Finish the OS as a sellable QADIR Shift platform.
Today is not for more research. The finish line is a public store, a clean execution plan, a working QADIR Shift shell target, a model-brain decision, and smoke-test gates that separate real product from fantasy.
Finish Today Board
Execute in this order. Anything outside this board is a distraction unless it directly proves QADIR Shift, sells the starter pack, or repairs the install/run path.
Convert store from placeholder to platform surface
Replace coming-soon store with QADIR Shift, agentic starter, free teardown pack, and first-party pack catalog.
Add execution page and public route
Make the roadmap operational: today board, gates, subscription decision, and website cleanup sequence.
Run QADIR Shift local shell inventory
Find the active Electron/native app folder, verify install commands, backend URL, Ollama status, and broken controls.
Repair the minimum viable shell
Home, Tasks, Tools, Models, Agents, Memory, Packs, Settings. Every control must be usable or disabled with a setup reason.
Make one task run end-to-end
Mission input, plan, approval, runner log, artifact output, changed-file list, and smoke-test note.
Capture proof
Desktop screenshot, health board screenshot, task run output, starter package screenshot, and 2-minute demo script.
Subscription Decision
Buy one monthly product for agentic coding first: Cursor Pro+ at $60/month. It gives daily agent use, frontier model access, MCPs, skills, hooks, and cloud agents without locking ABUZ8 to one model vendor. (This is an internal tooling decision — not a product we sell.)
- Do not buy Devin first. It is valuable for software teams, but it is narrower and higher commitment for our current platform build.
- Do not buy Manus first. Useful to study, not the main build engine for QADIR Shift.
- Keep ChatGPT/Claude as reasoning/review layers if already available, but the purchase priority for execution is Cursor Pro+.
- If usage crushes the limit, upgrade to Cursor Ultra only after one week of logged usage proves it saves time.
Benchmark Rule
Benchmarks are now marketing inputs, not purchase authority. The only benchmark that matters for ABUZ8 is whether the tool can modify our repo, run tests, explain the diff, and ship a verified artifact.
- Use public benchmarks to shortlist models.
- Use our repo tasks to choose tools.
- Score by pass/fail output, not leaderboard rank.
- Separate soul/persona from coding model quality.
Local Brain Stack
| Layer | Decision | Reason |
|---|---|---|
| Daily local coder | Best served Qwen coder tier available locally | Fast enough to run constantly on ABUZ8 hardware and compatible with local-first OS positioning. |
| Elite code verifier | Cloud frontier model only for review or stuck tasks | Use paid models where they create leverage, not as the whole product dependency. |
| Mytho/persona | Zait/Qadir doctrine and soul layer | Mytho-style tuning belongs in behavior and identity, not as the core code model. |
| Real benchmark | ABUZ8 task suite | Fix repo issue, run command, produce artifact, log proof. That beats mixed internet reviews. |