v0.0.4 — first public release

LeAgents v0.0.4

Agentic orchestration for the LeRobot pipeline: a deterministic collect → train → eval → improve loop where agents propose and a pure function decides.

Highlights

• M0 verified end-to-end on a real GPU: 3-cycle autonomous run (6.1 GPU-hours, zero intervention) on LIBERO — budgets held, weights carried across cycles, and the escalate_floor guard correctly refused to escalate a 0%-plateau. Full numbers in the README.
• M1: OKF knowledge wiki with per-claim provenance, provider-agnostic LLM adapter (Anthropic / any OpenAI-compatible endpoint / none), DexFlyWheel data path — success-filtered rollout harvesting, accumulated mix merging, adaptation training.
• M2: flow dashboard (leagents dash) — live cycle pipeline, decisions, eval chart, rollout videos, event log, knowledge browser.
• Constitution safety gate: sim-only, real-hardware and CVE-2026-25874 paths denied and audited.
• 65 tests, no GPU or lerobot needed to run them; verified against lerobot 0.5.1 (10 real-environment quirks cataloged in the design docs).

Grounding

The architecture is grounded in an adversarially-verified research pass over 2023–2026 robotics papers (24 claims confirmed, 1 refuted — the refuted one became a design rule). See DESIGN.md.