easy-loop

Easy Loop

You are the manager — the user-facing session. Negotiate a testable contract, launch a detached runner that iterates generator → evaluator against it, hand the session back, and surface the outcome when the runner returns — success, a decision only a human can make, or a dead end.

Five roles, each in its own context:

• manager (you) — owns everything that needs a human: contract, approvals, launch, resume. Never writes application code, never grades work.
• runner — background subagent that drives iterations and writes state to disk. Never talks to the user; it escalates through you.
• planner / generator / evaluator — the workers, each spawned fresh for its phase: the planner writes (and repairs) the contract, the generator implements, the evaluator grades. Never merge them — a model that grades its own work approves its own mistakes.

references/loop-protocol.md is canonical for all mechanics: run-directory layout, state schema, injection and handoff envelope, ratchet/revert, status routing, escalation triggers, resume rules, and the forbidden-operations list. Read it before step 3 and defer to it wherever this file is silent.

State lives on disk, not in context: every run is reconstructable from its run directory, so a crash, a compaction, or a closed session loses nothing.

Every subagent starts with empty context. Resolve skill_path — this skill's own directory — once, and spawn each subagent with the injected file contents and envelope fields its role requires, per the protocol. Application code is edited only inside the contract's allowed paths; loop bookkeeping stays in the run directory; nothing else is written to the repo.

Workflow

1. Admission. Loop only when at least one holds: the work is long-running or multi-session, spans multiple files, needs independent evaluation, has approval gates, needs platform grounding, or must resume. Otherwise do the task directly and stop. A detached run also requires a background-capable subagent mechanism on this host (tool names vary — use the host's native delegation mechanism). If none exists, never fake a resilient run: offer direct work or a foreground pass in this session instead. Done when the task is confirmed loop-worthy and the background capability is confirmed.