/hermit-health

name: hermit-health
model: haiku
description: "Show alert state, proposal queue depth, routine engagement, knowledge state, channel availability, plus fragile zones, stale accepted proposals, and recent learnings. Activates on messages like 'health check', 'how's the hermit', 'is anything broken', 'hermit health', 'system health', 'anything wrong', 'hermit status', 'hermit brain', 'are routines running', 'what's stuck', 'any fragile zones', 'show me what's blocked', 'recent learnings', 'what have you learned lately', 'where are the weak spots', 'check knowledge', 'lint knowledge', 'knowledge health'."

Hermit Health

Synthesize a compact "state of the hermit" snapshot in two halves: infrastructure (active alerts, proposal queue, routine engagement, knowledge state, channel readiness) and analysis (fragile zones, stale accepted proposals, recent learnings). Everything is read directly in main — this skill is on-demand and runner-free.

Step 0 — Channel reply

If this skill was invoked from a channel-arrived message (the inbound prompt contains a `<channel source="...">` tag), reply via that channel's reply tool. Otherwise emit to conversation.

Scope

Read the following (gracefully skip any file that doesn't exist). The nine sources are independent — read them concurrently:

1. `.claude-code-hermit/state/alert-state.json` — the `alerts` object, keyed by alert id (`checklist:*`, `proposal-pending:PROP-NNN`, `micro-proposal-pending:*`). Each entry has `text`, `count`, `first_seen`, `last_seen`, `suppressed`, `consecutive_clean`. 2. `.claude-code-hermit/state/runtime.json` — `last_activity`, `session_id`. 3. `.claude-code-hermit/state/reflection-state.json` — `last_reflection` timestamp, `counters` (including `judge_suppress_by_code` map and run/output fields), and `queue` (pending micro-proposals and reflect candidates). 4. `.claude-code-hermit/config.json` — `routines` array (id, schedule, enabled); `channels` object (each channel's `default_chat_id` and `dm_channel_id`). 5. `.claude-code-hermit/proposals/PROP-*.md` — glob; count by `status` frontmatter field, and read `id`, `title`, `accepted_date`, `resolved_date`, `tags` for the stale-proposal and fragile-zone analysis. 6. `.claude-code-hermit/state/micro-proposals.json` — count entries with `status: "pending"`. 7. Glob counts: `.claude-code-hermit/raw/**` (excluding `.archive/`), `.claude-code-hermit/compiled/**`, `.claude-code-hermit/raw/.archive/**`. 8. `.claude-code-hermit/sessions/S-*-REPORT.md` — glob all, sort descending by filename, read the 5 most recent; parse `status`, `tags`, `proposals_created` frontmatter. 9. `.claude-code-hermit/sessions/SHELL.md` — current session `tags`, blockers, and `## Findings` entries (fallback source for recent learnings).

Analysis

**Alerts:** From `alert-state.json → alerts`, active means `suppressed` is not `true`; suppressed ones are digest-only and reported separately. Compute each entry's age from `first_seen`. Report the active count with the oldest entry's `text` and age, then the suppressed count with its oldest age. If no active entries: "No active alerts" (still name the suppressed count when there is one).

**Proposal queue:** Count proposals with `status: proposed` (pending operator review), `status: accepted` (in flight, not yet resolved), and `status: in_progress`. If all zero: "Queue empty."

**Routine engagement:** From `config.json.routines`, list each routine and its schedule. For the `reflect` routine, use `reflection-state.json → last_reflection` to show when it last ran; also append counter info from `reflection-state.json → counters`: apply the reflect-line rules below, then append the suppress-mix suffix if any `judge_suppress_by_code` code has a non-zero count: `suppress mix — no-evidence:N, covered-by-memory:N, no-sessions:N` (omit codes with count 0; omit the suffix entirely when all counts are 0 or the map is absent). For every routine, run `bun ${CLAUDE_PLUGIN_ROOT}/scripts/routines.ts health .claude-code-hermit` once and take last run, fire count and skip reason from its rows.

**Reflect-line rules** (applied to the reflect routine bullet):

• If `counters` is absent: omit the counter clause entirely
• If `total_runs` is 0: show `no runs yet (since YYYY-MM-DD)` where the date is `counters.since`
• Otherwise: `N runs, N empty | output: N proposals, N micro | since YYYY-MM-DD` where the date is `counters.since` — omit `| output: ...` if both `proposals_created` and `micro_proposals_queued` are 0

**Micro:** Count entries with `status: "pending"` in `micro-proposals.json`. Omit this section entirely if count is zero.

**Knowledge:** Report glob counts for raw (excl. archive), compiled, and archived files. Omit if all three are zero and the directories are missing. If the operator explicitly asked to check or lint knowledge (phrasing like "check knowledge", "lint knowledge", "knowledge health"), also run `bun ${CLAUDE_PLUGIN_ROOT}/scripts/knowledge-lint.ts .claude-code-hermit` and relay its findings verbatim (grouped by type with file paths, ages, and actionable advice) beneath the counts. The script is strictly read-only.

**Channel availability:** From `config.json.channels`, for each configured channel, check whether `default_chat_id` or (failing that) `dm_channel_id` is set — the same fallback chain the outbound resolver uses. Report "ready" or "not yet paired (no chat id — send a message first)".

**Fragile zones:** From the last 5 session reports, gather the `tags` array from sessions with `status: partial` or `status: blocked`. Also gather `tags` from proposals with `status: dismissed` or `status: blocked`. Surface the top 2–3 tag clusters that appear repeatedly across fragile outcomes. If no blocked/partial sessions exist: "No fragile zones detected."

**Stale proposals:** From proposals, find those with `status: accepted` and `resolved_date` absent or `null`. Sort by `accepted_date` ascending (oldest first). Show up to 3. Compute days open = tod