Repository navigation
NexaPipe v0.2.0
NexaPipe exposes HTTP, HTTPS, WebSocket, TCP and UDP services that live behind NAT through a
single iroh/QUIC endpoint — without terminating TLS, without a control plane, and without a
server to rent.
v0.2.0 is a hardening release. A full review of the tree — server, client library, Android and
desktop apps — produced a batch of security and correctness fixes, alongside configurable
health checks, a draining shutdown, and access logs on the iroh tunnel path. It carries
breaking changes: default_backend is removed, and the server now refuses some configs it
used to "fix" silently. See Upgrading before restarting.
Highlights
Security hardening across the whole tree
Ordered by what an attacker could reach first:
- An empty enrollment token completed enrollment.
pending_enrollment = ""— the natural
way to clear a token — accepted any peer: the secret was rotated and handed to the caller,
and the real devices were dropped. Both sides of the comparison must now be non-empty, and a
blank token refuses to load. - Config hot reload logged the credentials. The endpoint's ed25519 secret key, the relay
auth token and TOTP secrets were printed on every reload. Hand-writtenDebugimpls now
redact them. - The client auth handshake had an unbounded allocation and no deadline. A 4-byte
0xFFFFFFFFlength made the client allocate 4 GiB, and one silent server could block every
request for that backend. Reads are capped at 64 KiB and the handshake has a 10 s timeout. - The desktop in-process path validated nothing. The frontend could publish an
unauthenticated LAN proxy that the privileged service path deliberately blocks. The loopback
and TUN-subnet checks now run before the service/process split, and a minimal CSP replaces
csp: null. - Request handling got deadlines and limits. The request head read has a 30 s total
deadline (an idle deadline is reset by one byte every 30 s); concurrent streams per
connection are capped at 256, refused rather than queued; request bodies are capped at
64 MiB checked before allocating; response compression switches to streaming past 1 MiB
instead of buffering unboundedly. - Android and desktop credentials are sealed. TOTP seeds, enrollment tokens and the relay
token used to sit in plaintextSharedPreferences/ localStorage, inside the Android backup
scope. They are now encrypted through the platform Keystore (AES-256-GCM), with old
plaintext values migrated on startup.
No protocol, wire-format or public API change.
Correctness fixes that were silently breaking traffic
- Chunked request bodies were silently dropped on the iroh tunnel path — only
content-lengthbodies were read, so the backend received a body with no framing and no way
to know it was truncated. They are now decoded as they arrive and forwarded with a computed
content-length. - Health checks were orphaned by a config reload: an unchanged route kept pointing at a
backend pool nobody probed, so it was never marked unhealthy again. Pools are now reused
across reloads, and probes are tracked and aborted with their route. - A reload could resurrect a spent enrollment token by overwriting live auth state with a
file snapshot read outside the lock; the read now happens inside it. - Host wildcards matched without a label boundary (
*example.commatchednotexample.com);
IPv6Hostheaders never matched a route; hop-by-hop headers were forwarded verbatim; a 502
echoed internal error details (backend URL, parse errors) to the client. - Plus fixes across TOTP validation, connection-pool lifecycle, TUN DNS parsing (RFC 1035
compression pointers), UDP flow limits, and a DNS cache for the TUN proxy.
Health checks are configurable, and the threshold is finally honoured
[health_check] is new, and every key is optional — absent means the defaults, so existing
configs keep working:
[health_check]
enabled = true # false disables probing entirely
interval = 10 # seconds between two rounds
timeout = 5 # seconds before one check is abandoned
threshold = 3 # consecutive failures before a backend leaves the pool
path = "/health"