Pynchy Ops

Pynchy Ops

The live Pynchy host and checkout path are deployment-specific. Public repo instructions must not assume a private hostname or home-directory layout. Set PYNCHY_HOST and PYNCHY_REMOTE_ROOT from local memory, environment, or the operator before running remote commands.

Auto-deploy: Never Restart Manually

Pynchy self-manages. Two mechanisms trigger automatic restarts:

1. Git changes on main — the polling mechanism detects new commits, pulls, and restarts (with container rebuild if source files changed).
2. Config file changes — editing config.toml, litellm_config.yaml, or other settings files triggers an automatic deploy on the next host git-sync poll. The default interval is 300 seconds; check [scheduler].git_sync_interval_seconds before deciding it was missed.

Do not manually restart containers or the service. This includes docker restart, systemctl restart, and direct container management (docker kill/stop/rm). Manual restarts bypass lifecycle management and can leave things in a bad state.

Only use manual commands when the service is unhealthy and needs fixing. See references/server-debug.md for diagnostic steps.

Quick Status Check

Preferred: the authenticated control-plane CLI. It uses the permission-restricted Unix socket on the live host:

# On the live host directly:
cd "$PYNCHY_REMOTE_ROOT"
uv run pynchy status

# Remotely over SSH:
PYNCHY_HOST="${PYNCHY_HOST:?set the live host}"
PYNCHY_REMOTE_ROOT="${PYNCHY_REMOTE_ROOT:?set the live checkout path}"
ssh "$PYNCHY_HOST" "cd '$PYNCHY_REMOTE_ROOT' && uv run pynchy status"

Returns JSON with: service (uptime), deploy (SHA, dirty, unpushed), channels (slack/whatsapp connected), gateway (LiteLLM health), temporal (cluster health, worker state, task queue, last scheduled workflow/result), queue (active containers, waiting groups), repos (per-repo worktree status — SHA, dirty, ahead/behind, conflicts), messages (inbound/outbound counts, last activity), tasks (scheduled tasks with status/next run), host_jobs, groups (total, active sessions).

Fallback: manual commands (when the HTTP server is down or you need logs):

# 1. Is the service running? (macOS live host)
launchctl print "gui/$(id -u)/com.pynchy"

# 2. Any running containers?
docker ps --filter name=pynchy

# 3. Any stopped/orphaned containers?
docker ps -a --filter name=pynchy

# 4. Recent errors in the dedicated error log?
tail -n 100 "$PYNCHY_REMOTE_ROOT/logs/pynchy.error.log"

# 4a. Need surrounding application context from macOS launchd?
tail -n 200 "$PYNCHY_REMOTE_ROOT/logs/pynchy.stdout.log"

# 5. Is Slack/WhatsApp connected?
tail -n 200 "$PYNCHY_REMOTE_ROOT/logs/pynchy.stdout.log" | grep -E 'Connected to|Connection closed|Slack'

# 6. Are groups loaded?
tail -n 200 "$PYNCHY_REMOTE_ROOT/logs/pynchy.stdout.log" | grep groupCount

Deploy & Observe

Before deploying source changes, commit one logical change on a feature branch and merge it into main. Do not leave the production checkout dirty or deploy an uncommitted implementation. Deployment-specific ignored configuration may change separately when needed, but source changes always go through a commit.

# Trigger a deploy through the live host's Unix socket. From containers, use
# mcp__pynchy__deploy_changes instead.
PYNCHY_HOST="${PYNCHY_HOST:?set the live host}"
PYNCHY_REMOTE_ROOT="${PYNCHY_REMOTE_ROOT:?set the live checkout path}"
ssh "$PYNCHY_HOST" "cd '$PYNCHY_REMOTE_ROOT' && uv run pynchy deploy"

# Observe (always safe)
ssh "$PYNCHY_HOST" 'launchctl print gui/$(id -u)/com.pynchy'
ssh "$PYNCHY_HOST" "tail -n 100 '$PYNCHY_REMOTE_ROOT/logs/pynchy.stdout.log'"
ssh "$PYNCHY_HOST" "tail -n 100 '$PYNCHY_REMOTE_ROOT/logs/pynchy.error.log'"
ssh "$PYNCHY_HOST" 'docker ps --filter name=pynchy'

# Manual restart — ONLY for unhealthy/stuck service
ssh "$PYNCHY_HOST" 'launchctl kickstart -k gui/$(id -u)/com.pynchy'

Monitoring Live Agent Activity

Service logs only show lifecycle events (container spawn, session create/destroy, errors). They do NOT show agent output (tool calls, thinking, text broadcasts). To monitor what an agent is actually doing, query SQLite:

# Recent activity for a specific group (replace <JID> with e.g. slack:C0AFR6DB0FK)
ssh "$PYNCHY_HOST" "cd '$PYNCHY_REMOTE_ROOT' && sqlite3 data/messages.db \"
  SELECT timestamp, message_type, substr(content, 1, 120)
  FROM messages WHERE chat_jid = '<JID>'
  ORDER BY timestamp DESC LIMIT 15;
\""

# All recent activity across all groups
ssh "$PYNCHY_HOST" "cd '$PYNCHY_REMOTE_ROOT' && sqlite3 data/messages.db \"
  SELECT timestamp, chat_jid, message_type, substr(content, 1, 80)
  FROM messages ORDER BY timestamp DESC LIMIT 15;
\""

Temporal Scheduler

Scheduled work runs through Temporal. Pynchy reconciles active agent tasks, database host jobs, and config cron jobs into Temporal schedules or delayed workflows. Pynchy owns the worker in the host process; Temporal owns workflow durability and wake-ups.

macOS launchd deployment:

Safe checks:

ssh "$PYNCHY_HOST" 'launchctl print gui/$(id -u)/com.pynchy.temporal'
ssh "$PYNCHY_HOST" 'temporal operator cluster health --address 127.0.0.1:7233'
ssh "$PYNCHY_HOST" 'lsof -nP -iTCP:7233 -sTCP:LISTEN'
ssh "$PYNCHY_HOST" "cd '$PYNCHY_REMOTE_ROOT' && uv run pynchy status"

data/temporal.db is durable scheduler state. Make sure host backups include it with the rest of data/.

Runtime DB Backups

macOS deployments can use scripts/backup_runtime_dbs.sh for SQLite-safe runtime DB snapshots. It backs up messages.db, neonize.db, and temporal.db into data/backups by default or into the explicitly configured SSH destination. Remote backups stage locally, verify checksums on the destination, and publish atomically. The script briefly unloads and reloads the Temporal LaunchAgent around the temporal.db snapshot; never run an online SQLite backup against the active Temporal development server because a write collision can leave its transaction state wedged.

Live service:

Safe checks:

ssh "$PYNCHY_HOST" 'launchctl print gui/$(id -u)/com.pynchy.backup'
ssh "$PYNCHY_HOST" 'launchctl print gui/$(id -u)/com.pynchy.backup | grep PYNCHY_BACKUP_'
ssh "$PYNCHY_HOST" 'tail -n 50 ~/Library/Logs/pynchy/backup.log'
ssh "$PYNCHY_HOST" 'tail -n 50 ~/Library/Logs/pynchy/backup.err.log'

When to use what:

Exercising Message Ingress

Pynchy does not expose a production HTTP endpoint for injecting user messages. Send a test message from a real account through a configured channel so the test crosses the channel's authentication and ingestion boundaries. Inspect the resulting messages and agent activity in SQLite as described in server debugging.

Service Management Reference

macOS:

launchctl load ~/Library/LaunchAgents/com.pynchy.plist
launchctl unload ~/Library/LaunchAgents/com.pynchy.plist

Linux:

systemctl --user start pynchy
systemctl --user stop pynchy
systemctl --user restart pynchy
journalctl --user -u pynchy -f          # Follow logs

Systemd unit template: config-examples/pynchy.service.EXAMPLE

Workspace GitHub CLI Access

GitHub CLI access requires a selected type = "workspace" tool whose required_env includes GITHUB_TOKEN. Pynchy does not discover gh auth credentials or inject a broad token into admin agents. Host-side repository operations retain their separate scoped-token resolution.

Verify only that the managed Pynchy host process receives GITHUB_TOKEN; never print the value. See Tool access and secrets for the canonical configuration.

Production Secret Materialization

Production credentials must enter the managed Pynchy host process through Proton Pass. Keep pass:// references in data/proton-pass/pynchy.env; the managed service must start scripts/run_pynchy.sh, which invokes pass-cli run when that template exists. Do not put resolved values in a launchd plist, systemd unit, workspace file, container argument, or generated env directory.

After updating the Pass items or tool requirements, use the normal managed deployment flow. Verify requirement names and tool availability through status, logs, or a canary without printing raw task environments or credential values.

Container Build Cache

Apple Container's buildkit caches the build context aggressively. --no-cache alone does NOT invalidate COPY steps. To force a truly clean rebuild:

container builder stop && container builder rm && container builder start
./src/pynchy/agent/build.sh

Verify: container run -i --rm --entrypoint python pynchy-agent:latest -c "import agent_runner; print('OK')"

LiteLLM Gateway

Runs as pynchy-litellm Docker container with PostgreSQL sidecar (pynchy-litellm-db). Access at http://localhost:4000 on the Pynchy host, or via Tailscale at port 4000.

Resolve the master key only through an approved secret mechanism into $KEY. Never echo, log, or paste it. Pass it only in the Authorization: Bearer $KEY header.

Config: $PYNCHY_REMOTE_ROOT/litellm_config.yaml. Editing it triggers an automatic deploy on the next host git-sync poll (300 seconds by default). Do not manually restart containers.

Dashboard: http://$PYNCHY_HOST:4000/ui/

Warning: /spend/logs is quarantined for routine live diagnostics regardless of requested limit. Do not use it or /global/spend/logs as a substitute.

• Safe gateway diagnostics, readiness, and failure evidence: references/litellm-diagnostics.md
• MCP server management API and gotchas: references/litellm-mcp-api.md

Zombie Processes (LiteLLM)

If SSH login reports zombie processes, check whether they live inside the LiteLLM container:

ssh "$PYNCHY_HOST" 'docker exec pynchy-litellm ps -eo pid,ppid,stat,args | awk '\''$3 ~ /Z/ {print}'\'''

Note: use args, not cmd — cmd can appear empty for zombie processes.

MCP Server Containers

MCP tool servers (e.g., Playwright) run as separate Docker containers managed by McpManager. They start on-demand when an agent needs them and stop after the configured idle_timeout.

See src/pynchy/host/container_manager/mcp/ and MCP management.

Database Files

All databases live in data/:

Quick inspection (run on the live host or prefix with ssh "$PYNCHY_HOST" "cd '$PYNCHY_REMOTE_ROOT' && ..."):

# List registered groups
sqlite3 data/messages.db "SELECT name, folder, is_admin FROM registered_groups;"

# Recent messages across all channels
sqlite3 data/messages.db "SELECT timestamp, chat_jid, sender_name, substr(content, 1, 80) FROM messages ORDER BY timestamp DESC LIMIT 10;"

# Active sessions
sqlite3 data/messages.db "SELECT * FROM sessions;"

# Scheduled tasks
sqlite3 data/messages.db "SELECT id, group_folder, status, next_run FROM scheduled_tasks WHERE status = 'active';"

For the full query cookbook (traces, tool calls, cross-table debugging), see the pynchy-dev skill's sqlite-queries.md.

Server Debugging

For specific failure scenarios — container timeouts, agent not responding, mount issues, WhatsApp auth — see references/server-debug.md.

Docker logs are useful for runtime errors (container crashes, process failures) where the issue occurs before messages reach the database. For agent behavior, use the pynchy-dev skill's SQLite query reference instead.