Agentic development environment

The development environment for coding agents.

Sona runs your coding agent through one pipeline: spec, plan, implement, audit, review, validate, pull request. You decide what starts and what leaves your machine. Every run is stored locally, so work resumes after a crash, a usage limit or a change of agent.

Install Sona
curl -fsSL …/SonaADE/main/install.sh | bash

Works with the agent CLI you already use

  • Claude Code
  • Codex
  • Cursor
  • Junie

01Pipeline

Every change takes the same path.

The agent writes the code. The runtime sets the order, runs the gates and records what each phase produced. A phase that doesn't meet its contract stops the run and says why.

  1. 01

    Spec

    An issue and its acceptance criteria become a spec bundle you can read before any code changes.

    spec.md
  2. 02

    Plan

    Large work is split into dependency-ordered subtasks. Each runs in a fresh context and ends in one commit.

    spec_subtask_N.md
  3. 03

    Implement

    The agent builds the subtask against the spec, the plan and the decisions recorded by earlier work.

    agent/decisions.md
  4. 04

    Audit

    The result is checked against every acceptance criterion. Gaps go back to the agent for repair.

    audit
  5. 05

    Review

    Specialist reviewers for your platform report findings. Each finding is verified before it is fixed.

    review · verify_findings
  6. 06

    Validate

    Your project's own build, tests, lint and formatting run. Failures are repaired, not skipped.

    validate
  7. 07

    QA

    Declared targets are exercised: endpoints, web journeys, the desktop app or a device.

    qa · qa_fix
  8. 08

    Pull request

    The branch is pushed, the PR opened and CI watched. Failing checks and merge conflicts are repaired.

    pr · monitor

Each phase can also run on its own: review a branch, validate it, QA it or open its PR.

02Control

You start the run. You approve what leaves.

Sona is built around operator decisions. Agents work inside the run; anything that edits, tags or pushes waits for you.

  • Pause, stop or resumeStop after the current subtask or immediately. Resume from the last finished phase.
  • Explicit confirmationProposals show what will change before anything is applied.
  • Review in placeRead each subtask's diff and leave inline comments for the agent to address.
  • A record of every phaseEach phase leaves a named artifact. Instruction files the agent saw are recorded by hash.

03Continuity

Runs outlive the session.

Progress lives in a local database, not a chat window. Close the laptop, hit a limit or switch agents halfway through; the run continues from the last finished phase.

  • Resumable at every boundaryEach phase writes its output before the next one starts.
  • Decisions carry forwardSubtasks record their decisions next to the code. The next plan reads them first.

04Details

Configured as source, not settings.

The parts you'd want to change are plain files in your repository and on your disk.

learnings

Learning from review

Reject a review finding and confirm it, and it becomes a learning. Later reviews apply it.

config.json

Model per phase

Choose the model and effort for each phase. Plan with one, implement with another.

phases/review.md

Phase instructions

Add a markdown file to give one phase your team's rules and conventions.

platform-packs/

Platform packs

Go, iOS, Kotlin, KMP, PHP, Python, Rust and TypeScript. Packs declare routing, review areas, checks and QA.

skills/

Your own skills

Add-ons and custom skills are written as plain source files, versioned with everything else.

telemetry

Telemetry you can audit

Anonymous, full or off, set at install. Every field is documented and the relay can be self-hosted.

05Local-first

Your machine, your agent, your credentials.

The runtime, your specs and every run's history stay on your disk. Sona uses the agent CLI you have installed and the repository you already work in.

No accountNothing to sign up for
No backendNothing hosted
Your credentialsYour agent CLI login
~/.skill-bill
~/.skill-bill/
├── runtime/
├── projects/<project-id>/
│   ├── specs/ORD-142-csv-export/
│   │   ├── spec.md
│   │   └── spec_subtask_1_streamed-writer.md
│   └── phases/review.md    # your rules for one phase
└── review-metrics.db       # every run, every phase

06Install

Install the runtime and the desktop app.

One command installs the runtime, the desktop app and the skill for each agent you choose. Binaries are checksum-verified. No JDK or build tools required.

$ curl -fsSL https://raw.githubusercontent.com/Sermilion/SonaADE/main/install.sh | bash
macOS, Linux and Windows Pre-1.0 Read install.sh before running it