FileSystemWatcher + DuckDB Skill

Status: ✅ Production Ready Trit: 0 (ERGODIC - coordination/observation) Principle: Observe filesystem changes, persist to DuckDB, query temporally GF(3) Balanced Triad:

• fswatch-duckdb (0) [Coordinator: Watch/observe]
• duckdb-temporal-versioning (+1) [Generator: Store events]
• clj-kondo-3color (-1) [Validator: Verify integrity]

---

Overview

Resilient filesystem monitoring with:

1. fswatch or watchman for native file events
2. DuckDB for persistent event storage with time-travel
3. DuckLake patterns for partitioned, versioned storage
4. Auto-start on every Amp session

Auto-Initialization (MANDATORY)

On every Amp session start, run:

# Initialize watcher database
duckdb ~/.agents/fswatch.duckdb <<'SQL'
CREATE TABLE IF NOT EXISTS fs_events (
  id INTEGER PRIMARY KEY,
  path VARCHAR NOT NULL,
  event_type VARCHAR NOT NULL,  -- 'created', 'modified', 'deleted', 'renamed'
  old_path VARCHAR,             -- for renames
  size BIGINT,
  mtime TIMESTAMP,
  checksum VARCHAR,             -- md5 for dedup
  session_id VARCHAR,           -- Amp thread ID
  created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
  trit INTEGER DEFAULT 0        -- GF(3) color
);

CREATE INDEX IF NOT EXISTS idx_fs_events_path ON fs_events(path);
CREATE INDEX IF NOT EXISTS idx_fs_events_session ON fs_events(session_id);
CREATE INDEX IF NOT EXISTS idx_fs_events_time ON fs_events(created_at);
SQL

Watch /tmp with fswatch

Start Watcher (Background Daemon)

# Using fswatch (macOS/Linux)
fswatch -0 -r /tmp | while IFS= read -r -d '' path; do
  event_type="modified"
  if [ ! -e "$path" ]; then
    event_type="deleted"
  elif [ ! -s "$path.prev" 2>/dev/null ]; then
    event_type="created"
  fi
  
  size=$(stat -f%z "$path" 2>/dev/null || echo 0)
  mtime=$(stat -f%m "$path" 2>/dev/null || date +%s)
  checksum=$(md5 -q "$path" 2>/dev/null || echo "")
  
  duckdb ~/.agents/fswatch.duckdb \
    "INSERT INTO fs_events (path, event_type, size, mtime, checksum, session_id) 
     VALUES ('$path', '$event_type', $size, to_timestamp($mtime), '$checksum', '$AMP_THREAD_ID')"
done &

Babashka Watcher (Recommended)

#!/usr/bin/env bb
;; scripts/fswatch-daemon.bb

(require '[babashka.process :refer [shell process]]
         '[babashka.fs :as fs]
         '[clojure.java.io :as io])

(def db-path (str (System/getenv "HOME") "/.agents/fswatch.duckdb"))
(def watch-paths ["/tmp" "/var/tmp"])
(def session-id (or (System/getenv "AMP_THREAD_ID") (str (random-uuid))))

(defn record-event! [path event-type]
  (let [size (if (fs/exists? path) (fs/size path) 0)
        mtime (if (fs/exists? path) 
                (-> path fs/last-modified-time str) 
                "1970-01-01")
        checksum (when (and (fs/exists? path) (fs/regular-file? path))
                   (-> (shell {:out :string} "md5" "-q" path) :out str/trim))]
    (shell "duckdb" db-path
           (format "INSERT INTO fs_events (path, event_type, size, checksum, session_id) 
                    VALUES ('%s', '%s', %d, '%s', '%s')"
                   path event-type size (or checksum "") session-id))))

(defn watch-loop []
  (let [proc (process {:out :stream} 
                      "fswatch" "-0" "-r" (first watch-paths))]
    (with-open [rdr (io/reader (:out proc))]
      (loop []
        (when-let [line (.readLine rdr)]
          (let [path (str/trim line)
                event-type (cond
                             (not (fs/exists? path)) "deleted"
                             :else "modified")]
            (record-event! path event-type))
          (recur))))))

(println "🔍 Starting FileSystemWatcher for" watch-paths)
(println "📊 Persisting to" db-path)
(println "🧵 Session:" session-id)
(watch-loop)

DuckLake Partitioning

For high-volume /tmp monitoring, partition by hour:

-- Create partitioned view
CREATE VIEW fs_events_hourly AS
SELECT 
  date_trunc('hour', created_at) as hour_bucket,
  path,
  event_type,
  size,
  checksum,
  session_id
FROM fs_events;

-- Export to DuckLake-style Parquet partitions
COPY (
  SELECT * FROM fs_events 
  WHERE created_at >= CURRENT_DATE
) TO '/tmp/ducklake/fs_events' 
(FORMAT PARQUET, PARTITION_BY (date_trunc('day', created_at)));

DuckLake Catalog Integration

-- Attach DuckLake catalog (if available)
ATTACH 'ducklake:fswatch' AS lake (DATA_PATH '/tmp/ducklake/data');

-- Sync local events to lake
INSERT INTO lake.fs_events 
SELECT * FROM fs_events 
WHERE session_id = current_session();

Queries for Resilience

Recent Events (Last Hour)

SELECT path, event_type, size, created_at
FROM fs_events
WHERE created_at > CURRENT_TIMESTAMP - INTERVAL '1 hour'
ORDER BY created_at DESC
LIMIT 50;

Events by Session (Cross-Amp Thread)

SELECT session_id, COUNT(*) as event_count, 
       MIN(created_at) as first_event,
       MAX(created_at) as last_event
FROM fs_events
GROUP BY session_id
ORDER BY first_event DESC;

Find Duplicates by Checksum

SELECT checksum, COUNT(*) as copies, 
       array_agg(path) as paths
FROM fs_events
WHERE checksum IS NOT NULL AND checksum != ''
GROUP BY checksum
HAVING COUNT(*) > 1;

Time-Travel Query

-- What was in /tmp at a specific time?
SELECT * FROM fs_events
WHERE created_at <= '2025-12-24 10:00:00'
  AND path LIKE '/tmp/%'
  AND event_type != 'deleted'
ORDER BY created_at DESC;

Session Persistence

Save Session State on Exit

# Add to ~/.zshrc or shell profile
fswatch_save_session() {
  if [ -n "$AMP_THREAD_ID" ]; then
    duckdb ~/.agents/fswatch.duckdb \
      "UPDATE fs_events SET trit = 1 WHERE session_id = '$AMP_THREAD_ID'"
  fi
}
trap fswatch_save_session EXIT

Restore Previous Session Context

-- Get context from previous session
SELECT DISTINCT path, event_type, MAX(created_at) as last_event
FROM fs_events
WHERE session_id = (
  SELECT session_id FROM fs_events 
  WHERE session_id != current_setting('amp.session_id')
  ORDER BY created_at DESC LIMIT 1
)
GROUP BY path, event_type;

GF(3) Trifurcated Monitoring

For Copy-on-Interact resilience, spawn 3 watchers:

;; Trifurcated watcher pattern
(defn trifurcated-watch [paths]
  (let [minus  (future (watch-with-trit paths -1))   ; Validate/filter
        ergodic (future (watch-with-trit paths 0))   ; Coordinate/aggregate
        plus   (future (watch-with-trit paths +1))]  ; Generate/store
    ;; GF(3) sum: -1 + 0 + 1 = 0 ✓
    {:validators @minus
     :coordinators @ergodic
     :generators @plus}))

Integration with Amp

Auto-Start on Thread Creation

Add to ~/.agents/AGENTS.md:

## FileSystemWatcher Auto-Start

On every session, run:
```bash
bb ~/.agents/skills/fswatch-duckdb/scripts/fswatch-daemon.bb &

Query recent activity:

duckdb ~/.agents/fswatch.duckdb "SELECT * FROM fs_events ORDER BY created_at DESC LIMIT 20"

### Justfile Commands

```just
# Start filesystem watcher daemon
fswatch-start:
  bb ~/.agents/skills/fswatch-duckdb/scripts/fswatch-daemon.bb &

# Stop all watchers
fswatch-stop:
  pkill -f fswatch-daemon

# Query recent events
fswatch-recent count="20":
  duckdb ~/.agents/fswatch.duckdb \
    "SELECT path, event_type, size, created_at FROM fs_events ORDER BY created_at DESC LIMIT {{count}}"

# Export to DuckLake
fswatch-export:
  duckdb ~/.agents/fswatch.duckdb \
    "COPY fs_events TO '/tmp/ducklake/fs_events.parquet' (FORMAT PARQUET)"

# Database stats
fswatch-stats:
  duckdb ~/.agents/fswatch.duckdb \
    "SELECT COUNT(*) as total_events, COUNT(DISTINCT path) as unique_paths, COUNT(DISTINCT session_id) as sessions FROM fs_events"

Surrounding File Context

When interacting with files, query related events:

-- Find all events for files in same directory
SELECT * FROM fs_events
WHERE path LIKE (
  SELECT dirname(path) || '%' 
  FROM fs_events 
  WHERE path = '/tmp/myfile.txt'
)
ORDER BY created_at DESC;

Cleanup & Maintenance

-- Archive old events (older than 7 days)
COPY (
  SELECT * FROM fs_events 
  WHERE created_at < CURRENT_DATE - INTERVAL '7 days'
) TO '/tmp/ducklake/archive/fs_events_archive.parquet';

-- Delete archived events
DELETE FROM fs_events 
WHERE created_at < CURRENT_DATE - INTERVAL '7 days';

-- Vacuum database
VACUUM;

QuickTime Recording Auto-Processor

Automatically processes screen recordings with triadic skill interleaving:

# Process single recording
bb ~/.agents/skills/fswatch-duckdb/scripts/quicktime-processor.bb ~/Desktop/recording.mov

# Watch mode (auto-process new recordings)
bb ~/.agents/skills/fswatch-duckdb/scripts/quicktime-processor.bb &

Triadic Processing Pipeline

Stream	Trit	Skill	Voice Persona
MINUS	-1	Frame extraction	Anna (Emmy Noether)
ERGODIC	0	Thumbnail generation	Amélie (Sophie Germain)
PLUS	+1	Audio extraction	Ava (Premium)

DuckLake Integration

Outputs stored in /tmp/ducklake/:

• frames/ - Extracted video frames
• recordings/ - Thumbnails
• audio/ - Extracted audio tracks

Interaction Entropy Coloring

Each recording gets a deterministic color based on its path entropy:

(def entropy (interaction-entropy path))
(def color (gen-color entropy 0))
;; color includes {:L :C :H :trit :seed}

Query Processed Recordings

-- Recent recordings
SELECT * FROM recording_processing ORDER BY processed_at DESC LIMIT 10;

-- Skill interleaving log
SELECT * FROM skill_interleave WHERE triplet_id = 0;

-- By trit color
SELECT input_path, trit, frame_count FROM recording_processing WHERE trit = -1;

---

Skill Name: fswatch-duckdb Type: FileSystem Observer with Temporal Storage + Video Processing Trit: 0 (ERGODIC - coordination) GF(3): Balanced with duckdb-temporal-versioning (+1) + clj-kondo-3color (-1) Auto-Start: Yes - runs daemon on every Amp session Watch Paths: /tmp, ~/Desktop, ~/Movies (configurable) DuckLake: /tmp/ducklake/

Scientific Skill Interleaving

This skill connects to the K-Dense-AI/claude-scientific-skills ecosystem:

Dataframes

• polars [○] via bicomodule

Bibliography References

• general: 734 citations in bib.duckdb

Cat# Integration

This skill maps to Cat# = Comod(P) as a bicomodule in the equipment structure:

Trit: 0 (ERGODIC)
Home: Prof
Poly Op: ⊗
Kan Role: Adj
Color: #26D826

GF(3) Naturality

The skill participates in triads satisfying:

(-1) + (0) + (+1) ≡ 0 (mod 3)

This ensures compositional coherence in the Cat# equipment structure.