A lean, effect-first development Skill suite whose stable identity survives model upgrades.
Version 0.8 introduces the model-neutral agentic-* namespace while retaining the dependency-aware hub-and-spoke structure: one accountable coordinator, one explicit controller-led delivery workflow, and nine focused Skills. A clear task can use one focused Skill directly—or no Skill at all—while long-running work can opt into a durable current-truth → Spec → Plan → Execution → Review → Verification → Acceptance → Closure lifecycle. The full repository installer also adds a concise global purpose-bound-rigor rule and an on-demand read-only audit agent.
The suite evolves through versioned model profiles instead of embedding a model generation in repository, plugin, or Skill identifiers. The current profile was shaped by OpenAI's GPT-5.6 prompting guidance, while the stable contracts remain outcomes, constraints, evidence, autonomy, validation, permissions, and stop rules. See model profiles for the distinction between design inputs, deterministic package checks, and live-model evidence.
This suite keeps the useful invariants—permission boundaries, project grounding, root-cause diagnosis, evidence before claims, preservation of user changes—and removes methodology ritual. It does not impose fail-first development, repeated broad suites, automatic worktrees, or per-task review chains. For repository changes, the Agent automatically owns task-branch selection, atomic commits, branch push, and PR creation or update without stepwise approval requests. The user reviews the PR and decides whether it may merge; the implementer cannot approve its own work. Explicit local-only requests stop before push, while merge, force push, history rewrite, unmerged-branch deletion, visibility changes, and comparable high-impact actions still require explicit authority.
| Skill | Use it for |
|---|---|
agentic-superpowers |
Two or more dependent development phases that need end-to-end coordination |
agentic-orchestrate-delivery |
Explicit standardized delivery from current-state discovery through acceptance and Git closure |
agentic-design-planning |
Consequential ambiguity in product, UX, architecture, interfaces, migrations, or scope |
agentic-writing-plans |
Approved Spec → ordered implementation Plans → execution handoff |
agentic-using-git-worktrees |
Explicit or justified worktree isolation that preserves existing user changes |
agentic-subagent-driven-development |
Implementation-ready plans with independently ownable outcomes under accountable controller ownership |
agentic-debugging |
Ambiguous, intermittent, environment-dependent, or multi-component failures |
agentic-verification |
Choosing proportionate evidence for material completion claims |
agentic-purpose-bound-rigor |
Requiring a concrete necessity case before adding hashes, gates, isolation, mocks, freezes, or repeated reviews |
agentic-delegation-review |
Genuinely independent parallel work or a focused independent review |
agentic-git-delivery |
Feature-branch commits, pushed PR evidence, independent approval, required checks, controlled merge, or cleanup |
All eleven are direct entry points. The orchestrated workflow is opt-in for explicit standardized delivery or long-running multi-module work; it does not force ordinary tasks through a fixed chain. The remaining execution-structure Skills may hand work to another focused Skill only at a real dependency boundary. Narrow trigger descriptions allow implicit routing without an always-on router, and explicit $skill-name invocation remains available.
For repository-changing work, the Agent automatically creates or selects an isolated feature branch from a verified baseline; it never asks the user to create the branch. It protects and safely migrates uncommitted default-branch work, asking only when separation is unsafe. After focused evidence passes, the Agent commits, pushes, and opens or updates a PR recording scope, implementation, evidence, risks, dependencies, and unresolved items. These mechanical steps are preauthorized by the change request. An independent authorized reviewer or the user must approve and required checks must pass before merge. Direct default-branch work requires prior user authorization and a recorded reason. After confirmed merge, cleanup is automatic only when the worktree is clean and no unmerged commits or open dependencies remain.
The comparison baseline is a local 14-Skill installation from obra/superpowers at commit b55764852ac78870e65c6565fb585b6cd8b3c5c9.
| Measure | Baseline | Version 0.8 | Reduction |
|---|---|---|---|
| Workflow Skills | 14 | 11 | 21.4% |
Total SKILL.md words |
15,737 | 5,474 | 65.2% |
| Always-trigger router | Yes | No | Removed |
A focused route loads one body of 233–850 words. The longest Skills carry concrete planning, orchestrated-delivery, worktree, or subagent contracts and load only when those structures materially help. The full 5,474-word package is never a mandatory prompt chain. Word count is a structural proxy, not a model-quality or tokenization guarantee.
Clone the repository and run the installer once:
git clone git@github.com:bells0/agentic-superpowers.git
cd agentic-superpowers
./scripts/install-local.shThe full installer:
- installs all eleven Skills;
- adds a managed purpose-bound-rigor block to the effective global
~/.codex/AGENTS.mdorAGENTS.override.mdwithout replacing existing guidance; - installs
execution-efficiency-auditorunder~/.codex/agents/; - removes only its managed guidance and Agent when
./scripts/restore-original.shrestores the transaction.
Start a new Codex task after installation. Global AGENTS.md guidance is then loaded for every repository; the full Purpose-Bound Rigor Skill expands only when relevant, and the audit Agent remains available on demand.
To install only the eleven Skills without global guidance or the custom Agent:
python3 ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py \
--repo bells0/agentic-superpowers \
--path \
skills/agentic-superpowers \
skills/agentic-orchestrate-delivery \
skills/agentic-design-planning \
skills/agentic-writing-plans \
skills/agentic-using-git-worktrees \
skills/agentic-subagent-driven-development \
skills/agentic-debugging \
skills/agentic-verification \
skills/agentic-purpose-bound-rigor \
skills/agentic-delegation-review \
skills/agentic-git-deliverygit clone git@github.com:bells0/agentic-superpowers.git
cd agentic-superpowers
./scripts/install-local.shThe installer treats all eleven gpt56-* Skill IDs as legacy migration inputs. It moves them into a restorable transaction, replaces the old global guidance marker with the model-neutral marker, and installs only the eleven agentic-* Skill links. It does not leave discoverable aliases that would duplicate routing metadata. See the migration guide.
The same installer also replaces an existing obra/superpowers installation. It validates the package, preserves legacy or conflicting Skill entries, and records the managed global runtime state. Restore the newest READY transaction with:
./scripts/restore-original.shYou can also pass a specific backup directory. Existing gpt56-superpowers version-0.7, version-0.6, version-0.3, and version-0.1 backups remain restorable.
Implicit routing handles strong matches. Invoke a Skill explicitly when you want a specific lens:
Use $agentic-debugging to diagnose this intermittent cross-service failure.
Use $agentic-orchestrate-delivery to run this long-running feature through controller-led discovery, planning, bounded execution, review, verification, acceptance, and closure.
Use $agentic-verification to choose proportionate evidence for this release claim.
Use $agentic-purpose-bound-rigor to decide whether these proposed gates and isolation steps are actually necessary.
Use $agentic-writing-plans to turn this approved Spec into ordered, repository-grounded implementation Plans and define the controller-owned execution handoff.
Use $agentic-superpowers to coordinate this migration end to end.
Use `docs/implementation-closed-loop.md` as a reusable template for long-running Spec → Plan → Execution workflows (including Plan A/B/C style handoffs and model-split recommendations).
Ordinary questions still use Codex directly. Repository-changing outcomes implicitly add Git & Delivery: the Agent handles branch, commit, push, and PR mechanics; the user or an independent reviewer controls approval, and merge remains explicitly authorized. Explicit $skill-name invocation remains available.
The full installer makes the read-only execution-efficiency-auditor Agent available across repositories. Invoke it only when a user asks to audit process bloat or repeated execution drift; it is not a routine reviewer or implementation gate.
The repository uses deterministic package checks, a fourteen-case routing specification, global-runtime tests, and install/restore transaction smoke tests. The scenarios constrain intended behavior; they are not a live model benchmark:
make validateRun the canonical Codex Skill and plugin validators with:
make dev-deps
make local-codex-validateSee architecture, model profiles, migration from gpt56-superpowers, migration from obra/superpowers, and evaluation.
Agentic Superpowers 是一套不绑定具体模型版本的轻量开发 Skill：1 个主控核心 + 1 个显式标准交付流程 + 9 个聚焦 Skill。普通任务不加载，单领域任务只加载一个；长期多模块任务可以显式选择主控编排、独立审查、真实路径验证、用户验收和 Git 收口的完整闭环。
模型迭代记录在独立的 Model Profile 中，不进入仓库、插件或 Skill 的永久命名。升级安装会备份并移除旧的 gpt56-* 调用名，只暴露新的 agentic-*，同时保留旧事务的恢复能力。
本版本彻底移除了开发方法论强制，不要求先写失败测试、不要求 RED/GREEN/REFACTOR、不要求重复跑全量测试。保留的是更薄的“声明—证据”验证：文档看 diff/schema/link，Bug 复查原始症状，行为跑最相关检查，视觉实际渲染，发布遵守项目门禁；无法验证就明确缺口。
仓库开发默认从功能分支开始。每个独立可验收、可独立回滚且验证通过的结果形成原子 commit；完成后推送功能分支，并通过 PR 记录范围、实现、验证证据、风险与依赖。开发者不能审批自己的实现，只有独立审批者或用户批准且必需检查通过后才能合并。禁止直接在 main 或默认分支开发、提交或推送；紧急例外必须由用户明确授权并记录原因。明确要求仅保留本地时，在 push 前停止。
MIT. This is an original rewrite inspired by obra/superpowers; see third-party notices.