| name | autoresearch |
| description | Optional, operator-gated metric experiments under .scratch/ for Home Assistant Homelab. Prefer prototype for design questions. Use only when a measurable eval exists and the operator wants a short experiment log — not a swarm research loop or docs/research PR. |
| disable-model-invocation | true |
Autoresearch — Home Assistant Homelab
Honest fit check first: this is a Home Assistant config repo, not a
benchmark-heavy code swarm. Most questions here are better answered by
prototype, live HA traces, or
architecture-review.
There is no docs/research/ surface and no research PRs. Git is
operator-owned (operator-owned-git.md).
This repo does not use git worktrees — experiments stay under .scratch/ in
the main checkout.
When NOT to use this skill
Stop and route elsewhere when:
| Situation | Use instead |
|---|
| "Does this automation/state model feel right?" | prototype in .scratch/ |
| Package layout / coupling / design debt | architecture-review |
| Vague idea needs scoping | alignment |
| Broken YAML / automation | config-validate / automation-debug |
| No measurable metric or eval command | Do not invent one — hand off |
Operator wants a durable research writeup under docs/research/ | Decline — that path is not first-class here |
When it is appropriate (rare)
Only when all of these hold:
- Operator explicitly wants metric-driven experiments (or find-work surfaces
this as a last-resort brief with a clear metric).
- There is a real eval command that produces a numeric metric (script
runtime, entity count query, validation exit code over a corpus, etc.).
- Experiments stay under
.scratch/ — never
under packages/ as throwaway.
Slim workflow
- [ ] 1. Confirm fit (or refuse and route)
- [ ] 2. Agree hypothesis, in-scope scratch paths, eval, metric
- [ ] 3. Baseline under .scratch/<slug>/
- [ ] 4. Run a short experiment loop (keep notes in .scratch; no commits)
- [ ] 5. Stop when proven / disproven / exhausted / operator interrupts
- [ ] 6. Hand off: findings + recommended next skill; operator decides what to promote
Contract (keep tiny)
| Field | Required |
|---|
| Hypothesis | Yes — testable claim |
| In-scope paths | Yes — only under .scratch/<slug>/ |
| Eval command | Yes — no embedded secrets |
| Metric name + direction | Yes — lower or higher |
| Constraints | Optional |
Do not change the contract mid-run. If the hypothesis evolves, stop and start
a new slug.
Rules
- No commits, pushes, or PRs
(
operator-owned-git.md).
- No edits to
packages/ as experimental throwaways — copy ideas into
scratch, measure, then promote via implement-change if warranted.
- Log experiments as plain markdown/JSON under
.scratch/<slug>/ (gitignored).
- Cap the loop: prefer a handful of deliberate trials over an autonomous swarm.
- At the end, summarize findings for the operator. Do not open a docs PR.
Hand off
## Autoresearch handoff
**Hypothesis:** …
**Result:** proven | disproven | exhausted | interrupted
**Best metric:** … (baseline …)
**Scratch path:** `.scratch/<slug>/`
**Recommend next:** prototype | alignment | implement-change | drop
Related
prototype · architecture-review ·
alignment · .scratch/README.md
