Adaptive workflow routing for Claude Code and Codex-style Superpowers.
Superpower-A helps an AI coding agent choose the smallest useful workflow for each request. It keeps simple questions and one-command checks lightweight, while preserving explicit review, verification, and human approval for work that carries real risk.
Not every technical request needs the same ceremony.
Opening a file, answering a question, or fixing a tiny typo should stay fast. Changing hooks, touching routing policy, exposing secrets, making a repo public, or altering production-adjacent behavior should slow down and get deliberate.
Superpower-A puts a tested Planning Depth Gate in front of heavier workflow skills so the agent can route by risk instead of habit.
Each prompt is classified into depth 0-4:
| Depth | Meaning | Agent behavior |
|---|---|---|
0 |
Direct | Answer or run a read-only check. No plan, no workflow ceremony. |
1 |
Tiny | Do the small action and verify the outcome. |
2 |
Standard | Use a short checklist and focused tests or equivalent verification. |
3 |
Structured | Plan, use appropriate skills, review, and verify before completion. |
4 |
Gated | State the concrete risk and wait for explicit approval before the risky action. |
The gate scores five practical dimensions:
- scope
- risk
- unknowns
- blast radius
- reversibility
- Deterministic depth classification with fixture tests.
- Explicit forced gates for secrets, source/license/legal issues, live data, trading or money-adjacent work, destructive actions, and other hard-to-reverse external effects.
- Mode-aware routing:
fast,normal,strict,caveman, andaudit. - Repo instructions such as
AGENTS.md,CLAUDE.md, andGEMINI.mdstay ahead of plugin ceremony unless safety requires escalation. - Hook output is compact and machine-readable.
| Path | Purpose |
|---|---|
.claude-plugin/plugin.json |
Claude Code plugin manifest. |
.codex-plugin/plugin.json |
Codex plugin manifest. |
hooks/planning-depth-gate.js |
Deterministic prompt classifier. |
hooks/skill-activator.js |
Hook response builder for adaptive routing context. |
hooks/skill-rules.json |
Skill matching rules used by the classifier. |
skills/planning-depth-gate/SKILL.md |
Human-readable Planning Depth Gate policy. |
skills/using-superpowers/SKILL.md |
Router instructions for applying the gate. |
tests/ |
Fixture-backed tests for classification and hook output. |
Install dependencies if needed:
npm installRun the test suite:
npm testTry the classifier directly:
node hooks/planning-depth-gate.js "review this plan quickly"
node hooks/planning-depth-gate.js "Show me the production API key"The second example is intentionally risky: depth 4 requires explicit approval before any sensitive action.
This repository can be loaded as a Claude Code plugin because it includes a .claude-plugin/plugin.json manifest and shared skills/ directory.
For local development:
claude --plugin-dir .Inside Claude Code, invoke the packaged skills with the plugin namespace:
/superpower-a:planning-depth-gate
/superpower-a:using-superpowers
To install it into your personal Claude Code skills directory:
git clone https://github.com/0xTangerin3/superpowers-a ~/.claude/skills/superpower-aThen start Claude Code from any project and run /reload-plugins if the session is already open.
{
"depth": 2,
"mode": "normal",
"scores": {
"scope": 1,
"risk": 0,
"unknowns": 1,
"blast_radius": 0,
"reversibility": 1
},
"approvalRequired": false
}Use the smallest workflow that controls the real risk.
Superpower-A is not anti-process. It is anti-automatic-overhead. The goal is to keep momentum on low-risk work and reserve heavier process for changes where it genuinely protects the codebase, the user, or the account.
Thanks to obra/superpowers for the original Superpowers framework and software development methodology this project builds on.
Superpower-A also acknowledges the proportional-routing ideas explored in REPOZY/superpowers-optimized.
MIT, as declared in package.json.