Factory skills / v0.1Local pilot
Your first factory starts with one fix.
Turn the Guide into practice. Use your Skillz skills and an existing Codex session to take a small bug from a failing test to a reviewable fix.
A fix you can inspect.
An isolated branch, the failure before the fix, the passing checks afterward, and a handoff for human review.
- 01Failing regression
- 02Passing candidate
- 03Reviewable evidence
Start here
Run the example first.
You need Node.js 22.13 or newer, Git, and macOS or Linux. Extract the ZIP, then run these commands from the folder containing factory-starter. Use a new skillz folder.
git clone https://github.com/jaydubya818/skillz.git skillz
git -C skillz checkout --detach 4942dde4b2a442df3ee45879bb22734c658426df
node factory-starter/demo.mjs ./skillzWhat should happen?
The example counts only two batches for five items at capacity two. The regression expects three. The script records the mismatch, applies the known fix, and runs the checks again.
Look for “status”: “verified”. The output includes the evidence directory.
This is a scripted demonstration. It does not claim an agent solved a new task.
Your skills, one workflow
Five phases. One source of truth.
The kit references a pinned version of your Skillz repository. Its generated instructions guide your Codex session through the relevant skills, with the original supporting files and attribution intact.
- 01
- 02
- 03
- 04
- 05
Next: your repository
Give Codex a bounded task.
Start with a small, trusted, dependency-free Node repository. The README includes the task format and complete commands.
Define the contract
Name the bug, expected assertion mismatch, implementation files that may change, and existing tests. Keep the task JSON outside the repository.
Begin the run
The runner creates a worktree and branch, then checks the baseline. It writes a PROMPT.md with the selected skills and workflow commands.
Open the worktree in Codex
Ask your session to follow PROMPT.md. It writes the regression before the fix and keeps the same test for verification.
Review the handoff
Inspect the patch and logs. Check for stale evidence before committing. Your normal PR checks, merge approval, and release process still apply.
Know what the pilot checks.
The runner checks changed files, workflow order, test results, and later changes to the recorded candidate. It stops on missing or failing checks, rewritten regression tests, timeouts, and exhausted retry budgets.
It is not a security sandbox or independent verifier. Use your harness permissions and trusted repositories. Logs stay local and may contain private data. A passing run does not approve a merge or establish production readiness.
Version 0.1 supports Node’s built-in test runner on macOS and Linux. Other languages, automatic PR creation, and Mission Control admission are future work.
Keep going
Understand the method behind the workflow.
Download v0.1.0 · 38 KB · SHA-256 9a689038e08eb2a60c2f9c7254d56b22b61d6aebffefad042cd64d245d22487a