v0.5.57

The Codex advisor starts again

0.5.56 added -s workspace-write -a never to the Codex advisor's launch, and it stops Codex negotiating an interactive terminal. Every session with a Codex advisor died after conclave's banner, before a session id was printed — so there was nothing in conclave sessions to inspect either. Reproduced here and independently bisected on another machine.

Reverted. This release is 0.5.56 with those two flags removed, and nothing else changed.

Verified by starting a Codex advisor, not by a passing suite. That distinction is the whole reason the regression shipped: the change was entirely about how a seat launches, the flags were checked for whether they parse, and no test in this project spawns a real Codex — so a green suite could not observe the only thing at risk.

What the reverted change was for, and why it is now an open question

It removed Codex's per-command approval prompts, each of which becomes a pause someone must answer; one went unanswered for 29 minutes on a reported run and burned most of a 45-minute ceiling.

The project that hit the regression, though, already runs permissions: bypass — so those prompts were never being asked there, and the flags were pure addition: nothing to gain, the launch to lose. Conclave already has a bypass mechanism, and the run that reported the stall was not using it.

So the gap may be an operator who did not know the mechanism existed, rather than a missing sandbox. That is cheaper to fix and cannot break a launch. If the sandbox is wanted anyway, the deliverable is a test that starts a real Codex seat and asserts it reaches ready.