Proudly Made in Nebraska. Go Big Red! 🌽 https://xkcd.com/2347/

A model-agnostic RLM harness, written in Rust. Companion to CryptoJones/ReCLamO-Harness (the Qwen-tuned v0.1).

v0.1 works against Qwen3.8 on pluto Flash-Next and solves its self-authored tasks perfectly. On the independent roundtable-eval (#22, frozen at ../ReCLamO-Harness sha e17580f), it loses to plain on small contexts and solves only 5/16 medium and 1/8 large. The fix list lives in ../ReCLamO-Harness NEXT-STEPS.md. This repo builds the model-agnostic version that also takes the fixes end to end.

The orchestrator is Rust; the REPL is the same vendored Python worker that v0.1 used — std-only stdlib, no third-party deps in the child process.

See PLAN.md for the goals, architecture, and success criteria.

• Phase 1 (scaffold) — SHIPPED. reclamo-anl --version works; cargo test passes (55/55 unit tests green). MockProvider round-trips against AnthropicProvider and OpenAICompatProvider request builders, including chat_template_kwargs passthrough for Qwen/Strata and reasoning_content normalization for Anthropic-style thinking.
• Phase 2 (loop + REPL + first model run) is the next milestone.

• PLAN.md — what I'm building and why.
• BACKLOG.md — backlog (mirrors the GitHub Issues tab once any are filed).
• src/ — the Rust crate.
  • providers/{base,anthropic,openai_compat,mock,reasoning}.rs — the provider abstraction (one trait, three adapters).
  • rlm.rs — the loop with the five NEXT-STEPS fixes baked in.
  • router.rs — plain-vs-harness routing (fix #1).
  • repl/{in_memory_repl,subprocess_repl,worker.py}.rs/.py — REPL stub (in-memory) and full subprocess variant (vendored worker.py).
  • prompts.rs + prompts/system_prompt.template — capability-driven system prompt.
  • logger.rs — JSONL trajectory.
  • parsing.rs — fenced code + FINAL/FINAL_VAR detection.
  • error.rs — typed limit errors with partial_answer.
  • config.rs — TOML profile parser.
  • cli.rs — reclamo-anl {run, ping, version}.
• evals/ — frozen-generators rig (phase 4).
• examples/ — needle, oolong-lite, long-doc-QA (phase 4+).
• profiles/ — example profile TOMLs.

Requires the Rust 1.75+ toolchain. cargo test runs the unit suite.

# Phase-1 sanity (works today).
cargo run -- --version
cargo test

# Phase-2 use (loop + REPL on a real model — wired up next).
cargo run -- run --profile profiles/anthropic.toml \
    --context path/to/context.txt \
    -q "Find the needle."

--provider mock skips network for the ping subcommand.

Tracked in BACKLOG.md; summary in PLAN.md.

Apache-2.0. See LICENSE.

Design inspired by alexzhang13/rlm and alexzhang13/rlm-minimal (both MIT); see NOTICE and THIRD_PARTY_LICENSES.md.

Proudly Made in Nebraska. Go Big Red! 🌽 https://xkcd.com/2347/