Skip to content
Back to skills

Sdd Execute

ASecurity

Implement exactly one SDD Composy task in the ready state with dependency preflight, TDD RED evidence, allowed paths, real verification commands, and the guarded transition to qa_required. Use when a ready task should be executed — 'executar a TASK-003', 'implementar a próxima tarefa', 'run task 4'. Do NOT use for changes without an approved task (sdd-quick), for autonomous multi-stage correction (sdd-loop), or for QA, review, or verification.

  • 3 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added September 22, 2026
testing

Security analysis

A100/100

Pro scans all 2 files and shows the line behind each finding

Scanned September 22, 2026

npx -y skills add pwdev-solucoes/pwdev-claude-marketplace --skill sdd-execute --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Sdd Execute?

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

Security grade badge for Sdd Execute
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/pwdev-solucoes-sdd-execute/badge)](https://www.skillsdirectory.com/skills/pwdev-solucoes-sdd-execute)

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: sdd-execute
description: >
  Implement exactly one SDD Composy task in the ready state with dependency
  preflight, TDD RED evidence, allowed paths, real verification commands, and the
  guarded transition to qa_required. Use when a ready task should be executed —
  'executar a TASK-003', 'implementar a próxima tarefa', 'run task 4'. Do NOT use
  for changes without an approved task (sdd-quick), for autonomous multi-stage
  correction (sdd-loop), or for QA, review, or verification.
metadata:
  version: 0.1.0
---

# SDD Execute

This portable skill implements exactly one `ready` task. The bundled `scripts/sdd_execute.py`
functions `preflight`, `run_command`, and `finish` are the deterministic execution boundary;
preserve their result fields and never create test-only bypasses.

Language: before writing human-facing prose, run `scripts/sdd_language.py <repo-root>` and use the persisted language; on `not_initialized`, return it with `next_action: run_init`. Localization rules: `references/language.md`.

## Contract

- Require exactly one `ready` task and a successful dependency preflight.
- Require TDD RED evidence before implementation, enforce the task's allowed paths, and run the
  declared verification commands, each as a real command. Record sanitized stdout/stderr, exit codes,
  SHA-256 evidence, environment ownership, and cleanup.
- Never modify approved contracts or user-owned processes.

On success, record the passing verification run with `scripts/sdd_tasks.py evidence <state>
<TASK-ID> tests --status passed --ref <command-record>` and request the guarded
`running -> qa_required` transition. A concrete dependency, environment, QA, evidence, or cleanup
blocker becomes `blocked`; a scope violation or rejected contract becomes `rejected` (both through
`transition ... --reason <text>`, from `ready` or `running`). Record the exact reason and next action, and stop on any QA or review
blocker.

## Read when

- `references/execution.md` — before the first `run_command` of a task (preflight fields, TDD
  evidence, path and command rules, outcome classification).
- `references/states.md` — a transition other than `running -> qa_required` is needed.
- The approved task contract (`tasks/prd-<slug>/task-<id>.md`) — always, it is the input.

## Output

Return the execution result, evidence manifest, changed paths, ownership and cleanup status, and
the permitted next lifecycle action.

Safety: Do not commit, push, or publish. Do not read or expose `.env`, credentials, tokens, private keys, certificates, or fleet environment files. Full contract: `references/safety.md`.

Files in this skill

  • SKILL.md2.6 KB
  • agents/openai.yaml184 B

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.