Metadata-Version: 2.4
Name: dsh-rla
Version: 0.0.1
Summary: Recursive learning agent — a DeepSeek Harness plugin for proactive self-evolution. (Name reservation; the real release follows.)
License: Apache-2.0
Project-URL: Homepage, https://github.com/cedricwyh/dsh-rla
Keywords: dsh,deepseek-harness,self-improvement,skills,memory
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Operating System :: OS Independent
Requires-Python: >=3.10
Description-Content-Type: text/markdown

# dsh-rla

**Recursive learning agent —— a DeepSeek Harness plugin, not an agent runtime.**

> This release is a **name reservation** (`0.0.1`). The real plugin ships later;
> the API published today is a placeholder and will be replaced.

What it will do:

- Create skills from a finished session (success or failure), write them to a catalog
- Improve existing skills by fragment-level patches, gated by a **comparison probe** —
  never "the model says it's better"
- Persist lessons only with `cue` / `lesson` / `evidence`, each expiring and auditable
- Retrieve past sessions by semantics (FTS, not embeddings — no surveyed project uses vectors)
- Keep a cross-session model of the user (preferences, landmines, working style) with source + date

Boundary: context-window management is a different project (`deepkeel`).
This one is about what survives **between** sessions.

Repository: https://github.com/cedricwyh/dsh-rla
