define agentic Prolog RLM harness and dual-version MCP runtime #2

Merged

lost-rob0t merged 4 commits from agent/agentic-harness-research into main

2026-08-12 00:33:38 +00:00

lost-rob0t commented

2026-08-12 00:33:12 +00:00

(Migrated from github.com)

What changed

redefine prolog-rlm as a Prolog-native LM harness and agent runtime, not just a Python-RLM translation
make the first executable milestone a real provider + typed plan interpreter + external context + tool loop + depth-1 RLM
add first-class rlm_agent and rlm_mcp architecture alongside rlm_chain and rlm_graph
specify capability-narrowed recursive subagents, bounded worker pools, structured outcomes, tracing, and durable artifact state
target both MCP 2025-11-25 and 2026-07-28 behind one canonical internal representation
document legacy sessionful initialize/initialized behavior and new stateless server/discover / per-request metadata behavior
prefer 2026 when explicitly supported and fall back to 2025 for maximum interoperability
add research records 002–009 covering agentic harnesses, typed symbolic execution, PrologMCP repair loops, SWI agent runtime, dual-version MCP, LangChain/LangGraph semantics, adaptive recursion, and durable artifact context
fold in Recursive Agent Harnesses and Self-Reflective Program Search as additional evidence

Validation

Documentation/research-only change. No executable runtime exists yet. The records include acceptance criteria that define the implementation slices and conformance matrix.

## What changed - redefine `prolog-rlm` as a Prolog-native LM harness and agent runtime, not just a Python-RLM translation - make the first executable milestone a real provider + typed plan interpreter + external context + tool loop + depth-1 RLM - add first-class `rlm_agent` and `rlm_mcp` architecture alongside `rlm_chain` and `rlm_graph` - specify capability-narrowed recursive subagents, bounded worker pools, structured outcomes, tracing, and durable artifact state - target both MCP `2025-11-25` and `2026-07-28` behind one canonical internal representation - document legacy sessionful `initialize`/`initialized` behavior and new stateless `server/discover` / per-request metadata behavior - prefer 2026 when explicitly supported and fall back to 2025 for maximum interoperability - add research records 002–009 covering agentic harnesses, typed symbolic execution, PrologMCP repair loops, SWI agent runtime, dual-version MCP, LangChain/LangGraph semantics, adaptive recursion, and durable artifact context - fold in Recursive Agent Harnesses and Self-Reflective Program Search as additional evidence ## Validation Documentation/research-only change. No executable runtime exists yet. The records include acceptance criteria that define the implementation slices and conformance matrix.