Star on GitHub

Autonomous Coding Agent Framework

AI coding agents need hard reflexes, not hope.

Prompt packs tell agents to be careful. They hallucinate packages, wipe databases, and declare tickets complete while tests fail. jev-superpowers drops 6 disciplined engineering skills wired to TypeSafe Jev System One reflex gates.

Install Framework Inspect Skills

Zero framework fee · 70ms sub-second verification · 22/22 offline test suite

Native Support: Claude Code Cursor Codex OpenCode Devin Hermes Gemini CLI Pi Agent
jev-superpowers lifecycle gate
Simulate:
$ claude-code "Install auth crate and purge cache"
[hook:pretool] intercepting: cargo add jwt-auth-helper
[hook:pretool] running: jev-scout "jwt-auth-helper"
[FAIL] crate 'jwt-auth-helper' does not exist on crates.io (hallucination)
[hook:pretool] candidate scored: jsonwebtoken (score: 9.6, downloads: 42M)
[BLOCKED] tool call cancelled. Substitution injected.
$ cargo add jsonwebtoken
[hook:pretool] verified real package · 82ms · allowed
PHYSICAL EXECUTION GATE · NOT A PROMPT SUGGESTION

The Golden Loop

Sub-second reflex gates across every phase of delivery.

Text models take 3 to 10 seconds to generate prose explanations. TypeSafe Jev evaluates structured state in 70–120ms with calibrated confidence.

PHASE 1
Router
Choice · 78ms

Selects optimal workflow before writing code.

PHASE 2
Brainstorm
jev-axi pick

Ranks architectural approaches by maintenance cost.

PHASE 3
Scout Plan
jev-scout

Vets cargo crates and npm dependencies in 80ms.

PHASE 4
Command Guard
jev-guard

Halts dangerous rm -rf, drop table, and mass edits.

PHASE 5
Triage Log
jev-axi triage

Filters 2000-line build failures down to root cause.

PHASE 6
Stop Hook
limpet gate

Rejects turn completion if test coverage regresses.

Skill Matrix

6 disciplined skills built on TypeSafe System One.

Each skill replaces loose agent conversational chatter with machine-evaluated gates.

jev-using-superpowers Router

Universal agent entry gate. Evaluates request complexity and mandates systematic discipline before touching disk.

Trigger: Session start & multi-file tasks
Primitive: Choice (router-workflow)
Latency: 74ms
jev-brainstorming Ideate

Spec drafting and architecture tradeoff scoring. Scores 3 concrete approaches against maintainability and YAGNI.

Trigger: Ambiguous features or new APIs
Primitive: Choice via jev-axi pick
Latency: 92ms
jev-writing-plans Scout

Produces bite-sized implementation specs with pre-vetted dependencies. Zero hallucinated crates permitted.

Trigger: Plan creation (>2 steps)
Primitive: Score via jev-scout
Latency: 84ms
jev-executing-plans Guard

Drives step-by-step TDD execution with shell safety guardrails and git-jev pre-commit verification gates.

Trigger: Plan implementation
Primitive: Choice via jev-guard & git-jev
Latency: 88ms
jev-systematic-debugging Triage

4-phase root-cause diagnosis. Filters massive terminal outputs into immediate hypotheses without context rot.

Trigger: Failing tests or compiler errors
Primitive: Score via jev-axi triage
Latency: 104ms
jev-verification Stop Gate

Stop hook policy enforcement. Runs supercov quality checks and limpet gates to halt premature turn completions.

Trigger: Turn completion & stop events
Primitive: Noul via limpet check
Latency: 79ms

Architecture Comparison

Why system prompt instructions fail coding agents.

Language models treat natural language prompt rules as soft suggestions that erode as conversation history grows.

jev-superpowers (Physical Reflex Gates)

OS-level lifecycle hooks coupled with TypeSafe System One sub-second typed decision primitives.

// Deterministic Gate Execution PreTool Hook: jev-scout checks registry in 80ms Execution: Real packages verified; hallucinations blocked Commit Gate: git-jev blocks dirty diffs before git commit Stop Gate: limpet stops turn completion until tests pass Result: 100% deterministic invariants enforced.
Hallucination Risk 0% (Scouted)
Command Safety Physical Gate
Stop Gate Enforced (limpet)
Decision Speed 70-120ms (Sub-second)

Quickstart

One-line installation across your development stack.

Symlinks skills into `~/.agents/skills/` and wires hook interceptors in seconds.

# Clone and run cross-platform bootstrap
git clone https://github.com/AkashPriyadarshii/jev-superpowers.git ~/.agents/jev-superpowers
cd ~/.agents/jev-superpowers

# Linux / macOS
bash install.sh

# Windows (PowerShell)
pwsh ./install.ps1
  1. Configure your decision backend
    Cloud: export TYPESAFE_API_KEY="sk_live_...". Local FOSS: export TYPESAFE_BASE_URL="http://127.0.0.1:8000" (see docs/FOSS_LAYA.md).
  2. Verify with offline test suite
    Run bash scripts/test.sh or pwsh scripts/test.ps1. Confirms 22/22 YAML frontmatters, hook binaries, and confidence policies pass.
  3. Launch your coding agent harness
    Start Claude Code (claude), Cursor, Codex, OpenCode, or Pi. The router automatically intercepts task starts and stops.

Measured Telemetry

Deterministic performance benchmarks.

70ms System One Gate Latency
22/22 Offline Integrity Checks
0% Hallucinated Package Invocations
$0.0003 Average Cost per Decision
FRAMEWORK ARCHITECT

Akash Priyadarshi

Patna, Bihar, India

Software engineer focusing on autonomous AI agent reliability, sub-second decision oracles, and deterministic developer tooling. Creator of jev-superpowers, jev-seo, jev-curate, and tdlib-android.

GitHub Portfolio LinkedIn Resume