Skip to content

A forwarded TCP connection never signals end-of-stream to the caller #88

Open

Description

@czpython

What happened

A caller that reads a forwarded TCP connection to end-of-stream never gets it. The sandbox side closes its socket; the caller's channel stays open with no data and no end.

Proof

A node server in a docker-sbx sandbox, createServer(s => s.end(Buffer.alloc(N, 97))), read from the caller through the gateway's direct-tcpip forward with asyncssh, reader.read(65536) in a loop until an empty read, with a timeout per read:

size=200000   got=200000   OK end=silence 10.3s
size=1000000  got=1000000  OK end=silence 10.1s
size=3000000  got=3000000  OK end=silence 10.1s
size=8000000  got=8000000  OK end=silence 15.4s
size=16000000 got=16000000 OK end=silence 15.3s

Every byte arrives. The end never does: each read after the last byte waits until the timeout.

Why it matters

A client protocol that ends with the server's close, HTTP/1.0 style or "send everything and close", hangs through the gateway. The chat bridge in druks works because it ends its answers with a newline, and a bug in its reader was only found because of that (linked below).

Where to look

The forward relay in src/gateway/server.py: when the tunnel side reports end-of-data for a forwarded connection, the caller's channel should get write_eof, and a close on the caller side should close the tunnel side. sandboxd logs ssh: port forward closed with non-zero relay exit on every forward close, which may be the same half-close.

Activity

  1. czpython commented on Oct 27, 2026

    @czpython
    OwnerAuthor

    Found through czpython/druks#1021.

  2. czpython commented on Nov 2, 2026

    @czpython
    OwnerAuthor

    The cause is in sandboxd, not in the gateway relay.

    A probe without drukbox (asyncssh through sbx ssh proxy, direct-tcpip to a node server in the sandbox that sends 200000 bytes and closes):

    • Caller sends no EOF: all bytes arrive, then no EOF and no channel close. The channel stays open until the caller's timeout.
    • Caller sends EOF: the channel closes at once, again with no EOF first.

    So sandboxd closes the forward only when both sides have half-closed, and never forwards the sandbox server's close. The gateway gets no signal to pass on. A test with an asyncssh sandboxd shows the relay already passes both an EOF and a close to the caller.

    Decision: no drukbox change. This goes to Docker as a sandboxd bug; this issue stays open until sbx ships a fix.


    Generated by Claude Code

  3. czpython commented on Nov 2, 2026

    @czpython
    OwnerAuthor

    Upstream: docker/sbx-releases#803


    Generated by Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions