onbehalf 0.2.0

This version adds Claude Code and Codex as AI harnesses. The operator
selects one AI harness for each host.

Shared stack

• A stack source can have models.json as its model catalog, and
  AGENTS.md and skills/ at its root. Then the stack source does not
  depend on one AI harness. onbehalf writes the gateway providers of
  OpenCode from models.json. A stack source with
  opencode/opencode.json and no models.json works as before.

Model gateway

• onbehalf init --harness NAME selects the AI harness of the host. The
  value is in onbehalf.conf. A host without the value uses OpenCode.
• init, doctor and health check that the gateway serves the API of
  the AI harness. The check sends no model request.
• doctor --model-check, status --model-check and the health probe
  send their request with the API of the AI harness.
• sudo onbehalf harness NAME changes the AI harness of a host. It
  checks the command, the shared stack and the gateway first. Then it
  updates the routes of each gateway key and stops the runtimes of the
  old AI harness. It removes the links of the old AI harness and gives
  each user the links of the new AI harness. Each user keeps the gateway
  key, the personal settings and the private state.

Claude Code

• onbehalf init --harness claude selects Claude Code. The stack source
  must have models.json.
• stack install writes the managed settings of Claude Code from
  models.json: the gateway, the personal gateway key, the model catalog
  and the default model. /etc/claude-code/managed-settings.json is a
  link to the current release. Team settings go in
  claude/managed-settings.json of the stack source.
• Each user gets links in ~/.claude: CLAUDE.md to the shared
  AGENTS.md, and each shared skill, custom agent and command.
• A user can set a different default model in ~/.claude/settings.json.
• doctor and status warn about ~/.claude/.credentials.json.
• Claude Code has no service. onbehalf restart tells the user to start
  a new session.

Codex

• onbehalf init --harness codex selects Codex. The stack source must
  have models.json.
• stack install writes the system configuration of Codex from
  models.json: the gateway provider, the personal gateway key and the
  default model. /etc/codex/config.toml is a link to the current
  release. Team settings go in codex/config.toml of the stack source.
• Each user gets links in ~/.codex: AGENTS.md to the shared
  AGENTS.md, and each shared skill.
• A user can set a different default model in ~/.codex/config.toml.
• doctor and status warn about ~/.codex/auth.json.
• Codex has no service. onbehalf restart tells the user to start a new
  session.
• The gateway key of a user on a Codex host permits only the responses
  API.

Docs

• The guides give the setup and the use of each AI harness. The operator
  guide has a section about a change of the AI harness.
• examples/stack is now a stack source with models.json. It works for
  each AI harness. In OpenCode, its provider of the OpenAI API has the
  name onbehalf, not onbehalf-openai. A stack source that you copied
  from version 0.1.0 does not change.
• Recorded casts show the operator setup, the first login of a user and
  the report. The README, the site and the guides show them. make demo
  records them again from a clean test host. See demo/README.md.

Upgrade

Get the files of this version and run the installer again. A host keeps
OpenCode as its AI harness, and its stack source keeps its function. Then
run sudo onbehalf doctor. To use Claude Code or Codex, see the
operator guide.