another-harness Repository Snapshot

another-harness is a proving ground for Codex-native harness extensions that compete with the newer Hermes harness layer while preserving a Lean project scaffold as a first-class occupant of the repo.

Main architectural facts

• The repo’s stated target is the harness layer: operator surface, work objects, evaluator loops, automation contracts, resumable artifacts, and eventually coordination patterns.
• It explicitly avoids rebuilding a full personal-memory stack or swallowing the Lean surface with a giant runtime.
• Active work state lives in repo artifacts such as state/work-items/, state/evaluations/, and state/handoffs/.
• Cross-session memory is delegated to the shared Honcho workspace through the repo-local honcho-codex plugin.
• The architecture spec treats evaluation artifacts and reality-bearing checks as prerequisites for completion.

Current operational maturity

The repo already has durable work and evaluation artifacts, but its backlog still frames automation and heavier orchestration as pending or partial. The center of gravity remains work-item discipline, evaluation separation, and Lean-grounded harness semantics rather than RL training loops.

Relevance to Atropos

Nothing in the authoritative repo docs or backlog currently marks Atropos, RL, trajectories, or reward-driven training as an immediate need. That does not make RL substrates irrelevant forever; it means the likely near-term use of an Atropos-style environment would be optional and downstream of the repo’s existing work/evaluation artifact model rather than part of the repo’s core architecture today.