Skip to content

reference-internals-skill-dispatch.md

Latest commit

 

History

History
53 lines (35 loc) · 2.37 KB

reference-internals-skill-dispatch.md

File metadata and controls

53 lines (35 loc) · 2.37 KB
Raw

Internals — Skill Dispatch

This page traces what happens between an LLM response landing in the loop and a skill actually running.

Expected LLM output shape

The prompt demands a tuple of up to 5 skill s-expressions:

((skillName1 "arg1") (skillName2 "arg2") ...)

With the hard rules that every argument is a quoted string and no MeTTa variables may appear.

Step-by-step dispatch

From src/loop.metta:

  1. Raw LLM string → $respi.
  2. Parenthesis repair — helper.balance_parentheses $respi → $resp. Common LLM mistakes (missing closers) are fixed here.
  3. First-character check — if $resp does not start with (, the agent receives a reminder prompt instead of a real dispatch; the LLM tries again next turn.
  4. Parse — catch (sread $response) → $sexpr. On parse failure, HandleError records MULTI_COMMAND_FAILURE_....
  5. Fan out — (superpose $sexpr) produces one binding per skill call in the tuple.
  6. Evaluate each — (catch (eval $s)). On success, the result is normalized via helper.normalize_string. On failure, HandleError records SINGLE_COMMAND_FORMAT_ERROR_....
  7. Aggregate — all results are collapsed into RESULTS: ((COMMAND_RETURN: (cmd result)) ...).
  8. Feedback — stored as &lastresults, fed back into the next prompt as LAST_SKILL_USE_RESULTS.

How eval $s resolves to a skill

MeTTa evaluates the head of the expression against the AtomSpace. Skills are defined as plain equations:

(= (remember $str) ...)
(= (shell $cmd)    ...)
(= (metta $expr)   ...)

So (shell "ls") matches the equation for shell and runs its body.

Bridges enter via:

  • (py-call (module.function args)) for Python.
  • (translatePredicate (predicate ...)) and !(import_prolog_function name) for Prolog.

Errors propagate via &error

HandleError appends to the &error state. When addToHistory runs, if &error is non-empty it is concatenated as ERROR_FEEDBACK:. On the next turn the agent sees the error and has the opportunity to correct course.

See also