execute-plan

Executes approved implementation plans phase-by-phase with per-phase commits, independent reviews, and resumable run state.

Updated Apr 6, 2026
One-click install
npx skills add https://github.com/toderian/project_template --skill execute-plan-toderian
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: execute-plan
Source: https://github.com/toderian/project_template/tree/main/plugins/agents-core/skills/execute-plan
Command: npx skills add https://github.com/toderian/project_template --skill execute-plan-toderian

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Approved implementation plans often stall or degrade during execution: phases get skipped, commits mix unrelated work, reviews are missing or not independent, and an interrupted session loses all progress. This Skill turns an approved task or plan file into a disciplined execution run where every phase is implemented, reviewed, committed, and recorded on disk so any later session can resume. ## Core Features & Use Cases - Phase-gated execution with run state: Each phase gets a brief, an implementation pass, independent spec/quality/security reviews, a fix loop capped at three rounds, and its own commit, all tracked in docs/tasks_manager/_runs/<TASK-ID>/state.md. - Small and large modes: Small tasks are implemented in the main thread with one reviewer per phase; large tasks run as a thin orchestrator dispatching fresh implementer and reviewer subagents per phase, with an architecture review up front. - Resumable across sessions and runtimes: The run directory is the source of truth for resume, with runtime-specific guidance for Claude Code, Codex, and inline execution without subagents. - Use Case: You have an approved task file docs/tasks_manager/_todos/AUTH-001.md with three phases. Invoke the skill, and it opens a run state, implements and reviews each phase, commits each one separately, runs a final whole-task review, and lets a fresh session pick up exactly where an interrupted run stopped. ## Quick Start Ask the agent to execute the approved plan in docs/tasks_manager/_todos/AUTH-001.md phase by phase with per-phase commits and independent reviews.

Frequently Asked Questions about execute-plan

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

FAQPage Schema
How do I execute an approved implementation plan with an AI coding agent?

Point the agent at your task file in docs/tasks_manager/_todos or plan file in docs/_plans and invoke the execute-plan skill. It normalizes the plan into phases, implements each phase, runs independent reviews, and commits each phase separately.

How does the skill decide between small mode and large mode execution?

Small mode covers tasks with up to two phases and a scope fence of roughly ten files, implemented in the main thread with one reviewer per phase. Any large trigger, such as three or more phases, a security surface, a data migration, or a public interface, switches the run to large mode with dispatched implementer and reviewer subagents.

Can an interrupted execution run be resumed in a new session?

Yes. The run state in docs/tasks_manager/_runs/<TASK-ID>/state.md is the source of truth for resume. A new session verifies recorded commits, re-reads rulings and interfaces, and continues at the first phase row that is not committed.

Does execute-plan work with both Claude Code and Codex?

Yes. The skill detects the runtime and records it in the run state. Claude Code uses named subagents resumable by id, Codex spawns agents from .codex/agents/*.toml roles without resume, and inline mode runs everything in the main thread with self-reviews labelled not independent.

What happens when a phase fails review repeatedly?

The fix loop is capped at three rounds, with round three using a structural prompt on the strongest model. If the phase still does not converge, the orchestrator adjudicates each open finding with a Ruling line or marks the row blocked and stops to report instead of looping indefinitely.

When should I not use execute-plan for a coding task?

Do not use it for a trivial one-sentence diff, which should just be edited, tested, and committed directly, or for a rough idea or unapproved proposal, which must go through planning or task skills first. Plans lacking phases, acceptance criteria, or checks are normalized before execution.