dev-loop

Drives an approved single-step plan through implement, test, and review stages until ready to commit.

Updated Apr 28, 2026
One-click install
npx skills add https://github.com/byron1st/personal-harness --skill dev-loop-byron1st
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: dev-loop
Source: https://github.com/byron1st/personal-harness/tree/main/skills/dev-loop
Command: npx skills add https://github.com/byron1st/personal-harness --skill dev-loop-byron1st

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Executing an approved development plan requires coordinating implementers, testers, reviewers, and fixers across multiple rounds with human approval gates. This Skill automates that orchestration as a state machine, checkpointing every stage to disk so work survives session interruptions and never loses track of findings or decisions. ## Core Features & Use Cases - Mode-based execution: Choose light (default, two review axes, no mutation testing), full (four review axes plus mutation testing), or noreview (no reviewers) at preflight; the mode is frozen for the entire run. - State machine with human gates: Drives IMPLEMENTING → TESTING → REVIEWING → FIXING rounds with exactly two human gates (triage of findings and READY_TO_COMMIT), plus escalation on budget exhaustion, blocked stages, or repeated failures. - Crash-safe checkpointing: Appends every stage result to a LOOP file after each return, so a session can resume from disk artifacts alone without relying on conversation memory. - Use Case: After approving a plan written by plan-dev, run the loop in light mode to have an implementer write the code, a tester verify it against the Acceptance Contract, and two reviewers check maintainability, stopping for your decision on any findings before reporting READY_TO_COMMIT. ## Quick Start Run the approved plan at docs/agents/dev/PLAN.md through the dev loop in light mode and stop at ready to commit.

Frequently Asked Questions about dev-loop

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run an approved plan through an automated implement-test-review loop?▼

Pass the plan file path to the dev-loop skill and optionally name a mode. It preflights the plan for an Acceptance Contract and Authority Boundaries, then drives implementer, tester, and reviewer personas through rounds until all termination predicates hold.

What is the difference between light, full, and noreview modes?▼

Light runs two review axes (maintainability and senior-generalist) without mutation testing. Full runs all four review axes plus mutation testing. Noreview skips reviewers entirely, relying on test findings and human review of the IMPL report.

Can I change the loop mode in the middle of a run?▼

No. The mode is frozen at preflight and recorded in the LOOP file. If a new instruction names a different mode than the file records, the loop refuses and asks rather than switching.

What happens if a session crashes during the loop?▼

The loop appends every stage result to a LOOP checkpoint file after each return. On resume, it trusts the file's recorded mode and next state over conversation memory, continuing from the last completed round.

When does the dev loop stop and ask a human?▼

Exactly two gates exist: triage of review or test findings (Fix or Accept decisions) and READY_TO_COMMIT. Escalations like blocked stages, budget exhaustion, or repeated failures abort the loop and report to the human.

Why does the loop refuse to run my plan?▼

Preflight refuses plans missing PlanType single-step, an Acceptance Contract, or Authority Boundaries, and refuses multi-step main plans. Run each STEP-N sub-plan through its own loop, or return to plan-dev to fix the plan.