Ait

Optional second tier of the transcript memory pipeline: replace the rule-based heuristic summarizer with an LLM that can compress decisions, abandoned approaches, and error chains the heuristic cannot infer.

Added

• Pluggable LLM-backed transcript summarizer. Off by default — set summarizer.kind to "llm" in .ait/memory-policy.json to opt in.
• Two providers, both stdlib urllib (no SDK dependency):
  • anthropic — /v1/messages, default model claude-haiku-4-5-20251001.
  • openai-compat — <base_url>/chat/completions. Works with OpenAI, Azure OpenAI, Together, OpenRouter, vLLM, and Ollama (base_url=http://localhost:11434/v1).
• LLM failures (missing API key, network, malformed response) log a warning and transparently fall back to the heuristic. Misconfiguration never blocks the attempt lifecycle.

Configuration

{
  "summarizer": {
    "kind": "llm",
    "llm": {
      "provider": "anthropic",
      "model": "claude-haiku-4-5-20251001",
      "api_key_env": "ANTHROPIC_API_KEY",
      "max_chars": 600
    }
  }
}

API keys are never written to memory-policy.json — only the env var name. ait reads the secret at summary time so it stays in shell config / direnv / your system keychain.

Install: pipx install ait-vcs or npm install -g ait-vcs.