Skip to content

v0.5.6

Latest

Choose a tag to compare

@gabelul gabelul released this 20 Aug 13:14

Fixed

  • edit --help no longer advertises a flag it rejects. _add_common_flags() is shared by generate, edit and compose, so edit inherited -i's "repeatable (-i a -i b -i c)" wording — then cmd_edit exited with edit requires exactly one -i/--image. The enforcement was right (edit takes one source, compose takes 2-16); only the inherited help string was wrong. -i and --size now carry per-subcommand wording. If you want several references, that has always been compose.

  • --size says out loud when it won't be honoured. It works for generate and is silently dropped for edit/compose: the Responses API returns the input's aspect the moment there's an input image. Measured — generate 1024x1536 → 1024x1536, edit 2160x3840 → 852x1846, compose 2160x3840 → 853x1844, the latter two at roughly 850px on the short edge for portrait inputs. pixeltamer does send the value, so this is API behaviour rather than a dropped flag, but the caller got no signal at all: ask for 4K, receive 850px, size your downstream crop off a number that never arrived. Both subcommands now warn on stderr, --help is honest about it, and references/codex-backend.md carries the measurements. --backend api is unaffected — /v1/images/edits accepts size properly.

    Nothing auto-upscales to compensate. That would trade a silent surprise for a silent lie, and the caller is better placed to decide how to resize. There's a note in codex-backend.md on the trap that follows: an edit comes back small, so anything you upscale afterwards gets softer — text especially. If you composited text in before the edit, composite it back on after the upscale rather than round-tripping it through the model, because image models re-render text rather than preserving it.

  • The index-staleness test stopped failing for reasons that had nothing to do with the index. It listed *.md with a bare extension filter, so macOS AppleDouble sidecars (._codex-backend.md) counted as real reference files and got reported as missing from the routing map. Any checkout on a non-APFS volume — an external drive, a USB stick, a shared mount — grows those on every edit, so the suite broke for anyone in that situation the first time they touched a doc. Dotfiles are skipped now.