Skip to content

Releases

v2.5.2 — Codex fallback runs stay tracked

Choose a tag to compare

@PythonShe PythonShe released this 23 Sep 15:34
30d20ad

Codex fallback runs stay tracked. Every skill with a cross-model Codex seat falls back to the bare codex exec CLI when the plugin's companion script isn't found, and that fallback ran with -s read-only --ephemeral. --ephemeral tells Codex not to write its session rollout under ~/.codex/sessions, and local usage trackers (the tokens CLI and similar) count Codex usage only from those rollouts. So every fallback Codex review or investigation spent real quota without showing up in local usage stats. A Coda run in a downstream project surfaced it: its whole-PR Codex review used about a tenth of a five-hour window, and the local record showed nothing.

The fallback shape drops --ephemeral everywhere it appeared: the shared reviewer contract (shared/codex-reviewer-core.md), the canonical statement in shared/invariants.md, and Tuner's investigator fallback. The contract now says not to add the flag back, and why. -s read-only still keeps Codex from editing the tree, and < /dev/null, the prompt and verdict files and the timeout are unchanged. Routing surfaces are unchanged, so no eval re-run is needed. Run /plugin marketplace update to pick it up.

v2.5.1 — Codex fallback stops hanging

Choose a tag to compare

@PythonShe PythonShe released this 04 Sep 19:02
188686c

The Codex fallback stops hanging. Every skill that seats a cross-model Codex agent falls back to the bare codex exec CLI when the plugin's companion script is absent, and that fallback has been quietly the least reliable seat in the line: codex exec drains stdin to EOF whenever stdin is not a TTY — prompt argument or not — and a subagent or background shell hands it an open pipe nobody writes to, so the process wedges forever at 0% CPU with at most Reading additional input from stdin... on the log (upstream openai/codex#20919, still open, no --no-stdin flag). Six separate runs in a downstream project hit it between June and September, and it is the hang that cost Jam its Codex seat in v2.3.0. The cure was found and re-found by hand each time; this release writes it down once. The shared contract's fallback becomes one exact shape — prompt written to a scratchpad file, -o verdict file, -s read-only --ephemeral, < /dev/null on the call, the Bash tool's ~10-minute timeout, foreground inside a background subagent, never in the conductor's own Bash call, because a hung codex exec blocks whoever ran it until a timeout fires or a human interrupts, and in the conductor's hands that is the whole run stalled — with the empty-output-is-failure rule extended to it, the stdin line named as the hang signature, and no relaunch: a second attempt doubles the wall-clock the panel already paid, so the shape has to be right on the first try. Tuner's investigator fallback, which until now told the conductor to run a bare codex exec in a background Bash call — the exact recipe for the hang — is rewritten as a marked copy of the same shape, and shared/invariants.md gains the canonical statement so the two cannot drift. The README's Codex prerequisite says the same in one sentence. Routing surfaces unchanged; no eval re-run needed.