CW · Codex Workflow by Queopius

Build with autonomy.
Advance with evidence.

AI development with gates, not guesses.

CW gives AI-assisted development a controlled workflow: plan, implement, validate, independently review, and advance only with verified evidence.

Get started (opens in a new tab)View on GitHub (opens in a new tab)
Open sourceApache-2.0CW CLI · v0.18.3
Real product flowRecorded from a real CW workflow

Recorded from a real CW workflow using CW version 0.18.3.

Goal: Add a deterministic greeting function with automated tests in one development phase using Python 3

  1. $ cw init
  2. ✓ Project initialized
  3. $ cw plan --goal "Add a deterministic greeting function with automated tests in one development phase using Python 3"
  4. ✓ Plan proposed
  5. $ cw plan show
  6. → 01-implement-greeting · Implement and test deterministic greeting
  7. $ cw plan approve
  8. ✓ Plan approved
  9. $ cw start
  10. ✓ Codex process started
  11. ✓ Codex session initialized
  12. → Implementation active
  13. → Running python3 -m unittest discover -v
  14. ✓ python3 -m unittest discover -v completed
  15. ✓ Project files updated · 2 changes
  16. ✓ Project files updated · 1 change
  17. $ cw review
  18. ✓ Deterministic validation passed
  19. ✓ Independent review approved
  20. ✓ Approval gate verified
  21. ✓ Completion review satisfied
  22. ✓ Completion target satisfied

Final result: completed with 1 approved phase and 1 valid gate.

Built in the open

Real signals. Verifiable engineering.

Adoption appears only when public GitHub data is meaningful. Until then, the product's engineering proof speaks for itself.

Release downloads
746
Downloads of CW distribution artifacts from stable GitHub Releases.
Apache-2.0
OPEN SOURCE
CW's source, architecture and safety model are publicly inspectable.
Read-only review
INDEPENDENT
Implementation and approval run as separate responsibilities.
Python versions in CI
5
Tested on Python 3.10, 3.11, 3.12, 3.13, 3.14.
REAL PRODUCT FLOWRecorded from a real CW workflow
  • Implementation
  • Deterministic validation
  • Independent review
  • Verified gate
Inspect the evidence (opens in a new tab)

Public GitHub data · updated periodically

The control layer

AI can write code.
CW controls what happens next.

Speed is not the same as control. Professional development still needs explicit scope, checks, independent judgment and a defensible reason to advance.

  • Scope
  • Validation
  • Independent review
  • Evidence
  • Controlled advancement
  • Recovery
Codeximplements.
CWsupervises.

Different responsibilities.
One controlled workflow.

A gated execution model

Every phase earns the next.

CW turns autonomous work into a sequence of bounded, verifiable transitions.

Plan

Define bounded work.

A read-only planner proposes explicit phases from bounded repository evidence. The plan is validated locally and waits for approval.
TRANSITION POLICYBOUND → CONTINUE
No valid gate.No next phase.

Separation of responsibility

Implementation is not approval.

CW deliberately separates the agent that changes the workspace from the agent that evaluates the result.

Independent responsibilities make every approval inspectable.

CW
CONTROL PLANECW Supervisor
STATE · POLICY · EVIDENCE
IMPLEMENTATIONImplementerworkspace-write
INDEPENDENT REVIEWReviewerread-only · ephemeral
Deterministic validation
Independent review
GATEVerified

Control without killing velocity

Room to work.
Boundaries that hold.

CW preserves agent momentum while keeping advancement tied to explicit, inspectable evidence.

Bounded autonomy

Authorize one phase—or a controlled batch with explicit phase, time and review budgets. Never unlimited autopilot.

Independent review

Implementation and review run as separate responsibilities with different permissions and isolated sessions.

Deterministic validation

Run approved project checks before semantic review, without shell interpretation or agent-injected commands.

SHA-256 gate evidence

Approved artifacts are hashed into gate evidence and rechecked before dependent phases begin.

Fail-closed recovery

Contradictory state and evidence stop execution. Backup-first repair reconciles validated evidence without inventing approval.

Live execution

Follow meaningful Codex checkpoints, elapsed time and redacted run records instead of watching a silent process.

The workflow, visible

Quiet when healthy.
Precise when it matters.

CW turns state, live execution and retained evidence into a coherent terminal surface—with stable output for people and versioned JSON for automation.

Codex Workflowproduct console
local workflowv0.18.3

~/code/sample-app

● connected

$ cw status

CW · Codex Workflow                                   v0.18.3by Queopius sample-app                                                 dev WORKFLOW                                              ● ACTIVE Progress    ##########--------------------   33%Approved    1 / 3 phasesState       IN PROGRESS CURRENT PHASE→ 02 · Build Authentication  Attempt       0 / 3  Readiness     NOT READY  Gate          PENDING
selectable outputNO_COLOR awareJSON available

Safety model

Autonomy with boundaries.

Serious controls, designed for forward motion. CW rejects ambiguous transitions instead of turning uncertainty into progress.

No valid gate. No next phase.

Implementation ≠ approval.

Evidence beats cached state.

Repair never fabricates approval.

Completed means completed.

FAIL-CLOSED
Code
Validate
Review
Verify
Advance

Developer experience

From repository to controlled workflow.

Install CW once. Initialize it inside a Git repository. Approve a bounded plan, then let the gate model do its work.

Project-localWorkflow state stays with the repository.

Explicit planningNo reliable goal means no invented work.

No silent updatesInstall and update actions stay intentional.

Quick start
git clone --branch v0.18.3 --depth 1 git@github.com:Queopius/cw.gitcd cw./install.sh cd ~/code/my-projectcw initcw plan --goal "Implement subscription billing"cw plan showcw plan approvecw # Existing managed installationcw update --version 0.18.3
RequiresGitPython 3.10+Codex CLI
PUBLIC REPOSITORY

CW is built in the open.

Inspect the architecture. Read the implementation. Review the safety model. Contribute.

View repository (opens in a new tab)View v0.18.3 release (opens in a new tab)Download v0.18.3 wheel (opens in a new tab)Report an issue (opens in a new tab)
Q
Queopius /cw
Public

A gate-driven workflow supervisor for controlled AI-assisted software development.

LATEST RELEASEv0.18.3Stable release (opens in a new tab)
PythonApache-2.0Public

Documentation

Understand every gate.

Installation, workflow design, planning, reviews, batch execution, integrations, recovery and architecture.

Read the documentation (opens in a new tab)

Getting started

Installation and first workflow

(opens in a new tab)

Workflow

States, phases and advancement

(opens in a new tab)

Gates & reviews

Evidence and approval contracts

(opens in a new tab)

Architecture

Roles, state and execution model

(opens in a new tab)

Troubleshooting

Diagnostics and safe recovery

(opens in a new tab)

Resources

Go deeper.

Documentation

Read the full CW guide.

(opens in a new tab)

GitHub

Source, issues and contributions.

(opens in a new tab)

Releases

Download and inspect releases.

(opens in a new tab)

Changelog

See what changed.

(opens in a new tab)

Security

Security policy and reporting.

(opens in a new tab)

License

Apache-2.0 terms.

(opens in a new tab)

FAQ

Before you hand over a phase.

What CW controls, what it leaves alone, and how it behaves at the boundaries.

What exactly is CW?

CW is a standalone, gate-driven workflow supervisor for explicit, reviewable AI-assisted development. It coordinates planning, current-phase implementation, deterministic validation, independent review and verified advancement.

Is CW a Codex wrapper?

No. CW is the control layer around a development workflow. Codex performs bounded planning, implementation and review roles; CW owns the state graph, validation order, permission boundaries, audit history and gate verification.

Does CW replace CI?

No. CW runs the approved deterministic checks configured for each phase before review. Your CI remains the repository's shared integration and delivery boundary; CW governs local phase advancement.

Why does CW use an independent reviewer?

Implementation and approval are intentionally separated. The reviewer is a new ephemeral Codex process with read-only access, hooks disabled and a structured evidence contract.

Can CW run multiple phases automatically?

Yes, within explicit limits. `cw run N` uses the same single-phase supervisor repeatedly and stops on human review, invalid evidence, integration failure, exhausted budgets, interruption or completion. There is no unlimited override.

What happens when a workflow becomes inconsistent?

CW fails closed and explains the integrity conflict. `cw repair` first creates a backup, then reconciles from validated authoritative evidence. It never fabricates an approval gate.

Does CW modify my global Codex configuration?

No. CW does not store MCP credentials, inject partial MCP server definitions or silently change `~/.codex/config.toml`. Managed processes retain your effective Codex configuration.

What happens when the workflow is complete?

The final valid gate writes `COMPLETED`. CW derives completion from the full contiguous chain of valid gates and never invents a successor phase: all valid gates, no next phase.

Codex Workflow

Build with autonomy.
Advance with evidence.

Give AI room to work—without giving up control.

Get started (opens in a new tab)View on GitHub (opens in a new tab)
No valid gate. No next phase.