dev-looplisted

# Dev Loop Long-running AgentHub work uses a loop: read current governance, choose the next scoped slice, implement, review, verify, record progress, and push. Short one-file fixes do not need this skill. ## Canonical State Read these before acting: 1. `AGENTS.md` 2. `docs/progress/MASTER.md` when it exists 3. `docs/roadmap.md` and affected `docs/architecture/*` 4. The relevant issue or PR Do not create a second roadmap or state file. Active spec-driven work lives in `docs/progress/MASTER.md`; long-term product direction lives in `docs/roadmap.md`. ## Loop 1. Define the narrow task slice, allowed paths, forbidden paths, and evidence gates. 2. Use a worktree only for parallel or risky branch work; small documentation or focused fixes can stay on the current branch. 3. For subagents, pass only the needed files, allowed write scope, forbidden scope, verification commands, and expected report format. 4. Review every subagent result before integrating it. 5. Run the smallest useful gates first, then broader gates when the change affects shared behavior. 6. Update `docs/progress/MASTER.md`, the issue/PR, and affected architecture/roadmap docs in the same pass. 7. Commit and push small verified changes. ## Agent Routing Do not hard-code provider names or local aliases in task cards. Assign work by capability: | Need | Route | |---|---| | Architecture, security, or high-risk tradeoff | strongest reasoning agent available | | Narrow Go/TypeScript implementation | focused cod