/reflect

name: reflect
description: Reflect on recent work and propose improvements if patterns are noticed.

Record notes only inside an open record's turn, using `bun ${CLAUDE_PLUGIN_ROOT}/scripts/task.ts note .hermit <id>` with the note on stdin. Otherwise skip record notes. Never edit a task file directly.

Reflect

Pause and think about your recent work.

**Silent by default.** Only notify the operator (per § Operator Notification) if reflect produces an outcome: a proposal task, a graduated observation, or a cost spike.

Rare-branch procedures live in `${CLAUDE_PLUGIN_ROOT}/skills/reflect/branches.md`. "Read branches.md § X" means: read that section now and follow it exactly — it is normative.

Commands

• `reflect-precheck`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/reflect-precheck.ts .hermit ${CLAUDE_PLUGIN_ROOT}`
• `update-reflection-state`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/update-reflection-state.ts .hermit/state/reflection-state.json`
• `routines-health`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/routines.ts health .hermit`
• `task-report`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/task-report.ts .hermit --recent --limit 3`
• `observations-patterns`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/observations.ts patterns .hermit`
• `observations-observe`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/observations.ts observe .hermit`
• `proposal-anchor`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/proposal.ts anchor .hermit`
• `proposal-gate`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/proposal.ts gate .hermit`
• `task-note`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/task.ts note .hermit`

Before writing a `quick-deferral` or `reflect-noticed` row, run `observations-patterns` (Commands) once per run. Reuse an existing label that names the same component and failure; create a short stable label only when none matches. Use that exact label on stdin for each observation.

Quick mode

If `$ARGUMENTS` contains `--quick`, read `${CLAUDE_PLUGIN_ROOT}/skills/reflect/quick.md`, follow it, and stop afterward.

Single-check mode

If `$ARGUMENTS` contains `--check-id` or `--check`, read branches.md § Scheduled checks and follow it exactly, then **stop** (no precheck or numbered steps below). The invocation is `reflect --check-id <id> --check <cmd…>`: everything after `--check` is the skill command, including its arguments, verbatim. A routine owns the cadence and may pin the model for this invocation.

Judge evidence handoff

For every `hermitd:reflection-judge` call, include `Task records:` with the fresh normalized adapter rows already read, preserving their `source_path`, outcome, title, lessons, and waiting reason. Candidate `Sessions:` identifiers for record evidence are task IDs from those source paths. A candidate outside that bounded evidence set needs another targeted adapter read, not an archive fallback.

Scheduled reflect

1. Determine whether a full run is warranted. With `--precheck-verdict '<verdict>'`: the reflect routine already ran the precheck in bash; use `<verdict>` directly, do **not** re-run it (always status `run`; the routine stops on `empty` without loading this skill). Otherwise run it yourself and read its JSON:

   bun ${CLAUDE_PLUGIN_ROOT}/scripts/reflect-precheck.ts .hermit ${CLAUDE_PLUGIN_ROOT}

• status `empty` → nothing due; the precheck already updated `reflection-state.json`. Emit `reflect: no candidates` and stop.
• status `run` → continue. `phases` lists due phases (`cost_spike`, `behavior`, `compute`, `digest`, `newborn`, `observations_fresh`); skip sections for phases not listed. `observations_fresh` means the ledger has rows newer than `last_run_at`; run step 3b even if `compute` is absent.

2. Run `bun ${CLAUDE_PLUGIN_ROOT}/scripts/task-report.ts .hermit --recent --limit 3` for the three most recent done, cancelled or unconfirmed records. Read `task.ts list .hermit --open --json` for the execution observation. Read fresh; never fall back to session archives. 3. If `cost_spike` is listed: the precheck already detected the spike and wrote the row (`cost-spike:<YYYY-MM-DD>` — the measured day, which is **yesterday**, carrying `day_total` and `median_7d` as fields), so there is nothing to read or record here — it graduates via step 3b like any other observation. Do **not** re-read cost-log.jsonl to restate a number the precheck already computed.

3a. **Behavioral digest** — if `behavior` is listed: run `bun ${CLAUDE_PLUGIN_ROOT}/scripts/transcript-digest.ts .hermit --record-observation` (bounded ~20-line JSON of ground-truth counters — tool failures, tool_rejections by kind, wakes vs productive_wakes, compaction_events, subagent_dispatches; never `Read` a transcript directly). Two uses of the counters:

• **Defer-loop auto-row**: `--record-observation` makes the script write the `defer-loop` row itself when its own counters cross the threshold, citing the observed `window.from→to` span. Nothing to record here. (The flag is opt-in so an ad-hoc digest run stays a pure read.)
• **Anomaly checklist** — carry the JSON as inherited context to the think-hard step and form one candidate per hit, citing the digest JSON (machine-written-state evidence class): any `tool_rejections.automode-blocked > 0`; any single tool with `tool_failures[tool] ≥ 5`; `compaction_events ≥ 3`. A hit is a **question for the gates** (triage → judge), never a verdict — a quiet week with mostly unproductive wakes is legitimate, so the gates own false-positive pruning.

3b. **Observations ledger** — prune, then graduate recurring patterns.

• Run `bun ${CLAUDE_PLUGIN_ROOT}/scripts/prune-observations.ts .hermit` (fail-open).
• Run `bun ${CLAUDE_PLUGIN_ROOT}/scripts/observations.ts graduate .hermit`. It prints the patterns that cleared the session threshold and the graduation cursor this run: `{"candidates":[{"pattern", "sessions": [...], "origin": "own-work"|"external-content", "rows": N}]}`. Empty `candidates` means nothing graduated; continue.
• Build one candidate per entry: `Evidence Source: archived-se