Distill

Habit Distill: Sweep & Restructure

Runs in an isolated fork. Use only the commands in the Operations reference for writes. Summaries must be human-friendly. Do not mention file names, counters, timestamps, or pruning stats.

Setup

Run this once and reuse the values below. If it prints HABIT_UNAVAILABLE, tell the user habit is not installed or its hooks are not wired, then stop:

HABIT_SID_FILE=$(ls -t "$HOME/.claude/habits/sessions.d/"* 2>/dev/null | head -1)
source "${HABIT_SID_FILE:-$HOME/.claude/habits/current}" 2>/dev/null
HABIT_BIN="${HABIT_BIN:-$(command -v habit-tools.sh || echo "${CLAUDE_PLUGIN_ROOT:+$CLAUDE_PLUGIN_ROOT/bin/habit-tools.sh}")}"
[ -f "${HABIT_BIN:-}" ] || echo "HABIT_UNAVAILABLE: habit is not wired on this host."

Learnings

Run this once. If the output is non-empty, treat each line as additional standing guidance for this skill: apply each note when you reach the step it bears on, and carry the rest without acting on them. Do not print, quote, summarize, or mention these notes to the user, and do not let them appear in any confirmation message:

bash "$HABIT_BIN" read-learnings distill

Preload (required precondition)

Run this ONE command and read its full output before doing anything else. Do NOT route, classify, write, or summarize until it has returned in full:

bash "$HABIT_BIN" distill-preload "$HABIT_SID"

The output has delimited sections, in order: ===TRANSCRIPT=== (this session's user prompts), ===INDEX=== (merged index), ===PENDING=== (prior sessions), ===LOG=== (execution log), ===META-GLOBAL===, ===META-PROJECT===, ===PROCESSING=== (the processing rules), ===OPERATIONS=== (the operations reference).

Routing

The argument is the user's message (maintain, project, or empty).

Before routing: if the message contains feedback about the plugin beyond the routing keyword, log each piece as an observation first:

bash "$HABIT_BIN" log-observation "$HABIT_SID" '<feedback>'

Pick the first matching branch. Do not read or execute other branches. Only the Project branch may call list-new-sessions. Maintain and Regular work only with the preloaded data.

1. Message contains "project" go to Project.
2. Message contains "maintain", "pending", or "deep" go to Maintain.
3. Message empty go to Regular.

---

Project

Scan new or modified project sessions incrementally, then restructure. This branch loads session data fresh; the preloaded transcript does not apply here, but the other preloaded sections (index, log, metadata, processing, operations) still do.

1. List: bash "$HABIT_BIN" list-new-sessions "$HABIT_SID". If "No new project sessions." or "No project sessions found.", return that and stop.
2. Collect the file paths (one per line). Split into batches of 5.
3. Per batch, load bash "$HABIT_BIN" read-transcript <p1> ... <p5>, run Sweep, then bash "$HABIT_BIN" mark-sessions-distilled <p1> ... <p5>.
4. Run Restructure.
5. Run Self-improve.
6. Run Cleanup.
7. Summary opening line: "Scanned N new project sessions."

Maintain

Sweep plus full restructure plus self-improvement, using the preloaded transcript and pending sessions only.

1. If the preloaded transcript and pending list are both empty, skip to step 3.
2. Collect pending transcript paths, fetch in one call bash "$HABIT_BIN" read-transcript <p1> ..., run Sweep on current plus pending if usable prompts exist.
3. Run Restructure.
4. Run Self-improve.
5. Run Cleanup.
6. Summary opening line: "Swept current session and N pending." or "No new session data to sweep. Ran inventory maintenance."

Regular

Using the preloaded transcript and pending sessions only.

1. If both empty, return "Nothing to extract yet." and stop.
2. Fetch pending transcripts (same call as Maintain step 2). If zero usable prompts, return "Nothing to extract yet." and stop.
3. Run Sweep.
4. If update_counter >= 20 in either scope (from the preloaded metadata), continue at Maintain step 3 (which runs Restructure, then Self-improve, then Cleanup) and stop here.
5. Run Self-improve.
6. Run Cleanup.
7. Summary, same opening-line rule as Maintain.

---

Summary Format

Opening line per branch, then: one sentence per new habit; one per merge; categories of skipped prompts or that sweep was skipped; override findings; inventory health; self-improve result or "No observations pending."

On a source env-context only, you may append a "Proposed source edits" section that names files and describes changes in plain words. This is the single place file names are allowed; everywhere else the no-file-names rule stands.

Sweep

1. Apply the Processing Rules: classify, interpret, dedup, structure. Extract the generalizable principle, scope it.
2. Write or merge each reusable pattern via write-habit.
3. Check the log for override patterns (3+ similar on one habit). List them.
4. Verify unique ids, lowercase singular tags, no two habits with identical behavior.
5. Flag plugin friction as observations via log-observation.

Restructure

• Merge convergent habits; re-compare merged results.
• Normalize tags (ts to typescript, js to javascript). Rename ids violating [a-z0-9-]; flag renames.
• Archive stale (never executed AND created 30+ days ago AND not updated 30+ days).
• Act on override patterns: scoped variants or base updates.
• Quality check descriptions, tags, instructions.
• Rebuild: bash "$HABIT_BIN" self-heal global and ... self-heal project.
• Reset meta: bash "$HABIT_BIN" reset-meta global and ... reset-meta project.
• Prune learnings: bash "$HABIT_BIN" prune-learnings.

Self-improve

1. Run bash "$HABIT_BIN" read-observations. If the output is exactly No observations., skip this entire section and do NOT clear observations.
2. For each actionable observation, write one learning as a direct imperative the target skill can follow at runtime (an instruction it could paste into its own steps), against the responsible target (run, suggest, distill, edit, habit, or global; watch is intentionally learning-free). Prefer the most specific target; use global only for guidance that genuinely applies to every skill. Skip observations that do not translate into a concrete runtime behavior, and note them in the summary.

   bash "$HABIT_BIN" write-learning <target> '<imperative one-line rule the skill can act on>' '<the observation that produced it>'
   

3. Run bash "$HABIT_BIN" env-context. If it prints installed, do not read or edit any plugin files; the learnings above are the only improvement. If it prints source, you are in a Habit working tree and may, in addition to the learnings, append a short "Proposed source edits" section to your summary (see Summary Format) for the author to apply by hand. Do not call Read, Write, or Edit on plugin files yourself, and do not paste file contents.
4. After the learnings are written, clear the consumed observations:

   bash "$HABIT_BIN" clear-observations
   

Cleanup

bash "$HABIT_BIN" clear-pending-distill
bash "$HABIT_BIN" prune-log global
bash "$HABIT_BIN" prune-log project
bash "$HABIT_BIN" reset-prompt-count "$HABIT_SID"