Mac Mini Machine Context

Auto-generated setup snapshot. Update after any configuration changes.

Identity

Field	Value
Hostname	Nhan's Mac Mini
macOS	12 (Monterey)
Node type	mini (orchestrator)
Tailscale IP	100.87.151.9
Tailscale hostname	mini2015
Tailnet	zhinnyshop

Docker Services (compose/mini.yml)

All services bind to Tailscale IP only (no public ports).

Service	Image	Port	Status
postgres	pgvector/pg16:16	5432	healthy
redis	redis:7-alpine	6379	healthy
n8n	n8nio/n8n:latest	5678	ready
telegram-bridge	custom build	7700	polling mode
light-router	custom (python:3.11-slim)	4000	active
grafana	grafana/grafana:latest	3000	up
prometheus	prom/prometheus:latest	9090	up
iphone-pwa	custom build	8080	up

Pipeline

Field	Value
Process	/usr/local/bin/devstation-pipeline
PID file	.run/pipeline.pid
Port	8001
LLM mode	cloud-only (via light-router)
Providers	DeepSeek V4 Flash (primary), NVIDIA Nemotron 70B (fallback)

Telegram

Field	Value
Bot polling	active (no public URL needed)
Chat ID	configured
Notify endpoint	POST :7700/notify {message: string}
Commands	/status, /pause, /resume, /approve, /reject, /llm, /help

Network

Service	Internal URL	External URL
n8n	http://127.0.0.1:5678	http://100.87.151.9:5678
Pipeline	http://127.0.0.1:8001	http://100.87.151.9:8001
Telegram bridge	http://127.0.0.1:7700	http://100.87.151.9:7700
Light Router	http://127.0.0.1:4000	http://100.87.151.9:4000
Grafana	http://127.0.0.1:3000	http://100.87.151.9:3000
Prometheus	http://127.0.0.1:9090	http://100.87.151.9:9090
PWA	http://127.0.0.1:8080	http://100.87.151.9:8080

Known Config

• Tailscale CLI: Wrapper at /usr/local/bin/tailscale → /Applications/Tailscale.app/Contents/MacOS/Tailscale (brew formula NOT used — builds from source, very slow)
• Docker Desktop: Headless mode (openUIOnStartupDisabled: true), 2GB RAM, 2 CPUs, no K8s
• Python venv for pipeline: /tmp/pipeline-venv (created by setup script)
• n8n encryption key: Generated at setup, stored in .env (do not lose)
• Postgres password: Generated at setup, stored in .env (do not lose)
• DEEPSEEK_MODEL: deepseek-chat (set in .env)
• Pipeline workspace: /tmp/devstation-workspace (set via WORKSPACE_DIR in .env)

Workflow Integration

Linear issue (tag: "plan" / "implement")
  → n8n Trigger (webhook)
  → POST :8001/webhook
  → Pipeline (6 agents, cloud LLMs)
  → GitHub PR
  → Telegram notification

CLI Commands

./bin/devstation.sh setup       # One-command setup
./bin/devstation.sh status      # Check services
./bin/devstation.sh doctor      # Diagnostics
./bin/devstation.sh down        # Stop all
./bin/devstation.sh restart     # Restart all

Troubleshooting

Symptom	Fix
n8n "Mismatching encryption keys"	Delete n8n volume and recreate: docker volume rm devstation-mini_n8n_data && docker compose up -d n8n
Postgres auth failures	Delete db volume: docker volume rm devstation-mini_pg_data && docker compose up -d postgres
Port binding errors	Check Tailscale is connected: tailscale status
Pipeline not starting	Check .run/pipeline.log, run: /usr/local/bin/devstation-pipeline
light-router not starting	Rebuild and restart: docker compose -f compose/mini.yml up -d --build light-router
Pipeline "All LLM providers failed"	Check .env has DEEPSEEK_API_KEY set and DEEPSEEK_MODEL=deepseek-chat
Pipeline "Invalid format specifier"	Check for unescaped {} in f-strings inside pipeline.py agent prompt templates
LiteLLM (ai-router) crash with merged_lifespan	Remove LiteLLM and use light-router instead (see compose/mini.yml and light-router/)

Setup Session Log (May 2, 2026)

Bugs Fixed

1. devstation.sh parse_args(): Last line [ -z "$ACTION" ] && ... returned non-zero with set -e, causing immediate exit. Fix: appended ; true.
2. devstation.sh run() function: Used eval "$*" which broke with spaces/special chars. Fix: replaced with "$@".
3. Pipeline f-string format bug: Unescaped { in agent agent_pm prompt caused Invalid format specifier. Fix: escaped inner braces as HTML entities.
4. Pipeline workspace: Default /Volumes/work didn't exist. Fix: set WORKSPACE_DIR=/tmp/devstation-workspace.

Issues Resolved

1. Telegram CHAT_ID: Was wrong (8774371348 → 1121576537), causing 403 Forbidden.
2. n8n Linear Trigger: IF node array contains operator incompatible with n8n version. Replaced with Code node.
3. LiteLLM crash: ghcr.io/berriai/litellm:main-stable had upstream breaking changes (routing strategy renamed, Prisma P1012 on Wolfi Linux). Replaced with custom light-router (60-line Python proxy, 3.71 GB freed).

Pipeline Test

• PR #20 created successfully: https://github.com/thanhnhan2tn/mini-dev-station/pull/20
• Full flow: webhook → DeepSeek LLM → spec generation → git clone → branch → PR ✓

Files Changed

• bin/devstation.sh: Fixed run() and parse_args() bugs
• compose/mini.yml: Replaced ai-router (litellm) with light-router; added n8n env vars
• ai-router/config.yaml: Fixed routing_strategy: usage-based → usage-based-routing (then replaced entirely)
• pipeline.py: Fixed f-string format bug, added _safe_json_parse(), pre-computed files_md/code_files_md
• projects.yaml: Hardcoded github_repo instead of env var with default
• .env: Added DEEPSEEK_MODEL=deepseek-chat, WORKSPACE_DIR, fixed TELEGRAM_CHAT_ID
• light-router/proxy.py: Created (new)
• light-router/Dockerfile: Created (new)
• telegram-bridge/bot.py: Added polling loop (was webhook-only)
• bin/pipeline-wrapper.sh: Fixed env loading (export $(grep ...) → set -a; source .env; set +a)
• linear-pipeline/linear_to_opencode.json: Replaced IF node with Code node, hardcoded URLs