Skip to content

🏹 v0.1.11 : Harness Generalization, Repository Context & Demo

Latest
Latest

Choose a tag to compare

View all tags
@Shashikant86 Shashikant86 released this 20 Jul 19:26
v0.1.11
6c4b8b9

RLM Code v0.1.11

This release brings the harness-generalization ideas from Alex Zhang’s July 2026 RLM post into RLM Code, alongside a maintained version of our AI Engineer World’s Fair 2026 demo.

Highlights

  • Added reusable repository-context construction with bounded, ranked evidence.
  • Added reference, repo_evidence, and lid RLM execution profiles.
  • Added configurable root observations: raw, metadata-only, opaque, or explicit.
  • Added structural and offloaded history modes for longer trajectories.
  • Added decomposition guidance and bounded subcall tracing.
  • Added trajectory-similarity metrics based on edit distance and trigram overlap.
  • Preserved explicitly supplied benchmark context instead of replacing it.
  • Added task-family, domain, split, expected-answer, and context-length benchmark metadata.
  • Improved runner metrics with prompt sizes, hashes, structural actions, and subcall counts.

New demos

Harness Generalization Demo

examples/july_harness_generalization/

An API-free, deterministic demo showing that the same RLM harness can generalize from a small commerce task to an eight-times-larger support task while:

  • keeping private context out of the root prompt;
  • decomposing work through bounded subcalls;
  • retaining compact subcall traces;
  • offloading growing history;
  • preserving structurally similar trajectories;
  • producing the correct final answers.

Run it with:

uv run --frozen python examples/july_harness_generalization/demo.py

#### AI Engineer World’s Fair 2026 Demo

examples/aie_world_fair_2026/

A maintained version of the live RLM probe from our San Francisco conference talk, updated for the
current RLM Code APIs.

It includes:

- an Ollama-ready local demo;
- Gemini configuration instructions;
- current repository-evidence construction;
- Docker-first sandboxing;
- prompts and suggested presentation use cases.

Run it with:

uv run --frozen python examples/aie_world_fair_2026/rlm_probe.py

The complete presentation and original conference materials remain available in the AI Engineer World’s
Fair talk repository (https://github.com/Shashikant86/rlm-codebase-aie-wf26-talk).

### Additional improvements

- Added CLI and configuration support for context, observation, history, decomposition, and execution
  profiles.

- Expanded environment and configuration documentation.
- Included examples in source distributions.
- Corrected the RLM paper reference.
- Updated package metadata and documentation for v0.1.11.
Assets 2
Loading