# Spec Execute Next

> Picks the highest-value approved spec, drives it to an outcome-labeled pull request, and resets the workspace. Autonomous mode only.

- Skill: `ainova-systems/spec-execute-next` (Agent Skill)
- Install (CLI): `npx skillmds@latest add ainova-systems/spec-execute-next`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ainova-systems/spec-execute-next/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: ainova-systems (https://skillmd.com/u/ainova-systems)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/ainova-systems/spec-execute-next

---


# Execute Next Best Task

Orchestrator-of-orchestrators: select ONE ready item, delegate the whole execution to `spec-execute` / `spec-continue` (their rules apply verbatim - this skill adds only selection and cleanup), then reset the workspace. Designed for autonomous and scheduled runs - this is `execution_mode: autonomous` machinery; its queue is filled by `spec-approve`, and a supervised project does not schedule it.

## Steps

1. **Preflight.** `git status --porcelain` clean (dirty - halt and surface; never sweep foreign work into a task). `git switch <base> && git pull && git fetch --prune`.
2. **Inventory.**
   - Candidates: `approved` specs in the project's spec area (profile `specs_dir`; else detect) - the primary pool - plus `in-progress` specs with no open PR (to finish; finishing beats starting). Never `proposed` (unreviewed - the owner approves via `spec-approve` first), `completed`, `cancelled`, or `superseded`. Skip idea-collection notes without a concrete ticket-sized item.
   - In-flight work: `gh pr list --state open --json number,title,headRefName,labels,updatedAt`; per-PR file scope via `gh pr view <n> --json files --jq '.files[].path'`.
3. **Select ONE.**
   - Value order: explicit priority marker - mid-flight spec with no open PR (finishing beats starting) - unblocks queued work - smaller effort at equal value. A focus-hint argument narrows the pool, never overrides the conflict gate.
   - Conflict gate, skip the candidate if: it already has an open PR (`ai:processing` means another run holds it; `ai:manual` / `ai:failed` need the owner, not a re-run); its file scope overlaps any open PR's files; it needs a decision the spec leaves open (that is `ai:manual` before even starting).
   - Interactive session - confirm the choice plus runner-up in one question. Autonomous run - proceed. No eligible candidate - report per-candidate skip reasons and stop (skip step 5's sync, still end clean).
4. **Execute until outcome.** Mid-flight spec - `spec-continue`; otherwise `spec-execute`. Run until the PR carries an outcome label (`ai:completed` with every declared factor, or `ai:manual` / `ai:failed`). Never merge.
5. **Reset (runs regardless of outcome).** `git switch <base> && git pull && git fetch --prune`. Close any spec whose PR merged since the last run (`spec-close`) so the queue and docs reflect reality. Re-render the project's generated intelligence outputs when it has an engine that produces them, using the command that engine documents; skip when it has none. Report: task picked + why; PR URL + outcome label; what needs the owner; runner-up for the next run.

## Verify

- Exactly one task progressed to an outcome label; workspace back on a clean, current base.

## Scope / hand-off

- All execution rules - `spec-execute`; the merge - `git-merge-pr` after owner accept. The approved queue this skill drains is filled by `spec-approve` (the owner's gate-1 decision).

## Constraints
- ONE task per invocation; the next starts only after reset and fresh context.
- The conflict gate is non-negotiable - when in doubt about overlap, pick the next candidate.
- Never block other agents' open PRs.