Workflow before code. Evidence before trust.

Make AI-assisted
delivery predictable

AI Workflow is an npm CLI that installs agents, commands, skills, policies and templates into your repository. It guides coding agents from request to implementation, validation and evidence.

For developers and engineering teams, AIWK adds scoped delivery, branch safety, observed validation, and release conformance around the coding runtime you already use.

Install AI Workflow Kit View npm package
  • ✓Scoped delivery
  • ✓Observed validation
  • ✓Release conformance
OpenCode PrimaryAntigravity SupportedCodex Supported with limitationsClaude Code Supported with limitations
The delivery problem

Coding agents need delivery controls around them

Fast output is not the same as a controlled delivery. Without an explicit path, teams have to reconstruct scope, branch state, validation, and evidence after the fact.

⊘

Uncontrolled diffs

A request can turn into more change than the team intended to review.

↯

Wrong-branch work

Without a branch gate, write-mode work can begin on a protected branch.

⊗

False success

A completion message can arrive without observed results from the required checks.

⌁

Rebuilt context

Decisions, ownership, and task constraints get reconstructed across sessions.

↻

Unbounded remediation

A failed check can trigger changes that quietly expand the approved scope.

What AIWK changes

A more controlled path from request to handoff

AI Workflow Kit adds a workflow layer around coding agents. The value is in the boundaries it makes visible and the evidence it expects—not in the size of its internal inventory.

⌁

Scope the change

Request classification and proportional policies make the intended work easier to define and review.

⌂

Protect the branch

The Branch Gate blocks write-mode work on protected branches and establishes a safe path when recovery is allowed.

◉

Make ownership explicit

A workflow owner routes delivery work while capability specialists contribute bounded analysis.

✓

Observe validation

A successful delivery needs real command results; an unobserved check cannot become a success claim.

↻

Bound remediation

Concrete validation findings can be repaired within a configured attempt limit instead of opening an uncontrolled loop.

▤

Protect consumer updates

Receipt-aware updates reconcile untouched AIWK assets while preserving consumer-owned or modified files.

AIWK does not replace your coding agent. It installs managed workflow assets into the repository and keeps runtime-specific integration boundaries explicit.

Canonical AIWK execution flow

Canonical AI Workflow Kit execution flow

1
RequestCapture the task and constraints
2
ClassificationChoose a proportional policy
3
Branch GateKeep writes off protected branches
4
Workflow OwnerRoute and own the delivery
5
ImplementationDeliver the scoped change
6
Observed ValidationUse real command results
7
Evidence + HandoffReturn outcome and limitations
Workflow contractrequest → evidence → handoffnot a live run
AIWK capabilities

Built as a coordinated workflow system

AIWK is not a single prompt. It combines explicit workflow ownership, bounded specialist analysis, reusable domain skills, and repeatable commands.

6
Primary agents
20
Specialist subagents
25
Reusable skills
14
Workflow commands

Verified against the installed AIWK package. These counts describe the capability surface, not product quality or behavioral parity.

Primary agents

Ownership stays explicit

6 roles

Primary agents own workflow stages and handoffs. They coordinate delivery responsibility rather than acting as interchangeable personalities.

  • AtlasOrchestration
  • NexusDiscovery & specification
  • OrionPlanning & release
  • AstraImplementation
  • SageValidation
  • PhoenixBounded remediation
Specialist subagents

Bounded domain analysis

20 roles

Specialists provide focused analysis when needed and return their findings to the active workflow owner. They do not receive workspace ownership.

  • Architecture
  • Security
  • Frontend
  • Backend
  • Database
  • QA
  • Performance
  • DevOps
  • Documentation
  • Product / discovery
View all 20 specialist roles
  • Architecture-Specialist
  • Backend-Engineer
  • Deployment-Specialist
  • Design-Specialist
  • Docs-Engineer
  • Frontend-Engineer
  • Frontend-Design-System-Specialist
  • Full-Stack-Engineer
  • Token-Economist
  • PR-Manager
  • Discovery-Analyst
  • Product-Planner
  • Memory-Guardian
  • Prompt-Engineer
  • QA-Engineer
  • Refactoring-Specialist
  • Release-Specialist
  • SDD-Specialist
  • Technical-Leader
  • UI-UX-Engineer
Reusable skills

Guidance when relevant

25 skills

Skills load focused technical guidance only when relevant to the current task, keeping domain context proportional to the work.

  • frontend-development
  • qa-workflow
  • release-workflow
  • documentation
  • performance
View all 25 reusable skills
  • architecture
  • backend-development
  • cyber-security
  • database
  • deployment
  • design-principles
  • devops
  • documentation
  • frontend-design-system
  • frontend-development
  • full-stack-development
  • localization
  • optimize-tokens
  • performance
  • pr-workflow
  • product-discovery
  • product-planning
  • project-memory
  • prompt-engineer
  • qa-workflow
  • refactoring
  • release-workflow
  • spec-driven-development
  • technical-leadership
  • ui-ux-design
Workflow commands

Repeatable entry points

14 commands

Commands provide repeatable entry points for planning, implementation, validation, evidence handoff, and workflow execution.

  • /plan
  • /implement
  • /validate
  • /release
  • /run
View all 14 workflow commands
  • /atlas
  • /audit
  • /deploy
  • /discover
  • /implement
  • /optimize-tokens
  • /plan
  • /release
  • /run
  • /spec-create
  • /spec-implement
  • /spec-review
  • /update-memory
  • /validate
Release conformance

A release is more than a green test run

AI Workflow Kit 2.13 adds executable release conformance and a strict release attestation. The qualified source, exact consumer tarball, observed validation, runtime result, and release identity are linked before publication.

Release chain

Evidence follows the exact candidate

Conceptual contract
1Source
2Exact candidate
3Conformance
4Real runtime qualification
5Release attestation
6Publication

The attestation binds the source commit, version, tag, package identity and SHA-256, observed commands, conformance results, and real OpenCode qualification. Missing or skipped required evidence blocks publication.

Conformance contract

Control boundaries that must pass

Each scenario must PASS
  • Protected branchWrite-mode work cannot bypass configured protected branches.Required
  • Read-only confinementRead-only work cannot mutate consumer files or commit changes.Required
  • Workflow ownershipDelegation must preserve the confirmed workflow owner and handoff.Required
  • Observed validationA failed or unobserved command cannot become a success claim.Required
  • Bounded remediationRecovery stays within the configured remediation budget.Required
  • Approval boundariesHigh-risk work stops at the required approval checkpoint.Required
  • Consumer lifecycle safetyExact packed-consumer installation and lifecycle checks remain valid.Required

This is the release contract, not a live release record. No fabricated hash, tag, or publication result is shown on this page.

Runtime compatibility

Use the integration that matches your runtime

AIWK is OpenCode-first and provides documented integration paths for Antigravity, Codex, and Claude Code. The level describes the evidence and boundary—not behavioral parity.

OpenCode

Primary
Integration
opencode.jsonc plus native agents, commands, and skills.
Material limitation
Direct-agent behavior depends on a validated model; use ai-workflow execute for CLI-enforced gates.

Antigravity

Supported
Integration
ANTIGRAVITY.md plus .agents/agents/, .agents/skills/, and .agents/commands/.
Material limitation
Agents are listed under /agents only while actively running through invoke_subagent.

Codex

Supported with limitations
Integration
AGENTS.md, .codex/agents/, .agents/skills/, and .agents/docs/policies/.
Material limitation
The App Server protocol remains experimental; execution is opt-in, trust-gated, and adapter-controlled rather than native custom-subagent identity.

Claude Code

Supported with limitations
Integration
CLAUDE.md plus .claude/agents/ and .claude/skills/.
Material limitation
Authenticated execution is opt-in; workspace-write delivery and full workflow evidence are not release-gated.
Example workflow

From request to evidence in one reviewable path

This terminal is illustrative, not a live run. It shows the shape of a delivery that records the request, safe branch, implementation, observed validation, and evidence.

  1. 1
    Request

    State the change and the constraints that shape it.

  2. 2
    Safe branch

    Pass the Branch Gate before write-mode work begins.

  3. 3
    Implementation

    Deliver the scoped change under explicit ownership.

  4. 4
    Validation

    Use observed command results, not an assumed pass.

  5. 5
    Evidence

    Return the outcome, artifacts, and limitations for review.

Illustrative terminalExample only
$ npx aw execute "Add contact form validation"
Example branch: feat/contact-form-validation

Request          ✓ classified
Branch Gate      ✓ safe branch established
Implementation   ✓ scoped change
Validation       ✓ observed commands
Evidence         ✓ delivery summary
Handoff          ✓ reviewable result

Example only — results depend on the repository, runtime, and policy.
Install

Install once. Choose your runtime.

AI Workflow Kit 2.13.0 requires Node.js 20.11 or newer and npm 10 or newer. Add the package, project the integration for the runtime you use, run the doctor check, and then start through that runtime's documented path.

Install AI Workflow Kit View npm package
Common dependencynpm install -D @williambeto/ai-workflow

OpenCode

Primary
npx aw init --yes
npx aw doctor
npx aw execute "your task"

Use opencode.jsonc and the generated OpenCode assets. ai-workflow execute is the CLI-enforced path for branch, delivery, validation, and evidence gates.

Codex

Supported with limitations
npx aw init --yes --codex
npx aw doctor
npx aw execute --runtime=codex "your task"

Trust the repository in Codex first. The App Server path is opt-in and adapter-controlled; doctor fails closed when the trust gate is absent.

Claude Code

Supported with limitations
npx aw init --yes --claude
npx aw doctor
claude --agent <name>

Use the generated CLAUDE.md, native agents, and skills. Authenticated execution is opt-in and full workspace-write evidence is not release-gated.

Antigravity

Supported
npx aw init --yes --antigravity
npx aw doctor
Select Atlas in the Agent selector

Use the generated ANTIGRAVITY.md and .agents/ projection. In the IDE, select Atlas or begin a prompt with /atlas.

Honest boundaries

What this is not

AIWK makes workflow controls and evidence more explicit. It does not remove the need for engineering judgment.

!

Not a correctness guarantee

Tests, review and production safeguards are still required.

!

Not a replacement for CI or security

AIWK complements existing engineering controls; it does not certify the software itself.

!

Not runtime-independent

Runtime and model capabilities still affect direct-agent behavior and available enforcement.

!

Not the same path in every runtime

Direct model interaction may not have the same enforcement as the AIWK control plane, especially where integrations are limited.

Build a delivery path you can inspect.

Make AI-assisted delivery more predictable, reviewable, and evidence-backed.

Install AI Workflow Kit View npm package ↗