domain-brainstorm

Domain Brainstorm

Kill criteria (read before running)

After ≥8 invocations, read state/proposal-metrics.jsonl and filter events where type:"brainstorm-emit" and skill:"ha-domain-brainstorm". Count CREATE vs total emits with ideas (triage-survival) and read status: of the resulting PROP files (PROP-acceptance). If triage-survival < 25% or PROP-acceptance < 30%, cut this skill rather than tune it — signal-to-noise isn't there.

Gate 0 — Gather inputs

Read these in parallel — all are cheap cached reads plus one lightweight listing. Do not run ha refresh-context, fetch individual automation configs in a loop, or actuate any device.

Entity inventory Read .claude-code-hermit/raw/snapshot-ha-normalized-latest.json → entity_index (the universe of areas, devices, and domains).

• If absent (ha-boot never ran): stop immediately, emit-zero with reason "no context snapshot — run ha-boot first". The entity universe is the required substrate.
• If mtime > 24h: note "stale snapshot", proceed — ha-boot owns refresh, do not refresh here.

Automation/script inventory Run:

${CLAUDE_PLUGIN_ROOT}/bin/ha-agent-lab ha list-automations
${CLAUDE_PLUGIN_ROOT}/bin/ha-agent-lab ha list-scripts

These are live REST calls. On error or timeout: do not retry — note "automation inventory unavailable" and continue with cached artifacts only. With no inventory, [coverage-asymmetry] and [unbuilt-intent] are unsupportable; a degraded run will usually emit-zero.

Operator intent Read .claude-code-hermit/OPERATOR.md (## HA hermit section) and the auto-memory MEMORY.md index. Note any stated wants or preferences ("close garage at night", "welcome light when arriving") that no automation name clearly implements.

Suppression filters (read to exclude ideas — never to source them)

• .claude-code-hermit/state/integration-health-degraded-domains.json — skip any idea whose target entity sits in a degraded domain.
• .claude-code-hermit/raw/snapshot-ha-pattern-analysis-latest.json / patterns-latest.md — already-flagged silent/dead items are ha-analyze-patterns' territory; do not surface them as capability gaps.

Gate 1 — Generate ideas (max 2)

Think across all four inputs. For each candidate, both constraints must pass before including it:

1. Concrete friction — state the operator pain in one sentence. What is the operator missing today because this automation or script doesn't exist? If no specific pain, discard.
2. ≥2 named grounding items — cite at least two by name (e.g. entity:binary_sensor.bedroom_motion, area:garage, script:bom_dia, operator:OPERATOR.md "close garage at night", health:no degraded domains). These support the friction; friction is the bar.

Map each passing idea to the closest HA capability-gap prefix:

• [automation-gap] — a device, sensor, or area present in the entity inventory but wired into zero automations.
• [coverage-asymmetry] — a paired-pattern gap (e.g. a bom_dia script exists but no boa_noite; a motion sensor turns lights on but nothing turns them off).
• [unbuilt-intent] — an operator-stated want (OPERATOR.md / auto-memory) that no automation or script name clearly implements.

Important precision note for [automation-gap]: a precise "wired into zero automations" coverage graph is not cheaply cached (ha-safety-audit persists entity references only for violating automations, not all). [automation-gap] therefore produces a candidate grounded in the alias inventory and entity_index — triage and the operator decision are the backstop. State this candidacy in the Evidence.

Drop any idea whose target entity/domain sits in a degraded integration (suppression filter). Cap at 2 ideas total. Emit-zero if none pass. Record discarded candidates (one line each) for Gate 4.

Gate 2 — Create proposals

For each idea, invoke /claude-code-hermit:proposal-create once:

Title: [<prefix>] <short idea title>
Evidence Source: capability-brainstorm
Evidence: <one paragraph: friction sentence + named grounding items>

Set frontmatter: source: auto-detected, category: improvement, tags: [domain-brainstorm, ideation].

Parse the verdict:

• CREATE — note PROP-NNN.
• SUPPRESS — <code> — record suppression code; don't retry.
• DUPLICATE:<PROP-ID> — record existing ID; don't create.

After each verdict, append a metrics event (Node stdlib, no deps). Tag skill:'ha-domain-brainstorm' to keep rows attributable to this plugin if other brainstorm skills ever share proposal-metrics.jsonl. Inlined here because append-metrics.js lives in the core plugin and is unreachable via ${CLAUDE_PLUGIN_ROOT}; JSON.stringify guarantees a valid line:

node -e "const fs=require('fs'); fs.appendFileSync('.claude-code-hermit/state/proposal-metrics.jsonl', JSON.stringify({ts:new Date().toISOString(),type:'brainstorm-emit',skill:'ha-domain-brainstorm',verdict:'<CREATE|SUPPRESS|DUPLICATE>',proposal_id:'<PROP-NNN or null>'})+'\n','utf-8');"

This event is what the kill-criteria audit reads — proposal-create's own created event does not carry per-skill provenance.

Do NOT invoke proposal-triage directly — /proposal-create handles it.

Gate 3 — Emit batch message

Send one message per the Operator Notification protocol in CLAUDE.md. Use the stored locale from OPERATOR.md (## HA hermit) for all user-facing text.

Zero-emit:

🏠 Domain brainstorm — 0 ideas emitted (<reason: thin context | automation inventory unavailable | all suppressed | all duplicates | stale snapshot>)

Non-zero:

🏠 Domain brainstorm (<N> idea(s))

1. **[prefix] <title>** — <one-line description>
   _Grounding: <item 1>, <item 2>_
   _Friction: <one-sentence pain>_
   PROP-NNN created  ·  (or: suppressed — <code>  ·  or: duplicate of PROP-NNN)

Gate 4 — Compiled artifact (non-empty runs only)

If ≥1 PROP was created (not suppressed or duplicate), write:

.claude-code-hermit/compiled/domain-brainstorm-YYYY-MM-DD-HHMM.md

---
title: Domain brainstorm — <ISO timestamp>
type: domain-brainstorm
created: <ISO timestamp with timezone>
tags: [domain-brainstorm, ideation]
source: interactive
proposals_created: [PROP-NNN, ...]
---

Body (150-line cap): ideas that passed (one paragraph each), discarded candidates (one line each), triage verdicts, inputs scanned (paths only, no content), note if inventory was degraded.

Do not tag foundational — this is a time-bounded ideation snapshot.

Zero-emit runs: skip the artifact entirely. Log one line to SHELL.md Findings: domain-brainstorm: 0 ideas emitted (<reason>)