Retire Ox Alpha, and make the checks tell the truth

[phall1]

dot-doctor was reporting three failures against a healthy system, which is how
a health check teaches you to ignore it. None of the three were real; all three
checks were wrong.

Ox Alpha is retired

It was a limited-time free stealth preview. The model is gone; the five-provider
fan-out it justified spanned eight files and five harnesses. Removed the launcher,
the per-harness catalogs (Pi, Hermes, Goose, Grok, OpenCode), the doc, and
checks/ox-alpha.sh. Both modify_ scripts existed solely to inject Ox
providers, so Hermes and Pi own their runtime files outright again.

Source removal alone leaves materialized copies behind, so the six applied targets
and the now-entirely-Ox ~/.pi/agent/models.json are appended to
provision/retired-files.sha256, each hash verified against its source first.

The checks now assert what the runtime resolved

• opencode.sh asserted permissions/providers/plugins and an invented
  update: auto against the raw file on disk. OpenCode migrates legacy key names
  as it loads (permission → permissions, provider → providers, attachment
  → media), so a working config could never satisfy it. It now asks
  opencode2 debug config what was actually resolved — and warns when several
  config documents silently shadow each other, which immediately found a redundant
  opencode.json duplicating one MCP entry.
• blackbird.sh required the MCP server at .mcp.servers.blackbird. OpenCode
  accepts both that and .mcp.<name>; the daemon was connected and the plugin
  loaded the whole time. It now reads opencode2 mcp list and plugin list.
• claude.sh compared ~/.claude/settings.json against a source file that
  stopped existing when settings moved to modify_settings.json. With the guard
  false, the conditional fell through to its else branch and printed
  "settings.json matches source" — green forever, about a file that was not
  there. It now derives the managed block from the modify_ script itself and
  proves the merge preserves herdr's unmanaged hooks.

A silent breakage the new checks caught

checks/skills.sh found seven dangling skill adapters across Claude, OpenCode
and Hermes on its first run — including cyclomatic-complexity.

Root cause: ~/.agents/skills is enrolled in live preferences, so .chezmoiignore
excludes it and chezmoi apply never writes there. dot_agents/skills/ is meant to
be a seed — but nothing was seeding it. Because the adapters are symlinks, the
failure is invisible at the filesystem level and shows up only as a skill that
never loads. scripts/bootstrap/skills.sh now seeds missing files into the live
tree, never overwriting an existing one, and runs inside converge.sh.

Pins are derived, not restated

checks/pi.sh restated the package list modify_settings.json declares — six pins
across three sites. The only drift that could detect was "you edited one file and
not the other." All three now derive from the declaring config. Doing so
immediately surfaced npm:blackbird-pi@0.1.1, declared managed but absent from the
hand-copied assertion, so its absence had been invisible.

Docs describe the repo that exists

dot_claude/settings.json has not existed since the move to a modify_ merge
script, so every instruction to "edit settings.json and apply" named a missing file
and described the wrong operation. Also retires ~/CLAUDE.md: an untracked
verbatim copy of this repo's CLAUDE.md at 14ff43f, which every Claude session
started under $HOME loaded as project instructions, two months stale.

Result

dot-doctor: 0 failures, one warning (ghostty missing, unrelated).
mise run check green. chezmoi status clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HvvVk5xnS9hRBL9u1WcP6p