Skip to content

row 166: strike the unserved plane FFI subprocess spawn path (its env secret resolved at the spawn) - #866

Merged
MattJackson merged 2 commits into
predevfrom
lane-row166-ffi-subprocess-strike
Oct 15, 2026
Merged

MattJackson merged 2 commits into
predevfrom
lane-row166-ffi-subprocess-strike

Conversation

@MattJackson

Copy link
Copy Markdown
Collaborator

Row 166, per the ARCHITECT ruling of 2026-10-13 (F11: unserved code is deleted with proof).

Reachability proof

  • pipe::open_subprocess has one production caller, plane_host/egress.rs:562, which passes &[] as the program allowlist.
  • command_admissible (pipe.rs:278) requires an absolute path that is ON the allowlist, so it is false for every input. It runs before resolve_child_env (pipe.rs:296-299 vs :309).
  • So the command decode, the env secret resolution (resolve_linked_string at the spawn, on a request path), the child lifecycle and the pipe registry were reached only by pipe_tests through a host-authored allowlist that no served path holds.
  • Nothing else opens a pipe: RawConn answers Unsupported (egress.rs).
  • The only other users of the pipe slots are the vtable entries (vtable.rs:71-72). They stay, and answer exactly as before for every pipe a plane could name.

Change

  • EgressKind::Subprocess answers Refused at the seam, before anything is decoded. Every plane saw the same answer before.
  • pipe.rs keeps only pipe_read / pipe_write: Refused on a bad host context or null out-pointer, otherwise Gone, Fault on a caught panic.
  • pipe_tests: the FFI refusal cell carries a secret env record and asserts the slots' answers. The cells that drove the struck mechanics go with them.

Gates whose ledgers name pipe.rs still match: one-memory-abi:c-unwind (the slots stay extern "C-unwind") and construction known_sites (catch_unwind stays).

Independent of #863/#864/#865.

… and is struck; a plane-driven subprocess open is refused at the seam

Row 166 (ARCHITECT 2026-10-13, F11). The only production caller of pipe::open_subprocess
(egress.rs:562) passed an empty program allowlist, and command_admissible requires an absolute path ON
that allowlist. So every plane-driven open was refused before resolve_child_env, and the command
decode, the host-side env secret resolution (resolve_linked_string at the spawn, a request path), the
child lifecycle and the pipe registry ran in no served path. Nothing in the tree opened a pipe any
other way: RawConn answers Unsupported.

- egress_open: EgressKind::Subprocess answers Refused at the seam, before anything on the descriptor
  is decoded or resolved (the same answer every plane got).
- pipe.rs keeps the two vtable slots. Each answers as it did for every pipe a plane could name (Gone),
  after the host-context and out-pointer checks, failing closed on a caught panic.
- pipe_tests: the FFI refusal cell now carries a secret env record and asserts both slots answer Gone
  for any pipe and Refused for a null out-pointer. The cells that drove the struck mechanics through a
  host-authored allowlist (a posture no served path held) go with them.
… the ten hits #866's strike drained (Law 9: the ceiling ratchets down to the measure)
@MattJackson
MattJackson merged commit a6b23ce into predev Oct 15, 2026
1 check was pending
@MattJackson
MattJackson deleted the lane-row166-ffi-subprocess-strike branch October 15, 2026 05:33
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant