v0.6.0

Support for oh-my-openagent 5.0.x. omo 5.0 split the deep category into deep-low and
deep-high, and the bundled suggestions now follow omo 5.0.1. If you are still on omo 4.19,
stay on omodel 0.5.1 — a 4.19 omo does not read deep-low.

Upgrading is omodel --update for the standalone binary, or the command it prints for a
pipx/uv/pip install.

Added

• show, candidates and check (--json) carry degraded_reason: why availability is
  unknown — opencode is not on PATH, or what opencode models did instead of answering
  (exited non-zero, printed nothing, timed out). degraded: true alone read the same for a
  missing opencode and a broken one, and only the first is fixed by installing it. Every verb
  now also says so on stderr.

Changed

• Bundled suggestions now come from omo 5.0.1. omo split the deep category into deep-low
  and deep-high, so there are 9 categories now, not 8. The fallback chains follow omo's new
  models (gpt-6-astra, gpt-6-luna-fast, claude-fable-5-1, claude-opus-5-5, …), and the
  vercel and quotio-openai providers are no longer suggested (chatgpt-subscription is new).
• A deep category in your config or in a saved preset now shows up as deep-low, the same way
  omo 5.0 reads it (if deep-low is also set, the old deep is dropped, as omo does). The new
  name is written with your next save. omodel set cat:deep is refused with a message that
  points to cat:deep-low.

Fixed

• A provider's -thinking, -free, -base, -instruct, -chat or -draft build no longer
  stands in for the bare model omo named as if it were that model. glm-5.2-thinking filled a
  glm-5.2 slot as an exact match — no ≈, no warning — the same hole mini fell through in
  0.5.1, one suffix family over. Such a build now shows as a marked ≈ substitute when it is
  the same family as the model omo named, and is not offered at all when it is not (omo files
  kimi-*-thinking as its own family; big-pickle has none) — never as the model itself.
• Saving no longer changes what kind of file your config is. A save used to reset its permission
  bits to 0600, replace a symlinked config (a dotfile manager pointing ~/.omo/omo.jsonc at a
  repo file) with a detached copy, turn every CRLF line ending into LF, and — for a file with a
  UTF-8 byte-order mark — drop every comment in the document. Bits, link, endings and comments
  all survive now (a file that mixes endings keeps each line's own), and the preview diff
  compares exactly what the save compares, listing only the lines that change.
• --json always answers with JSON. A config that could not be parsed — or could not be opened
  at all: --config ~/.omo, one component short of the file, or a file you cannot read — made
  every verb exit 1 with an empty stdout (the unreadable case as a raw traceback). Both now
  return the standard failure payload with error: bad_config.
• omodel preset use +-1 (any preset reference int() refuses) is an unknown_preset refusal
  instead of a traceback.
• --dry-run reports changed: true whenever the real run would write. It looked at the config
  diff alone, so a write that only adopts a foreign config into the active preset (a sync
  conflict) previewed as a no-op and then wrote.
• apply --json reports the reasoning level it actually wrote: asked for none, it echoed
  none back for a file that now said off.
• omodel --check counts the targets omo actually defines (31), not an ultrawork sub-target
  on every agent (41) — ten of which every other verb refuses.
• omodel --restore no longer burns a backup slot when asked for a backup that does not exist.

Notes

A deep already in your config or presets keeps working — omo 5.0 reads it as deep-low, and
omodel shows it that way. It is written as deep-low the next time you save.

Full history: CHANGELOG.md