Skip to content
Back to skills

Acceptance Orchestrator

ASecurity

Use when a coding task should be driven end-to-end from issue intake through implementation, review, deployment, and acceptance verification with minimal human re-intervention.

  • 91 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added May 29, 2026
ai-agentsgogitapisecurity

Works with

  • api

Security analysis

A100/100

Scanned May 29, 2026

npx -y skills add kursku/skills --skill acceptance-orchestrator --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Acceptance Orchestrator?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Acceptance Orchestrator
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/kursku-acceptance-orchestrator/badge)](https://www.skillsdirectory.com/skills/kursku-acceptance-orchestrator)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: acceptance-orchestrator
description: Use when a coding task should be driven end-to-end from issue intake through implementation, review, deployment, and acceptance verification with minimal human re-intervention.
risk: critical
source: community
date_added: "2026-03-12"
---

# Acceptance Orchestrator

## Overview

Orchestrate coding work as a state machine that ends only when acceptance criteria are verified with evidence or the task is explicitly escalated.

Core rule: **do not optimize for "code changed"; optimize for "DoD proven".**

## Required Sub-Skills

- `create-issue-gate`
- `closed-loop-delivery`
- `verification-before-completion`

Optional supporting skills:
- `deploy-dev`
- `pr-watch`
- `pr-review-autopilot`
- `git-ship`

## Inputs

Require these inputs:
- issue id or issue body
- issue status
- acceptance criteria (DoD)
- target environment (`dev` default)

Fixed defaults:
- max iteration rounds = `2`
- PR review polling = `3m -> 6m -> 10m`

## State Machine

- `intake`
- `issue-gated`
- `executing`
- `review-loop`
- `deploy-verify`
- `accepted`
- `escalated`

## Workflow

1. **Intake**
   - Read issue and extract task goal + DoD.

2. **Issue gate**
   - Use `create-issue-gate` logic.
   - If issue is not `ready` or execution gate is not `allowed`, stop immediately.
   - Do not implement anything while issue remains `draft`.

3. **Execute**
   - Hand off to `closed-loop-delivery` for implementation and local verification.

4. **Review loop**
   - If PR feedback is relevant, batch polling windows as:
     - wait `3m`
     - then `6m`
     - then `10m`
   - After the `10m` round, stop waiting and process all visible comments together.

5. **Deploy and runtime verification**
   - If DoD depends on runtime behavior, deploy only to `dev` by default.
   - Verify with real logs/API/Lambda behavior, not assumptions.

6. **Completion gate**
   - Before any claim of completion, require `verification-before-completion`.
   - No success claim without fresh evidence.

## Stop Conditions

Move to `accepted` only when every acceptance criterion has matching evidence.

Move to `escalated` when any of these happen:
- DoD still fails after `2` full rounds
- missing secrets/permissions/external dependency blocks progress
- task needs production action or destructive operation approval
- review instructions conflict and cannot both be satisfied

## Human Gates

Always stop for human confirmation on:
- prod/stage deploys beyond agreed scope
- destructive git/data operations
- billing or security posture changes
- missing user-provided acceptance criteria

## Output Contract

When reporting status, always include:
- `Status`: intake / executing / accepted / escalated
- `Acceptance Criteria`: pass/fail checklist
- `Evidence`: commands, logs, API results, or runtime proof
- `Open Risks`: anything still uncertain
- `Need Human Input`: smallest next decision, if blocked

Do not report "done" unless status is `accepted`.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.