# Subagent Execution > Use when a primary agent is executing an approved implementation plan with available delegated implementers or needs evidence-based recovery from stalled delegated work. - Skill: `coltonhyer/subagent-execution` (Agent Skill, multi-file: 38 files) - Install (CLI): `npx skillmds@latest add coltonhyer/subagent-execution` - Raw SKILL.md: https://api.skillmd.com/api/skills/coltonhyer/subagent-execution/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: coltonhyer (https://skillmd.com/u/coltonhyer) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/coltonhyer/subagent-execution --- # Subagent Execution The supervisor coordinates one approved task at a time. It does not implement, review its own task work, publish, merge, or turn a report into completion. Before coordination, confirm the preserved approved plan, authoritative spec, and execution authorization. Read [the execution state](../implementing-plans/references/execution-state.md), [implementation report contract](../implementing-plans/references/implementation-report.md), [implementer brief](references/implementer-brief.md), and, when diagnosis is justified, [diagnostic assignment](references/diagnostic-assignment.md). At review and verification gates, load [reviewing-implementation](../reviewing-implementation/SKILL.md) and [verifying-changes](../verifying-changes/SKILL.md), respectively. ## Dispatch Before creating execution scratch, add its exact directory to the local VCS exclusion. Establish the starting revision and separate pre-existing working-copy changes, including any pending Library preservation. That database change remains outside task ranges and the final recording gate. On resumption, reconcile every recorded revision, report, evidence location, and finding with repository state; do not repeat a completed task solely because session context was lost. When delegation is available and inline work was not requested, dispatch one fresh implementer with only the assigned task, exact applicable plan/spec excerpts, required interfaces and earlier decisions, workspace, report path, and the implementing-plans instructions. Do not require a new worker worktree unless repository instructions require it. Keep unrelated history, the whole plan, and supervisor state out of the worker context. Read the returned report and inspect the full recorded task range and cited evidence. Then dispatch `reviewing-implementation` for that exact range. Read the actual returned review before assigning findings. Send required corrections and fix verification back to an implementer; do not edit the task yourself. Only move to the next task after current verification and every material finding or verification gap is resolved. Record evidence-based dispositions in execution state; a report or review alone never establishes completion. ## Recovery Supply missing task context directly before diagnosing. A repeated unresolved defect, contradictory evidence, or plan/code conflict can justify a fresh diagnostic assignment. If an applicable Debugging skill is in the active inventory, include it; otherwise the diagnostic agent uses judgment under the same assignment. Do not add a fallback debugging procedure. Without subagent capability, the inline agent performs the diagnostic assignment and discloses that no fresh investigator was available. That exception does not turn diagnosis into implementation or authorize a scope change. Use supported causes, uncertainty, conflicts, and the proposed correction/check to steer the current implementer, replace it, or raise a consequential decision. An inconclusive diagnosis does not authorize a blind retry or a scope change. ## Finish After all tasks, arrange whole-change review and verification against the approved spec. Readiness requires covered approved requirements, resolved material findings, final-applicable evidence, recorded intended changes, and no unresolved conflict. Before cleanup, incorporate every approved material plan/spec decision into its preserved authoritative artifact and retain any lasting diagnostic or report outside scratch. Hand off the plan/spec, start and resulting ranges, evidence, review outcomes and dispositions, limitations, and any separately pending preservation change. Remove only this execution's scratch after those preservation and handoff gates succeed; retain it for incomplete, blocked, or awaiting-decision work.