- New egress-proxy/ container (tinyproxy) that only allows CONNECT to api.anthropic.com, refuses everything else. - docker-compose.yml: sde-meeting-toolkit is no longer on 'default' at all. Its only networks are 'proxy' (NPM, inbound) and a new private 'egress' network shared only with egress-proxy. egress-proxy is the only container with a real internet route. - server.js: route the one outbound Anthropic call through EGRESS_PROXY_URL when set, using undici's own fetch + ProxyAgent together (mixing them with Node's global fetch throws a version mismatch error - caught this locally before it hit a real deploy). - Add undici as a real dependency (package.json, package-lock.json), npm ci step in Dockerfile. - Add smoke-test-egress.sh + Step 9b in DOCKER_TESTING.md to verify this whole chain with real Docker. - Update CHANGELOG.
64 lines
2.9 KiB
Plaintext
64 lines
2.9 KiB
Plaintext
## Copy this file to .env and fill in real values. Do not commit .env.
|
|
|
|
# Set to true to test the AI draft button with no API key and no cost.
|
|
# Returns a canned draft instead of calling Anthropic. Leave unset or false
|
|
# for real drafts.
|
|
MOCK_AI=false
|
|
|
|
# Required for the AI draft button, unless MOCK_AI=true.
|
|
ANTHROPIC_API_KEY=your-anthropic-api-key-here
|
|
|
|
# Only meaningful in the Docker deployment (docker-compose.yml sets this
|
|
# automatically to http://egress-proxy:8888). When set, the one outbound
|
|
# call this app makes (to api.anthropic.com) is routed through that proxy
|
|
# instead of going out directly, and the container itself has no other
|
|
# route to the internet at all. Leave unset for local, non-Docker runs -
|
|
# the AI draft button then calls Anthropic directly, same as always.
|
|
EGRESS_PROXY_URL=
|
|
|
|
# Named users. Comma-separated user:password pairs. Each name doubles as the
|
|
# identity used for the per-user token quota below.
|
|
# Example: APP_USERS=alice:pass1,bob:pass2,carol:pass3
|
|
# Leave blank for solo local use with no login prompt (usage is then tracked
|
|
# under the identity "local").
|
|
APP_USERS=
|
|
|
|
# Legacy single shared login. Only used if APP_USERS is blank. Everyone who
|
|
# logs in with this pair shares one identity and one token quota. Prefer
|
|
# APP_USERS instead: a shared login means the audit log below can't tell
|
|
# which real person made a given request.
|
|
APP_USERNAME=
|
|
APP_PASSWORD=
|
|
|
|
# Login lockout (brute-force protection). After this many failed logins from
|
|
# one address within the window, that address is locked out for the lockout
|
|
# period, even with the correct password. Only applies when a login is
|
|
# required (APP_USERS or APP_USERNAME/APP_PASSWORD is set). Defaults: 10
|
|
# failed attempts per 10 minutes locks out for 15 minutes.
|
|
LOGIN_ATTEMPT_MAX=10
|
|
LOGIN_ATTEMPT_WINDOW_MS=600000
|
|
LOGIN_LOCKOUT_MS=900000
|
|
|
|
# Sign-in audit log. One line per login event (timestamp, source address,
|
|
# username, success or failure) - never the password. Failures are logged
|
|
# individually; successes are logged at most once per address+user within
|
|
# SIGNIN_LOG_DEBOUNCE_MS, since Basic Auth resends credentials on every
|
|
# request and logging each one would flood the file. Persists across
|
|
# restarts if AUTH_LOG_FILE's folder is a mounted volume.
|
|
AUTH_LOG_FILE=./data/auth-log.jsonl
|
|
SIGNIN_LOG_DEBOUNCE_MS=1800000
|
|
|
|
# Optional. Caps AI draft calls per source IP, to catch a runaway script
|
|
# fast. Defaults: 20 requests per 5 minutes.
|
|
RATE_LIMIT_MAX=20
|
|
RATE_LIMIT_WINDOW_MS=300000
|
|
|
|
# On the back burner: off by default. Caps combined input+output tokens per
|
|
# named user, per UTC calendar day. Leave at 0 (or unset) for unlimited,
|
|
# which also skips writing TOKEN_USAGE_FILE. Set a positive number (for
|
|
# example 50000, roughly 100-200 AI drafts with this tool's prompt size) to
|
|
# turn it back on. Usage then persists across restarts if TOKEN_USAGE_FILE's
|
|
# folder is a mounted volume.
|
|
TOKEN_LIMIT_PER_USER=0
|
|
TOKEN_USAGE_FILE=./data/token-usage.json
|