git:20260912.f8c5c0d to git:20260925.01c05e3

26 added, 21 removed. Audit B to B.

---

name: provider

- description: Configure Claude or Ollama for the next local Claude Code session, verify real API requests, list models, or update provider control.

+ description: Configure Claude, Ollama, or MiniMax for the next local Claude Code terminal or VS Code session, verify real API requests, list models, or update provider control.

disable-model-invocation: true

allowed-tools: Bash

---

- Validate the argument as exactly one of `claude`, `ollama`, `status`, `models`,

- `doctor`, or `update`. Empty means `status`. Reject other arguments, especially

- internal credential actions. Do not interpolate arbitrary text into a shell.

+ Validate arguments against this exact grammar before running anything:

- Execute the installed launcher with that validated literal argument:

+ - No argument means `status`.

+ - One argument: `claude`, `ollama`, `minimax`, `status`, `models`, `doctor`, or `update`.

+ - Two arguments: `models ollama`, `models minimax`, `doctor ollama`, or `doctor minimax` only.

+ - Reject everything else, especially internal credential actions. Do not interpolate arbitrary text into a shell.

+ Execute the installed launcher with only the validated literal arguments:

```bash

"$HOME/.local/bin/backs-provider" status

```

- Replace `status` above only with the validated action. The launcher supplies the

- saved BACKS project and Python interpreter used by the IDE credential helper.

+ Replace `status` above only with a validated action and optional literal provider.

+ The launcher supplies the saved BACKS project and Python interpreter used by the IDE credential helper.

- - `ollama` checks the installed credential helper and makes minimal real Messages

- API requests to selected models before saving the complete profile. These

- checks use Ollama inference quota. A failing check must not be called success.

- - `claude` restores saved provider configuration, not an authentication guarantee.

- - `status` reports saved configuration, not the running model's transport.

- - `models` lists catalog alternatives; they are not automatic request-time retry.

- - `doctor` tests the installed helper and real Messages API without changing settings.

- - `update` refreshes provider code without replacing the primary BACKS runtime.

+ - `minimax` resolves `MINIMAX_API_KEY` through the existing BACKS environment, tests real streaming tool use and a tool-result round trip, then switches the whole lineup. The shipped profile selects MiniMax-M3 for main, Opus/Sonnet/Haiku aliases, and subagents. Checks use MiniMax inference quota; a failed check must not be called success.

+ - `ollama` checks its installed credential helper and makes minimal real Messages API requests to selected models before saving the complete profile. These checks use Ollama inference quota.

+ - `claude` restores the original provider configuration across either alternate provider, not an authentication guarantee. Never change the OAuth credential store.

+ - `status` reports saved configuration and saved preflight results, not the running model's transport.

+ - `models` and `doctor` use the saved alternate provider, or Ollama when none is active. An explicit `minimax` or `ollama` argument selects which one to inspect.

+ - MiniMax's model list comes from its documented compatibility list, not account entitlement discovery. `models` does not test inference or guarantee request-time fallback.

+ - `doctor` tests the installed helper and real Messages API without changing settings. MiniMax additionally tests streaming and tool-result handling.

+ - `update` refreshes provider code, profiles, helpers, and this skill through the existing validated updater. It does not replace the primary BACKS runtime.

- A new local Claude Code process is required after a provider change. This skill

- cannot repair an already unauthenticated session or convert a cloud/Remote

- Control subscription session to Ollama. Use the terminal launcher in that case.

- Never run `__key`, print credentials, read a key file aloud, or copy `.env` values.

- Never change managed organization policy or the Claude OAuth credential store.

+ Start a new local Claude Code process after switching. In VS Code, run Developer: Reload Window, then open a NEW LOCAL Claude Code chat. An existing session is not hot-switched. Updating Claude Code itself or the general skills plugin alone does not update the separate provider-control installation.

+ The shared Claude Code settings and project overlay cover the CLI and Claude Code VS Code extension, not Copilot, Codex chat, or a cloud/Remote Control subscription session. An unauthenticated chat cannot run this skill; use the terminal launcher instead.

+ Inherited shell or VS Code Anthropic environment variables may override saved settings. Follow the launcher's warning; never claim a saved profile changed the running process. Never run the MiniMax setup wizard over this managed configuration.

+ Never run `__key` or `__minimax_key`, print credentials, read a key file aloud, or copy `.env` values. Never change managed organization policy. Preserve unrelated hooks, MCP settings, and user preferences.