Skip to content

Releases: vincx2000/opendreams

v0.0.1-alpha — pipeline + cross-domain eval

10 May 16:28
@vincx2000 vincx2000
v0.0.1a1
f9714b1

Choose a tag to compare

View all tags
v0.0.1-alpha — pipeline + cross-domain eval Pre-release
Pre-release

First public release. The pipeline is complete; the eval surfaced cross-project memory pollution as a real, measurable phenomenon.

What's in

  • Three-stage pipeline: trace → reflect → consolidate → memory
  • Three adapters on a polymorphic base: claude_code (flagship), aider, generic_jsonl
  • Dual-backend LLM client: OpenAI-compatible (default; covers OpenAI, Ollama, vLLM, Together, Groq, Fireworks) + Anthropic native
  • AGENTS.md export with idempotent <!-- OPENDREAM:BEGIN --> / <!-- OPENDREAM:END --> markers
  • 15-task FastAPI eval suite with deterministic scoring
  • 174 tests, CI on Python 3.11 + 3.12

Eval result (cross-domain)

+0.0pp aggregate — but the test was unfair. Memory was consolidated from sessions of building OpenDream itself, then injected as AGENTS.md while the agent ran against a different codebase (eval/fixtures/library_api/). That's not what dreaming is supposed to do.

The cross-domain run surfaced two −20pp regressions on feature tasks: irrelevant memory actively hurts an agent. This finding is the project's most useful signal so far. v0.0.2 fixes it with a domain-matched two-pass eval.

Install

```bash
git clone https://github.com/vincx2000/opendreams && cd opendreams
pip install -e .
```

PyPI release lands with v0.0.3, once the discriminating eval clears the ≥5pp aggregate bar unambiguously.

Loading