AI agents say tasks are done.

DoneSpec verifies they actually are.

The deterministic completion layer for AI coding agents.

Current release: v1.0.0.

DoneSpec turns "done" into a local, repeatable, machine-checkable contract.

The Problem

AI coding agents can claim completion while tests fail, required files are missing, forbidden paths changed, or documentation was skipped.

DoneSpec makes completion explicit: a task is not done until done.json passes.

Done means deterministically verified.

Demo

The core demo is a forbidden-file scenario:

1. An agent modifies a file protected by done.json.
2. donespec validate done.json fails.
3. The issue is fixed.
4. DoneSpec passes.

Run it locally:

./scripts/demo-forbidden-file.sh

Windows PowerShell:

.\scripts\demo-forbidden-file.ps1

See:

• docs/demo.md
• docs/demo-transcript.md
• docs/demo-recording.md

Install

pip install donespec

Check the installed version:

donespec --version

PyPI:

https://pypi.org/project/donespec/

Upcoming v1.0.0 release tag:

https://github.com/xryv/DoneSpec/releases/tag/v1.0.0

60-second quickstart

Create a completion contract:

donespec init --yes

Validate it:

donespec validate done.json

Use strict mode for serious agent workflows:

donespec validate done.json --strict

Inspect the contract without executing checks:

donespec explain done.json

Add a check safely:

donespec add-check done.json --type file_exists --name "README exists" --path README.md

Example done.json

{
  "$schema": "done.schema.json",
  "version": "1.0",
  "task_id": "ship-safe-change",
  "must_pass": [
    {
      "type": "file_exists",
      "name": "README exists",
      "path": "README.md"
    },
    {
      "type": "command",
      "name": "tests pass",
      "run": "pytest -q"
    },
    {
      "type": "regex_in_file",
      "name": "README explains DoneSpec",
      "path": "README.md",
      "pattern": "deterministic completion layer"
    }
  ],
  "must_not": [
    {
      "type": "file_not_modified",
      "name": "lockfile untouched",
      "path": "uv.lock"
    }
  ]
}

The contract is plain JSON. Agents, humans, hooks, and CI can all read the same definition of completion.

Example CLI Output

Passing:

DoneSpec validation: ship-safe-change

+ README exists  (0.4ms)
+ tests pass  (812.4ms)
+ README explains DoneSpec  (0.6ms)

Validation passed. 3 checks passed.
Exit code: 0

Failing:

DoneSpec validation: ship-safe-change

+ README exists  (0.4ms)
x lockfile untouched  (31.5ms)
  Forbidden path modified: uv.lock

Validation failed.
1 check failed.
Exit code: 1

DoneSpec uses Unicode status symbols when the output stream supports them and ASCII fallback when it does not.

GitHub Actions

Use DoneSpec as a completion gate in CI:

name: DoneSpec

on:
  pull_request:
  push:
    branches:
      - main

jobs:
  donespec:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - uses: xryv/DoneSpec@v1.0.0

      - name: Validate completion contract
        run: donespec validate done.json --strict

The same contract runs locally, in git hooks, in CI, and inside agent workflows.

Why deterministic validation matters

DoneSpec validates explicit contracts, not confidence.

It does not replace tests, CI, or human review. It gives them a small deterministic interface:

agent claim -> done.json -> deterministic checks -> exit code

That means:

• completion requirements are inspectable
• failures are machine-readable and repeatable
• CI can enforce the same contract agents see locally
• reviewers can inspect what "done" actually means

AI Agent Integration

DoneSpec works with coding agents because it has no agent runtime, API key, SaaS service, memory layer, or LLM dependency.

A useful agent instruction is:

Before claiming completion, run:

donespec validate done.json --strict

If validation fails, the task is not complete.
Do not weaken done.json to make validation pass unless explicitly requested by the human.

Useful with:

• Codex
• Claude Code
• Cursor
• Aider
• OpenAI Agents SDK
• GitHub Actions
• local shell workflows
• multi-agent coding workflows

See:

• docs/integrations.md
• docs/agent-integration.md
• examples/integrations/

Core Philosophy

DoneSpec stays small on purpose:

• local-first
• no cloud
• no LLM dependency
• no orchestration
• no database
• no hidden network calls
• deterministic exit codes

It is infrastructure, not a platform.

Architecture Overview

CLI
  reads done.json
  validates JSON Schema
  optionally applies strict contract hygiene
  runs deterministic checks
  emits human or JSON output
  exits 0, 1, or 2

The main pieces are:

• donespec CLI
• done.json completion contract
• deterministic checks
• packaged JSON Schema
• CI, git hooks, and agent instructions that call the same command

Supported checks

DoneSpec is intentionally small. The goal is not to model every workflow; it is to provide a deterministic completion contract that composes with existing tools.

Strict Mode

Strict mode validates contract hygiene before executing checks:

donespec validate done.json --strict

It catches:

• empty contracts
• unnamed checks
• duplicate check names
• duplicate check IDs
• invalid regex patterns
• absolute paths
• parent traversal paths

See docs/strict.md.

CLI Commands

donespec explain done.json

Explain a contract without executing it. Supports --json. See docs/explain.md.

donespec doctor

Inspect whether a project is DoneSpec-ready. Supports --json. See docs/doctor.md.

donespec schema

Print or export the packaged JSON Schema. Schema usage and v1 stability are documented in docs/schema.md.

donespec templates

List starter contract templates. See docs/templates.md.

donespec add-check

Safely add a check and validate the updated contract before writing it. See docs/authoring.md.

donespec init

Create DoneSpec files in a project. See docs/init.md.

Editor support is documented in docs/editor-support.md.

CLI output principles are documented in docs/cli-ux.md.

Verification

DoneSpec validates itself with the same contract model it asks agents to use.

The repository is checked through done.json self-validation, strict DoneSpec validation, cross-platform GitHub Actions, package build verification, and a local wheel smoke test. Normal validation remains local-first and does not depend on hidden network calls.

See docs/cross-platform-verification.md, docs/package-smoke-test.md, docs/v1-release-candidate-audit.md, and docs/v1-release-runbook.md.

Reliability And Field Notes

DoneSpec's own failures become permanent checks.

The first real Codex + VS Code pass found practical issues around release metadata, git hooks, line endings, and Windows output encoding. Those failures were turned into tests and DoneSpec checks.

See docs/field-notes/first-codex-vscode-experience.md.

Cross-platform verification is documented in docs/cross-platform-verification.md.

Package smoke tests are documented in docs/package-smoke-test.md.

Package build and local wheel verification also run in CI.

v1.0 Readiness

DoneSpec v1.0 is a stabilization milestone, not a platform expansion.

See:

• docs/v1-readiness.md
• docs/v1-freeze-report.md
• docs/v1-github-actions-verification.md
• docs/v1-release-candidate-audit.md
• docs/v1-release-runbook.md
• docs/launch/README.md
• docs/oss-launch.md
• docs/release-checklist.md
• docs/releases/v1.0.0.md

Roadmap to v1.0

The core is already capable. The path to v1.0 is stabilization:

• README polish
• CLI wording refinement
• schema freeze
• documentation hardening
• cross-platform verification
• release checklist
• changelog
• final GitHub and PyPI release
• OSS launch assets

DoneSpec should remain tiny, local-first, deterministic, composable, and infrastructure-focused.

Contributing

See CONTRIBUTING.md.

Before opening a change, run:

python -m ruff check .
python -m ruff format --check .
python -m pytest -q
donespec validate done.json
donespec validate done.json --strict

Contributions should protect the minimal core: determinism, clear CLI behavior, stable schema, cross-platform behavior, and documentation quality.

Avoid changes that turn DoneSpec into an orchestration platform, AI framework, SaaS service, agent runtime, workflow engine, plugin marketplace, or memory layer.

v1.0 release status

DoneSpec v1.0.0 is the first stable public release.

See:

• docs/v1-release-completion.md