oah helps you build an AI development workflow that fits you and keeps changing as AI improves. It is a working reference and a guide you can adapt.
A useful result needs more than generated code: the scope must be clear, the changes checked, and decisions kept for the next session. oah connects those steps so one request can carry through the work with a record of what happened.
002 / Why it keeps changing
Older rules and skills can hold a newer model back or make its results worse. Defaults should be easy to understand and change; a rule that gets in your way is a design problem to fix.
Keep the parts that help, remove what no longer helps, and revise the workflow as your tools improve. You can start with the full setup or take only the impact map, wiki, leases, evidence gates or review loop that your project needs.
003 / Why Claude builds and Codex reviews
This split comes from the author's own experience, not a measured comparison: Claude fills in details a beginner might miss in a first version; Codex checks and corrects within a fixed scope.
The current commands run Claude Code and Codex CLI. The ideas can also be adapted to another tool or harness, with one AI or several. Give each stage to the AI that suits your work.
004 / From one request to a checked change
Describe the result you want in plain language. The AI records the goal and scope, checks what the change touches, reserves its paths and chooses the tools for the task. Claude builds; Codex reviews, corrects and reviews again.
Gates record the checks and their results. The project wiki keeps the decisions and reasons beside the code. In v0.3.0, optional TypeSafe Jev can judge difficulty for model selection; rules remain the fallback when it is unavailable or uncertain.
005 / The decisions stay with you
You ask, answer when the AI needs a decision, and confirm before the change lands. If the same problem repeats or the work reaches a design choice, the loop stops to ask you.
Ask “Is the review done?” or “Did all the gates pass?” and inspect the evidence. The final confirmation belongs to you; writing code, checking it and approving its landing are separate steps.
006 / Free, open source, yours to adapt
oah is free, open-source software under the MIT license. Read the source, change it and bring the useful pieces into your own workflow. Your AI services and any optional services use their own accounts and billing.
The current v0.3.0 release supports macOS and native Windows, and expects familiarity with git and pull requests. Install it, run oah init in your project, then try the small hello-oah example. The installation and documentation sections explain each step.
New in v0.3.0: oah models report counts how model choices were made — tier source, Jev's difficulty status and confidence, the model judge's status, and how often a floor raised the tier. oah models export saves one JSON line per run as cases you can compare later. Counts alone do not prove Jev helped.
When you need it
Without a record of how each run's tier and model were chosen, you cannot compare the rules with Jev or tune a rule later.
How it works
oah models report reads this repository's run records (<state>/runs/<task>/<run-id>/result.json) and counts by stage, engine and in total; --json prints the same counts as one object with a format number.
oah models export prints one JSON line per run folder, ordered by task folder and run id, with only decision data (axes, stage, engine, tiers, Jev and judge fields, model choice) and the outcome (status, exit code, timeout, duration).
Prompts, briefs, task names, paths, commands, environment values and worker output are left out; values a correction record hashed stay hashed.
Both call no provider, reach no network and write nothing; redirect stdout to keep the export.
Example
Print the counts as one JSON object (from Upgrade).
oah models report --json
Requirements
Requires v0.3.0 or later.
Run records in this repository.
Limits and recovery
The counts show how choices were made, not whether Jev helped: no outcome or quality is compared.
A run that finished does not prove its choice was right.
Older runs lack newer fields and count as not recorded; a value that cannot be read counts as unreadable, never as zero.