A Plan-Do-Check loop plugin for OpenCode.
Three agents — Planner, Doer, Checker — iterate until the Checker passes the work.
open-looper provides an automated development cycle:
- Plan — Explore codebase and produce an actionable plan
- Do — Write code and tests (TDD red→green)
- Check — Review work and issue PASS/FAIL verdict
- Repeat — On FAIL, iterate with feedback until PASS
Uses isolated git clones for sandboxed development and produces conventional commits with full audit trail.
# Agents are in .opencode/agents/
# Skills are in .opencode/skills/looper/
# Scripts are in .opencode/skills/looper/scripts//looper add authentication to the login endpoint
The skill triggers on /looper <task> and orchestrates the PDC loop.
.opencode/
├── agents/ # PDC loop agents
│ ├── planner.md # Creates execution plans
│ ├── doer.md # Implements via TDD
│ ├── checker.md # Issues PASS/FAIL verdict
│ └── */ # Helper agents (debugger, simplifier, etc.)
└── scripts/ # Helper scripts (setup-clone, git-commit-loop, etc.)
| Agent | Mode | Purpose |
|---|---|---|
planner |
primary | Explores codebase, produces actionable plan |
doer |
subagent | Implements via TDD (red→green→simplify) |
checker |
subagent | Reviews and issues verdict |
debugger |
subagent | Root-cause analysis for stuck iterations |
simplifier |
subagent | Refines implementation |
check-* |
subagent | Specialized review (build, tests, code, runtime, adversarial) |
plan-* |
subagent | Plan review (feasibility, completeness, scope) |
- Clone isolation — Each task runs in
.clones/<task>/with only the work branch (no access to master) - TDD enforcement — Red phase (tests) must commit before green (implementation)
- Resume support — Pick up interrupted loops from last iteration
- Conventional commits — Full audit trail with
Loop-Phase:,Loop-Iteration:,Loop-Verdict:trailers - Parallel exploration — Agents spawn subagents for concurrent analysis
- Git repository
ghCLI authenticated for PR creation (optional)- Docker/docker-compose for integration tests (optional)
MIT