| # AIOS Cockpit |
|
|
| Local cockpit for the agentic operating system that is already spread across |
| Samuel's projects. |
|
|
| This repo is the cockpit layer only. It does not replace the existing sources |
| of truth: |
|
|
| - `samuel-os-v2` owns focus, inbox, decisions, and learning discipline. |
| - `KomandoZentral` owns deterministic routing decisions. |
| - `Zentrale` owns context packs, harness profiles, and run records. |
| - `ProjectDashboard` remains useful prior art for project scanning. |
| - `PromptGarage` owns prompt memory. |
| - `Sidecar-V8` owns assistant coaching and safety. |
| - Harness projects own evidence and eval results. |
|
|
| ## Run |
|
|
| ```bash |
| cd /home/smlflg/Projekte/AIOS-Cockpit |
| python -m aios_cockpit.cli snapshot |
| python -m aios_cockpit.app |
| ``` |
|
|
| Then open: |
|
|
| ```text |
| http://127.0.0.1:8092 |
| ``` |
|
|
| ## What the MVP does |
|
|
| - Shows today's focus from `samuel-os-v2`. |
| - Scans `~/Projekte` for active, source-of-truth, experiment, and archive |
| candidates. |
| - Provides project detail pages with README/PLAN excerpts and delegation |
| warnings. |
| - Reads harness profiles and latest runs from `Zentrale`. |
| - Provides run detail pages for Zentrale `RUN.json` and `log.json` artifacts. |
| - Reads PromptGarage stats without mutating the database. |
| - Reads first harness-eval signals from the SWE-bench Lite experiment DB. |
| - Mounts `CodexMaxHarnessHowTo` as `/codex-max-harness/`. |
| - Produces canonical implementation handoffs for Codex/OpenCode/Claude. |
| - Saves handoffs as auditable Markdown artifacts under `handoffs/`. |
| - Calls `KomandoZentral` in-memory when available; otherwise returns a |
| conservative fallback routing recommendation. |
|
|
| ## V2 flow |
|
|
| 1. Open the cockpit and inspect today's focus. |
| 2. Open a source-of-truth or active project detail page. |
| 3. Create a handoff from that project page or from the cockpit form. |
| 4. Review the saved handoff under `/handoffs`. |
| 5. Use the handoff as the bounded implementation prompt for Codex/OpenCode. |
|
|
| Generated handoffs are local repo artifacts. They are the first intentional |
| write flow in this cockpit and are kept small enough to inspect before use. |
|
|
| ## V3 workbench |
|
|
| V3 adds the first real working surface: |
|
|
| - `/workbench` is now the primary entry point. |
| - Pick a project and convert README/PLAN signals into sprint candidates. |
| - Add candidates to a cockpit-local task queue. |
| - Route tasks through `KomandoZentral`. |
| - Save agent-neutral handoffs from a task. |
| - Prepare and dry-run a `Zentrale opencode-prompt` launch command. |
|
|
| The run launcher is guarded: it writes instructions and preview records under |
| `workspace/`, uses `--dry-run`, and does not edit source projects. A future |
| version can add an explicit execute mode after the preview workflow proves |
| useful. |
|
|
| ## Working rule |
|
|
| The cockpit is allowed to aggregate and display truth. It must not become a |
| new competing truth store unless a field cannot reasonably live in an existing |
| system. |
|
|
| ## Still out of scope |
|
|
| - Starting autonomous source-mutating agents from the UI. |
| - Writing into upstream projects. |
| - Replacing `Zentrale` or `KomandoZentral`. |
| - Importing every historical project into a new database. |
|
|
