR #468) caught a regression I landed in
the uds fix: `UDSAddress.get_random()` put the per-call token AFTER
`@{pid}` (`no_runtime_*@{pid}.{token}.sock`), which breaks the
`tractor._testing._reap` matcher
`^(?P<name>.+)@(?P<pid>\d+)\.sock$` — so no-runtime orphan socks
stopped matching and never got reaped/attributed. Move the token
INTO the name (`{prefix}.{token}@{pid}.sock`) so the canonical
`@{pid}.sock` suffix stays intact for both that regex and the
`spawn._reap` reconstruction; also bump the token to 8 hex chars.
Also two robustness nits from the same review,
- `tests.conftest._measure_sustained_headroom()`: guard `frac <= 0`
before `1./frac` — a 0/parked-core freq read would
`ZeroDivisionError`, get swallowed by the broad `except` into a
1.0 (no-throttle), defeating the probe on the exact broken box it
should flag; read 0 as max throttle.
- `scripts/cpu-perf-check`: mark the burn procs `daemon=True` and
wrap sampling in `try/finally` so a Ctrl-C / error reaps them
instead of leaving stray CPU hogs.
Regressed-by: 09c50f49 (uds no-runtime token placed after `@pid`)
Found-via: Copilot review #4595803812 (`_testing._reap` regex)
Review: PR #468 (Copilot)
https://github.com/goodboy/tractor/pull/468#pullrequestreview-4595803812
(this patch was generated in some part by [`claude-code`][claude-code-gh])
[claude-code-gh]: https://github.com/anthropics/claude-code
Finish TODO #2 from PR #468: the last raw `cpu_scaling_factor()`
call-sites still used the static-only check (blind to the
sustained-load power-cap). Point them at the `cpu_perf_headroom()`
SUPERSET — it calls `cpu_scaling_factor()` internally + the
session-cached throttle probe, so it's always >= the old factor
(never LESS headroom, so CI stays green).
Migrated,
- `test_spawning`: the `fail_after(1 * ...)` smoke deadline
(#465-added).
- `test_inter_peer_cancellation`: `this_fast` budget.
- `test_docs_examples`: example-run `timeout`.
- `test_resource_cache`: fan-out `timeout`.
`test_cancellation` already used `cpu_perf_headroom()`; only its
explanatory comments still name the static helper.
(this patch was generated in some part by [`claude-code`][claude-code-gh])
[claude-code-gh]: https://github.com/anthropics/claude-code
W/o a live runtime `get_random()` named UDS socks purely by
`(prefix, pid)`, so two calls in one proc returned the SAME
`no_runtime_*@{pid}.sock` — the 2nd `.bind()` then tripped
`EADDRINUSE`. Append a per-call `uuid4().hex[:6]` token so
each call yields a distinct sockpath.
This fixes the 3 `tests.discovery.test_tpt_bind_addrs` uds
failures (one registrar + disjoint-bind, two non-registrar
binds) where `reg_addr` and a "random" bind addr aliased —
the uds CI job's only red, surfaced when this branch rebased
onto the newer base that carries those tests.
Scoped to the no-runtime branch ON PURPOSE: the runtime
`{name}@{pid}` convention stays deterministic so
`spawn._reap.unlink_uds_bind_addrs()` can still reconstruct
+ unlink a SIGKILL'd subactor's sock (#454).
Deats,
- `_uds.py`: add `uuid4` import + token in the no-runtime
branch
- `_testing.addr`: tighten `get_rando_addr()` docstring re the
intra-proc per-call token (not just pid-keyed namespacing)
(this patch was generated in some part by [`claude-code`][claude-code-gh])
[claude-code-gh]: https://github.com/anthropics/claude-code
Both `_burn()` impls (the `cpu-perf-check` script + the
`_measure_sustained_headroom()` probe in `tests.conftest`) grew `x`
~x**2 per iter, so the loop quickly went bigint alloc/mul-bound — a
noisy CPU load + needless memory across N procs. Mask each step to
64-bit for a steady, fixed-width ALU burn (still pegs every core,
which is all the freq probe needs).
Also, `_read_mhz()` opened the sysfs freq files without a ctx-mgr;
`with open(...)` so the FD closes deterministically (matches the
script's own `_read()`).
Review: PR #468 (Copilot)
https://github.com/goodboy/tractor/pull/468
(this patch was generated in some part by [`claude-code`][claude-code-gh])
[claude-code-gh]: https://github.com/anthropics/claude-code
`cpu_perf_headroom()` is a strict SUPERSET of `cpu_scaling_factor()`
— it folds in static cpu-freq scaling + the slow-CI bump AND the
sustained-load power-cap throttle probe. Point the timing-deadline
call sites at it so they're robust to sustained throttling too (the
gremlin behind mass `trio` deadline-miss flakes),
- `test_legacy_one_way_streaming`: `time_quad_ex` cancel-deadline +
`test_a_quadruple_example`'s `this_fast` smoke bound.
- `test_cancellation`: `test_cancel_via_SIGINT_other_task` +
`test_fast_graceful_cancel_*` deadlines.
`cpu_perf_headroom >= cpu_scaling_factor` always, so never less
headroom (CI stays green); `cpu_scaling_factor()` remains the
internal static component.
(this patch was generated in some part by [`claude-code`][claude-code-gh])
[claude-code-gh]: https://github.com/anthropics/claude-code
Standalone CLI companion to `cpu_perf_headroom()` (20cb99ec): idle
freq snapshots LIE — every static knob (`governor`, EPP,
`platform_profile`, `scaling_max_freq`) can read "performance"
while a firmware/EC power cap (AMD PPT/STAPM + friends) clamps the
package to ~30% the moment a sustained multi-core load lands,
masquerading as a `trio`-backend deadline-miss "regression" on
byte-identical code.
Deats,
- burns every core for `CPU_PERF_SECS` (default 4s) and samples the
ACHIEVED `scaling_cur_freq` steady-state (post boost-ramp) vs the
package max ceiling,
- exits 0 when the sustained fraction clears
`CPU_PERF_HEALTHY_FRAC` (default 0.45), 1 when throttled — so it
gates a suite run: `scripts/cpu-perf-check && pytest tests/ ...`,
- prints the static knobs first (to show they all read fine) then
the remediation list on failure (`platform_profile` bounce, USB-C
PD replug, `ryzenadj`, reboot) w/ the key reminder: do NOT bump
test budgets — the box is slow, not the code.
(this patch was generated in some part by [`claude-code`][claude-code-gh])
[claude-code-gh]: https://github.com/anthropics/claude-code
Mass `trio` deadline-miss failures on byte-identical code turned
out to be a firmware/EC power-cap (AMD PPT/STAPM) clamping the
all-core sustained clock while every static knob (`governor`,
`scaling_max_freq`, EPP, platform-profile) still read "performance"
— invisible to the existing `cpu_scaling_factor()` check. See
`scripts/cpu-perf-check` + the
`ai/conc-anal/trio_033_cancel_cascade_slowdown_depth3_issue.md`
notes.
Deats,
- add `_measure_sustained_headroom()` to `tests/conftest.py`: a
one-shot ~0.9s all-core burn (explicit `fork`-ctx `mp` procs)
sampling achieved-vs-max freq AFTER the boost window; under a 0.6
gate it returns the full inverse fraction (capped 4x), else 1.0;
best-effort 1.0 on non-linux or any error,
- add `cpu_perf_headroom()`: `max()` of the static scaling factor
and the (session-cached) sustained probe,
- inflate deadline budgets by it in `test_dynamic_pub_sub`, both
`test_clustering` cases, the
`test_multi_nested_subactors_error_through_nurseries` pexpect
waits + `test_nested_multierrors`,
- `xfail(strict=False)` `test_nested_multierrors` depth=3 under
throttle: the deep tree trips tractor's INTERNAL reap deadlines
(`soft_kill`/`hard_kill` `terminate_after=1.6`) minting a
`Cancelled` inside the runtime — not fixable by test-budget
inflation; auto-clears once the box un-throttles.
(this patch was generated in some part by [`claude-code`][claude-code-gh])
[claude-code-gh]: https://github.com/anthropics/claude-code
Gud Boi
Bd
dependabot[bot]
root
mahmoud