rein-openweights/SCOPE.md

SCOPE

One-liner

Agentic tool-use harness driving current open-weight models via OpenRouter — an alternative to the frontier-vendor harnesses (Claude Code, Grok CLI, Codex).

Core Idea

One agentic loop (plan → tool call → observe → repeat), own minimal OpenRouter client (not llm-connect's OpenRouterAdapter — it can't drive a multi-turn tool-calling loop; see workplans/REIN-OW-WP-0001-bootstrap.md T01), same starter tool surface as rein-aharness's green-commit-only, own OpenBao credential acquisition (glas-harness ADR-002, Option B).

In Scope

• openrouter_client.py: thin chat-completions client, own transport, not routed through llm-connect
• tools.py: read/write/edit/glob/grep + git status/diff/log/add+commit, path-traversal guarded
• loop.py: the agentic loop itself, turn- and budget-bounded
• credentials.py: own OpenBao AppRole/ambient-token acquisition
• runner.py/hub.py: commit-verified success criterion + State Hub reporting, mirroring rein-aharness's model

Out of Scope

• Model routing/selection policy — caller-supplied --model, not owned here
• Monitoring/evaluation/optimization as first-class capabilities — a deliberately deferred question (glas-harness ADR-002 part 2, "composable reins as middleware") rather than something this repo builds ad hoc
• Scheduling/task intake beyond a task-file/caller-driven CLI — no kaizen-agentic/issue-core-style coupling (glas-harness ADR-003)
• Being the glas-harness meta-framework itself — this repo is one concrete rein under it

Current State

REIN-OW-WP-0001 is done (5/5). Live-proven through glas-harness's gateway against the real OpenRouter API (qwen/qwen-2.5-72b-instruct), real tool-calling loop, real verified commit. Attached to its Forgejo remote and registered with State Hub.

Getting Oriented

• Start with: INTENT.md
• Agent instructions: AGENTS.md
• Workplans: workplans/