Support using installed CLI LLM tools instead of API keys for reflection

[nicholaswhittle]

Feature Request

Summary: Allow cm to use locally installed CLI-based LLM tools (e.g., claude, codex, gemini) for the reflection/summarization step, instead of requiring an API key (OpenAI/Anthropic) to be configured.

Motivation

Currently, cm requires an API key (e.g., OPENAI_API_KEY) to perform LLM-powered reflection (cm reflect, cm context, etc.). Many users already have CLI tools like claude (Claude Code), codex, or gemini-cli installed and authenticated — these tools handle their own auth, token management, and model selection.

It would be convenient to leverage these existing CLI tools as the LLM backend for reflection, avoiding the need to separately provision and manage API keys just for cm.

Proposed Behavior

• Add a configuration option (e.g., llm_backend: cli) that tells cm to shell out to an installed CLI tool for LLM calls instead of hitting an API directly.
• Allow the user to specify which CLI tool to use (e.g., claude, codex, gemini).
• Fall back to the current API-key-based approach if no CLI tool is configured or available.

Benefits

• No extra API key management — reuse existing CLI auth
• Model flexibility — users can use whichever model/provider their CLI tool is configured for
• Lower barrier to entry — users who have Claude Code or Codex installed can use cm reflection immediately without setting up a separate API key

Example

# In ~/.cass-memory/config.json
{
  "llm_backend": "cli",
  "llm_cli_command": "claude"
}

Then cm reflect would invoke the claude CLI to perform summarization rather than calling the OpenAI API directly.

[Dicklesworthstone]

Implemented in bfcf51e.

You can now use installed CLI LLM tools (claude, codex, gemini) instead of API keys:

{
  "provider": "cli",
  "cliCommand": "claude"
}

Or via environment variable: export CASS_CLI_COMMAND=claude

How it works:

• Shells out to the configured CLI tool with the prompt
• Requests JSON output and validates against the expected schema
• Supports auto-detection: if no API keys are set but claude/codex/gemini is on PATH, it's used as a fallback provider
• Integrated into the existing multi-provider fallback chain

[nicholaswhittle]

Hey — the CLI provider feature from this issue looks great, but the latest release binary (v0.2.3, Jan 7) doesn't include it yet. provider: "cli" gets rejected by the zod schema with Expected 'openai' | 'anthropic' | 'google'.

Would it be possible to cut a new release that includes this? I'd love to switch to using claude CLI instead of API keys for reflection. Thanks!