distill

/habit:distill: Sweep & Restructure

Runs in forked subagent. All data is pre-loaded below. Use only Bash commands from the Operations reference for writes. Summaries must be human-friendly. Do not mention file names, counters, timestamps, or pruning stats.

User prompts from this session

!bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh read-transcript ${CLAUDE_SESSION_ID}

Current Index (merged)

!bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh read-index merged

Pending Sessions (from prior sessions)

!bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh read-pending-distill

Execution Log

!bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh read-log

Global Metadata

!bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh read-meta global

Project Metadata

!bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh read-meta project

Processing Rules

!bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh read-shared PROCESSING.md

Operations

!bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh read-shared OPERATIONS.md

Routing

Arguments: $ARGUMENTS

Before routing: If the arguments contain feedback about the plugin beyond the routing keyword (complaints, questions about behavior, suggestions), log each piece as an observation before proceeding: bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh log-observation ${CLAUDE_SESSION_ID} '<feedback>'.

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

1. Arguments line above contains "project" → go to Project below.
2. Arguments line above contains "maintain", "pending", or "deep" → go to Maintain below.
3. Arguments line above is empty → go to Regular below.

---

Project

Scan new or modified project sessions incrementally and restructure the full inventory. Ignore the preloaded session transcript above; all session data is loaded fresh here.

1. List new/modified sessions: bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh list-new-sessions. If it returns "No new project sessions." or "No project sessions found.", return that message and stop.
2. Collect the file paths from the output (one per line). Split into batches of 5.
3. For each batch, load transcripts: bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh read-transcript <path1> <path2> ... <path5>. Run Sweep on the batch data. After each batch, mark processed: bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh mark-sessions-distilled <path1> ... <path5>.
4. Run Restructure.
5. Run Self-improve.
6. Run Cleanup.
7. Return summary per Summary Format below. Opening line: "Scanned N new project sessions." (Batch size: 5 transcripts per read-transcript call. Adjustable based on session length.)

Maintain

Session sweep followed by full inventory restructure and plugin self-improvement. Data sources: the preloaded session transcript and pending sessions above. Do not call read-sessions or load sessions by any other means.

1. If the current session transcript above is empty and the pending sessions list is empty, skip to step 3.
2. Gather prompt sources (current session transcript is already loaded above). Collect all transcript paths from the pending list. Fetch them in one call: bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh read-transcript <path1> <path2> .... Missing files are skipped automatically. Run Sweep on gathered data (current session + pending transcripts) if any usable prompts exist.
3. Run Restructure.
4. Run Self-improve.
5. Run Cleanup.
6. Return summary per Summary Format below. Opening line: "Swept current session and N pending." or "No new session data to sweep. Ran inventory maintenance."

Regular

Data sources: the preloaded session transcript and pending sessions above. Do not call read-sessions or load sessions by any other means.

1. If the current session transcript above is empty and the pending sessions list is empty, return "Nothing to extract yet." and stop.
2. Gather prompt sources (current session transcript is already loaded above). Collect all transcript paths from the pending list. Fetch them in one call: bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh read-transcript <path1> <path2> .... Missing files are skipped automatically. If zero usable prompts remain after fetching, return "Nothing to extract yet." and stop.
3. Run Sweep on gathered data.
4. Check the preloaded Global Metadata above: if update_counter >= 20 in either scope, continue to Maintain step 3 (skip step 5; Maintain ends with its own Cleanup).
5. Run Cleanup.
6. Return summary per Summary Format below. Use the same opening line rule as Maintain.

---

Summary Format

Each branch specifies its own opening line. After that, include:

• One sentence per new habit: what pattern it captures and its origin.
• One sentence per merge: what was merged and why.
• Categories of skipped prompts, or note that sweep was skipped.
• Override pattern findings from the execution log.
• Inventory health: size, archive/stale actions, or confirm clean.
• Self-improve: plugin edits made, or "No observations pending."

Sweep

1. Apply the Processing Rules: classify each prompt, interpret, dedup, and structure. Look for prompts that solve a specific problem but contain a generalizable principle. Extract the principle, decide if it applies globally or only to this project, and scope accordingly.
2. For each reusable pattern found, write or merge via the write-habit command (see Operations).
3. Check execution log for override patterns (3+ similar on same habit). List any patterns found. Restructure will act on this list later in the same flow.
4. For each habit written, verify: run bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh read-habit <id> and confirm frontmatter is valid, description starts with a verb and is under 120 chars, instruction is self-contained. If not, rewrite.
5. Verify: every new habit has a unique id, all tags are lowercase singular nouns, no two habits in the index would produce the same agent behavior. If any check fails, fix before proceeding.
6. Flag plugin friction in the swept data (routing errors, output complaints, habit execution struggles) as observations via log-observation. These feed Self-improve later in the flow.

Restructure

• Merge convergent habits (would produce the same agent behavior). Compare each pair independently. If A and B merge, re-compare the result against remaining habits.
• Normalize tags (ts→typescript, js→javascript).
• Rename IDs that violate the [a-z0-9-] format. Flag renames in summary.
• Archive stale (never executed AND created 30+ days ago AND not updated in 30+ days).
• Act on override patterns found during Sweep: create scoped variants or update base habits.
• Quality check remaining habits: verify each description accurately reflects actual usage (check execution log overrides), tags are specific and useful for matching, and the instruction reads as a clear standalone directive. Fix any that are vague, misleading, or drifted from how they're actually used.
• Run bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh self-heal global and bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh self-heal project to rebuild indexes.
• Reset meta: bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh reset-meta global and reset-meta project.
• Prune log: bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh prune-log global and prune-log project.

Self-improve

Read observations: bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh read-observations. If none, skip.

For each observation, identify the plugin source file responsible (under ${CLAUDE_PLUGIN_ROOT}/skills/ or ${CLAUDE_PLUGIN_ROOT}/bin/). Read the file, apply the smallest edit that resolves the friction. Preserve existing structure and style. If no source file in ${CLAUDE_PLUGIN_ROOT}/skills/ or ${CLAUDE_PLUGIN_ROOT}/bin/ clearly maps to the observation, skip that observation and note it in the summary instead of editing anything.

Cleanup

1. Clear pending: bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh clear-pending-distill.
2. Clear observations: bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh clear-observations.
3. Reset prompt counter: bash ${CLAUDE_PLUGIN_ROOT}/bin/habit-tools.sh reset-prompt-count ${CLAUDE_SESSION_ID}.