Fleet Command is a Home Assistant add-on that orchestrates a multi-stage AI worker pipeline to generate dashboards, YAML configs, and other outputs.
Runs a configurable pipeline of AI workers — Project Manager → Manager → Generator → Reviewer → Supervisor — each assigned to a model of your choice (local Ollama, Claude, OpenAI-compatible, or any custom HTTP endpoint).
- Universal chunking — messages are split to fit each worker's context window. Nothing is ever truncated.
- Python-only assembler — dashboard fragments assembled in Python, no AI worker needed for that stage.
- 3-pass reviewer — entity resolution, structure review, and card_mod validation as separate chunked passes.
- Escalation system — workers signal
ESCALATE: <reason>when stuck; the level above provides guidance and the worker retries. - Modelfile support — per-harness Ollama Modelfiles stored and pushed from the UI.
- Harness probe — test connection and auto-retrieve context window before saving a new worker.
- Role minimums — configure recommended context window / token allowance per role. Harness cards flag workers that fall below the minimum.
- Auto context window resolution — Ollama harnesses query
/api/showat load time to resolve their real context window.
- Modelfile generator — generate a complete Ollama Modelfile per harness from the UI. Fetches existing Modelfile from Ollama as base, merges in role persona, chain-of-command rules, escalation signal format, per-role parameters (repeat_penalty, top_k, top_p, seed), stop sequences, and few-shot MESSAGE examples. Editable before push. Warns before overwriting a previously pushed Modelfile.
- Token allowance respected in chunking —
available_input_tokensnow reservesmax(min_output_tokens, token_allowance)so input chunks are correctly sized for workers with large output budgets. - Role minimum tags — harness capability tags now only show roles the worker actually meets the context window minimum for.
- Role minimum warnings on role cards — assigned worker flagged with ⚠ if it falls below the minimum for that role.
- Dropdown cleaned up — role assignment picker shows clean model names only, no capability warnings in the list.
| Stage | Role | Purpose |
|---|---|---|
| Project Manager | Planning | Breaks job spec into blocks and card list |
| Manager | Task breakdown | Converts plan into one task per card |
| Generator | Code / YAML | Produces one card per task |
| Assembler | Python | Combines fragments into a full dashboard |
| Reviewer | QA (3-pass) | Entity resolution → structure → card_mod |
| Supervisor | Sign-off | Validates against spec, routes rejections |
- Add this repository to your Home Assistant add-on store.
- Install and start the Fleet Command add-on.
- Open the Fleet Command panel via the sidebar.
- Assign AI models to roles in the Staff tab.
- Submit a job in the Projects tab.
The add-on source lives in fleet-command/.