Skip to content

reference-internals-extension-points.md

Latest commit

 

History

History
79 lines (51 loc) · 2.77 KB

reference-internals-extension-points.md

File metadata and controls

79 lines (51 loc) · 2.77 KB
Raw

Internals — Extension Points

Where to plug in new behavior, in order of increasing depth.

Add a skill

Most common extension. Two edits:

  1. A line in getSkills (src/skills.metta) so the LLM knows the skill exists.
  2. A (= (my-skill $arg) ...) definition, either pure MeTTa or a py-call / translatePredicate.

Full walkthrough: tutorial-03-writing-a-custom-skill.md.

Add a remote skill

Same as above, but the body delegates to src/agentverse.py:

(= (my-remote-skill $arg)
   (py-call (agentverse.my_remote_skill $arg)))

Full walkthrough: tutorial-06-remote-agentverse-skills.md.

Add a channel

Three touch points:

  1. New Python module channels/myadapter.py implementing start_*, getLastMessage, send_message.
  2. A new branch in initChannels, (receive), and (send $msg) in src/channels.metta.
  3. New parameters declared via (= (MY_*) (empty)) and bound by configure.

Full walkthrough: tutorial-04-adding-a-channel.md.

Add an LLM provider

In src/loop.metta, the main dispatch is:

(if (== (provider) OpenAI)
    (useGPT ...)
    (if (== (provider) Anthropic)
        (py-call (lib_llm_ext.useClaude $send))
        (if (== (provider) ASICloud)
            (py-call (lib_llm_ext.useMiniMax $send))
            (py-call (lib_llm_ext.useAsi1 $send)))))

To add a provider:

  1. Implement a call function in lib_llm_ext.py (or a new module).
  2. Add a branch to the if chain.
  3. Use the new provider name in the configure provider ... line or via command-line provider=....

Change the prompt

The agent's identity and values are in memory/prompt.txt. The run-time prompt template that sandwiches it is in getContext in src/loop.metta. Edit carefully — the output-format instruction is what keeps the LLM producing valid skill s-expressions.

Change the embedding model

In src/memory.metta, the embed function dispatches on embeddingprovider:

(= (embed $str)
   (if (== (embeddingprovider) Local)
       (py-call (lib_llm_ext.useLocalEmbedding (string-safe $str)))
       (useGPTEmbedding (string-safe $str))))

To add a new backend, add a branch and implement the Python function.

Change the reasoning library

lib_nal.metta and lib_pln.metta are plain MeTTa files loaded by lib_omegaclaw.metta. Add new rule definitions directly, or swap in a different logic library entirely — the only required surface is whatever operator the LLM invokes through (metta ...).

See also