[Feature] Add support for Claude Code/CLI as Anthropic backend (alternative to API calls)

[rahit]

Motivation

Currently, Biomni integrates Anthropic models using their API. Some users may prefer, or be required, to use the claude CLI (Claude Code) instead of direct API access (e.g., for local dev, easier debugging, organization/IT policy, or to avoid managing API keys in the Python environment).

Proposal

Add the ability to use Claude via the official Anthropic CLI (claude) as an alternative backend to direct API calls. This mode would:

• Accept the same prompt/messages payload as the API backend.
• Call the CLI tool via a subprocess and capture the output as the model response.
• Map errors/timeouts into Biomni's existing retry and error-handling logic.

User-facing behavior

• When LLM_PROVIDER=claude_code (or similar), Biomni will use the CLI instead of the API.
• If the CLI is not installed or cannot be authenticated, Biomni returns a clear error.
• No Anthropic API key is required for this mode.

Implementation suggestions / Feasibility

• Detect CLI presence (e.g. shutil.which('claude')).
• Start with non-streaming output (collect full stdout); add streaming later if feasible.
• For tool-calling, either:
  • Require the prompt to instruct CLI to return a strict JSON block for parsing.
  • OR mark CLI backend as text-only, reserving tool-calling for API-based usage.
• Provide a config option to select between API and CLI.

Acceptance Criteria

• Users can select CLI mode and run Claude models without an API key.
• Descriptive error when CLI is not installed or authenticated.
• Unit/integration test verifies prompt formatting and output parsing.
• Documentation update for new mode.

---

Related issues: None found for this specific feature. Closest: #284, #264