[P1] Expand `rlm_chain` with messages, structured output, retries, streaming, and middleware #13

New issue

Closed

opened 2026-08-12 01:19:13 +00:00 by lost-rob0t · 0 comments

lost-rob0t commented

2026-08-12 01:19:13 +00:00

(Migrated from github.com)

Goal

Build the reusable Prolog-native model/tool abstraction layer needed by agents and RLMs without cloning LangChain's Python object model.

Scope

normalize message roles/content and multimodal metadata where feasible;
expose provider capabilities and model routing hooks;
add prompt templates/binding;
add structured-output schemas and validation;
add retry/backoff policy;
add streaming event protocol;
add middleware/hooks around request, model response, tool call, and completion;
expose tracing and usage/cost metadata consistently;
keep the API idiomatic Prolog terms/predicates.

Acceptance criteria

one provider request can stream normalized events;
structured output is validated before downstream use;
retry policy is bounded and observable in trace events;
middleware can inspect/transform a request without provider-specific branching;
agent/RLM code depends on canonical rlm_chain interfaces rather than direct provider details.

Dependencies

#5
#10
parent #3

Research

See RLM-RESEARCH-007-langchain-langgraph-port.org.

## Goal Build the reusable Prolog-native model/tool abstraction layer needed by agents and RLMs without cloning LangChain's Python object model. ## Scope - normalize message roles/content and multimodal metadata where feasible; - expose provider capabilities and model routing hooks; - add prompt templates/binding; - add structured-output schemas and validation; - add retry/backoff policy; - add streaming event protocol; - add middleware/hooks around request, model response, tool call, and completion; - expose tracing and usage/cost metadata consistently; - keep the API idiomatic Prolog terms/predicates. ## Acceptance criteria - one provider request can stream normalized events; - structured output is validated before downstream use; - retry policy is bounded and observable in trace events; - middleware can inspect/transform a request without provider-specific branching; - agent/RLM code depends on canonical `rlm_chain` interfaces rather than direct provider details. ## Dependencies - #5 - #10 - parent #3 ## Research See `RLM-RESEARCH-007-langchain-langgraph-port.org`.