Universal Edition: Support for Claude, OpenCode, Gemini, and Codex

[waleedtarbosh]

🚀 Update: Universal Edition Framework Available

Hello everyone! I hope you're all doing well.

Taking Mads' advice from Issue #45, I have built a Universal Edition of this framework.

🏗️ Architectural Improvements

To avoid duplicating workflow files and causing configuration drift, I implemented a zero-duplication modes/ architecture.

• Single Source of Truth: The framework now relies on a centralized AGENTS.md.
• Lightweight Pointers: Introduced specific pointer files (OPENCODE.md, GEMINI.md, CLAUDE.md, and CODEX.md) to route the agents.
• Runtime-Agnostic: Everything functions exactly as the upstream intended, but is now completely independent of the runtime environment.

✅ Testing & Validation

I have personally tested the full workflow end-to-end (setup ➔ scrape ➔ apply) using OpenCode, and it works flawlessly.

🔗 Try It Out

You can explore the fork and run it with your favorite AI CLI here:
👉 waleedtarbosh/ai-job-search

If anyone encounters any issues, needs help setting it up, or wants to suggest fixes, please don't hesitate to reach out. I'd be happy to help!

[MadsLorentzen]

This is genuinely impressive work, and I can see you took the #45 reasoning seriously rather than just porting around it. Before anything else, though, one thing to flag for your own sake:

Your config/01-candidate-profile.md on the public master branch has your real profile filled in - name, email, and location. That's /setup doing what it's designed to do (write your details into that file), but it's a tracked file, so on a public fork it's now visible to anyone, and the framework gives you no push-time warning - that part is on us, not you. A few things worth knowing:

• A fork of a public repo can't be made private on its own (GitHub's own note: you can't change a fork's visibility while it's a fork).
• Deleting the data in a new commit isn't enough - it stays in the git history. Scrubbing it properly means rewriting history (or, cleanest given the repo is only days old, recreating the fork from a placeholder-clean state).
• GitHub's guide for this: Removing sensitive data from a repository.

Now the fun part - the architecture. I looked closely and you did land real zero-duplication: modes/ holds the content once, .claude/commands/ are thin pointers into it, and the harness files all route to the same place. Testing setup->scrape->apply end-to-end in OpenCode is exactly the kind of validation I can't do in review, so thank you for that.

One honest note in case staying synced with upstream matters to you. The shape I'd sketched in #45 keeps .claude/commands/ as the canonical full content (untouched from upstream) and adds just a thin AGENTS.md pointing at it - the point being that git pull upstream then merges cleanly forever. Your version inverts that: canon moved into modes/ and the .claude/ files became pointers. It's clean internally, but it means future upstream changes to .claude/ will land on pointers and won't map onto modes/, so you'd be hand-porting each update. If you want the fork to track upstream with less friction, keeping .claude/ canonical and pointing into it is the lower-drift path.

On linking it so others can find it (which I did offer in #45): I'd love to once the profile file is placeholder-clean - I won't point people at a fork that's exposing personal data. Sort that out and give me a shout, and we can figure out the best place for it. Really nice work.

[waleedtarbosh]

Thanks so much for the detailed review and the critical heads-up regarding the personal data!

I have just completely rewritten the git history and force-pushed a clean state. The repository is now 100% placeholder-clean with no personal data remaining anywhere in the commit history.

Regarding the architecture: you were absolutely right about the upstream drift. I have completely reverted the modes/ structure and restored .claude/commands/ as the canonical source of truth exactly as you sketched in #45. Now, the AGENTS.md and the CLI-specific pointer files (OPENCODE.md, GEMINI.md, etc.) just point directly to the .claude/ directory. This keeps the fork perfectly aligned for clean upstream merges in the future while still supporting all the other AI CLIs with zero duplication.

I'd love it if you could link to the fork now that the data leak is secured and the architecture is perfectly aligned.

Please let me know if you need any further modifications. I'd love to stay in touch and contribute to maintaining and developing the project in the future, specifically focusing on the OpenCode integration since that's what I use the most these days! Thanks again for this masterpiece of a framework.

[MadsLorentzen]

Good turnaround, and I didn't just take the reply's word for it: I cloned the repo and swept every reachable commit in the rewritten history for the data that was exposed - zero hits. The rewrite is real and done right, and the .claude/commands/ restore fixes the drift problem completely; upstream changes to commands and skills will now merge cleanly into your fork. That was the hard part, and it's solid.

Two small things before the row goes up, both in the harness docs rather than the architecture:

1. The quickstarts still describe the old layout. OPENCODE.md (and the Gemini guide) tell users the setup will "populate the config/ files" - that directory no longer exists after your restructure - and use a /job-search mode="setup" syntax that isn't defined anywhere in the tree. Someone following the guide today hits dead references at step 2. Point them at what actually exists now (.claude/commands/ + the profile files under .claude/skills/), and it's fixed.
2. Heads-up rather than an ask: your top layer is the mirror image of the Setup for opencode #45 sketch - AGENTS.md is the canonical file and CLAUDE.md a thin importer, where the sketch kept CLAUDE.md canonical with one thin AGENTS.md pointer. Fine to keep if that's your preference, but upstream changes to CLAUDE.md won't map onto your structure, so the index row will note it as "tracks upstream commands/skills; top-layer restructured" - same honest-caveat treatment other restructured forks get.

Fix the quickstarts, ping me here, and I'll add the row. And the OpenCode-focused contribution offer is noted with thanks - .claude/commands/ staying canon is exactly what makes that sustainable.

[waleedtarbosh]

Thanks for the thorough review and for verifying the cleanup!

I've just updated the quickstarts in both OPENCODE.md and GEMINI.md. All dead references to the config/ directory and outdated syntax have been completely removed. The guides now correctly point users to .claude/commands/ and .claude/skills/.../01-candidate-profile.md.

Also, thanks for the heads-up on the AGENTS.md vs CLAUDE.md structure. I'm completely fine with the index row noting the top-layer restructure.

Everything is pushed and ready on my end. Feel free to add the row! And if there's anything else you need me to adjust, don't hesitate to reach out at any time.

[MadsLorentzen]

Saw the quickstart fixes land (checked both files directly): OPENCODE.md and GEMINI.md now route cleanly to AGENTS.md + .claude/commands/, the stale config/ references are gone, and the walkthrough steps match the real commands. That closes out everything from my last round - thank you for following through, and for the earlier history rewrite that made this listable in the first place.

Your Universal Edition is now listed in the #78 index as the multi-harness adaptation, with the agreed note that the top layer is restructured (AGENTS.md canon, CLAUDE.md as pointer - the inverse of upstream's #45 shape), so readers know core-file merges need care while commands and skills track upstream. Congrats - first four-harness edition in the index. 🎉

[waleedtarbosh]

Sounds good, thanks for the update and for your help getting this across the finish line! Really glad to see the Universal Edition officially listed in the index. 🎉
If you or anyone else using this run into any issues down the road or have ideas for improvements, just ping me here and we'll get them sorted out together.