Files
sde-meeting-toolkit/docker-compose.yml
Matt Mabrey 6eb4b59d97 Fix EAI_AGAIN: work around Docker's internal-network DNS bug with a static IP
- egress-proxy gets a fixed IP (172.28.5.10) on a new 172.28.5.0/24
  subnet for the 'egress' network.
- sde-meeting-toolkit gets a static extra_hosts entry for
  'egress-proxy' instead of relying on Docker's embedded DNS, which
  breaks for internal-network names on a container that's also on a
  normal network (confirmed via a real smoke test run - EAI_AGAIN).
- Update CHANGELOG with the root cause and the subnet-collision caveat.
2026-09-29 18:52:28 -07:00

98 lines
3.9 KiB
YAML

# Standalone compose file for this tool.
#
# Network layout:
# - "proxy": external, owned by Nginx Proxy Manager's own stack. Lets NPM
# reach sde-meeting-toolkit by container name. No internet route.
# - "egress": private to this compose file. The ONLY thing
# sde-meeting-toolkit can reach besides "proxy". Used to talk to the
# egress-proxy container below.
# - "default": Compose's own auto-created network. Only egress-proxy is on
# it, since it's the only container here that needs real internet access.
#
# sde-meeting-toolkit itself is NOT on "default" and has no direct route to
# the internet at all. Its one outbound call (the AI draft button, to
# api.anthropic.com) is routed through egress-proxy via EGRESS_PROXY_URL,
# which only forwards that one destination and refuses everything else
# (see egress-proxy/tinyproxy.conf and egress-proxy/filter). A compromised
# app container still can't reach anything but Anthropic's API.
services:
sde-meeting-toolkit:
build: .
container_name: sde-meeting-toolkit
restart: unless-stopped
ports:
- "5173:5173"
environment:
- MOCK_AI=${MOCK_AI:-false}
- ANTHROPIC_API_KEY=${ANTHROPIC_API_KEY:-}
- EGRESS_PROXY_URL=${EGRESS_PROXY_URL:-http://egress-proxy:8888}
- APP_USERS=${APP_USERS:-}
- APP_USERNAME=${APP_USERNAME:-}
- APP_PASSWORD=${APP_PASSWORD:-}
- RATE_LIMIT_MAX=${RATE_LIMIT_MAX:-20}
- RATE_LIMIT_WINDOW_MS=${RATE_LIMIT_WINDOW_MS:-300000}
- LOGIN_ATTEMPT_MAX=${LOGIN_ATTEMPT_MAX:-10}
- LOGIN_ATTEMPT_WINDOW_MS=${LOGIN_ATTEMPT_WINDOW_MS:-600000}
- LOGIN_LOCKOUT_MS=${LOGIN_LOCKOUT_MS:-900000}
- AUTH_LOG_FILE=${AUTH_LOG_FILE:-./data/auth-log.jsonl}
- SIGNIN_LOG_DEBOUNCE_MS=${SIGNIN_LOG_DEBOUNCE_MS:-1800000}
- TOKEN_LIMIT_PER_USER=${TOKEN_LIMIT_PER_USER:-0}
- TOKEN_USAGE_FILE=${TOKEN_USAGE_FILE:-./data/token-usage.json}
volumes:
# Persists the per-user token quota file across restarts and rebuilds.
# Without this, everyone's daily usage silently resets on every deploy.
- ./data:/app/data
networks:
- proxy
- egress
# A container attached to both a normal network ("proxy") and an
# internal: true network ("egress") can hit a real Docker bug where
# embedded DNS stops resolving names on the internal side (EAI_AGAIN),
# confirmed while testing this. Sidestep it with a static hosts entry
# instead of relying on DNS for this one name - egress-proxy's fixed
# IP is set below via ipv4_address.
extra_hosts:
- "egress-proxy:172.28.5.10"
depends_on:
- egress-proxy
# Hardening for a container that's reachable from the public internet
# (through Nginx Proxy Manager). Neither of these needs any Linux
# capability the app doesn't already have, and the app never needs to
# escalate privilege.
cap_drop:
- ALL
security_opt:
- no-new-privileges:true
egress-proxy:
build: ./egress-proxy
container_name: sde-meeting-toolkit-egress-proxy
restart: unless-stopped
networks:
egress:
# Fixed IP so sde-meeting-toolkit can reach this by static hosts
# entry (extra_hosts above) instead of DNS, working around the
# internal-network DNS bug. Must match that entry if ever changed.
ipv4_address: 172.28.5.10
default: {}
cap_drop:
- ALL
security_opt:
- no-new-privileges:true
networks:
proxy:
external: true
egress:
driver: bridge
# Without this, Docker still gives this network a NAT route to the
# public internet (every bridge network gets one by default, "default"
# isn't special) - which defeats the whole point. internal: true means
# nothing on this network, including sde-meeting-toolkit, has any route
# out except through egress-proxy, which also sits on "default" and IS
# meant to have one.
internal: true
ipam:
config:
- subnet: 172.28.5.0/24