Configuration

All settings are environment-style NIMI_* variables. The complete list with defaults is in the settings reference; nimi-service/env.example is a commented template.

Where settings come from

Later sources override earlier ones:

1. built-in defaults
2. ~/.config/nimi/service.env — outside the repo (created by db/bootstrap.sh)
3. nimi-service/.env — inside the repo folder, git-ignored (easy to find)
4. real environment variables

NIMI_ENV=/path/to/file replaces 2 and 3 with exactly that file. To check what is in effect and where each value came from (secrets masked):

.venv/bin/python -m nimi_service config

Database

Connection details are separate settings — not a URL — so credentials can be injected from a secrets manager (Vault Agent, Kubernetes secrets, …):

Setting	Default
NIMI_DB_HOST / NIMI_DB_PORT / NIMI_DB_NAME	127.0.0.1 / 5432 / nimi
NIMI_DB_USER	nimi_service
NIMI_DB_PASSWORD or NIMI_DB_PASSWORD_FILE	unset: libpq falls back to ~/.pgpass / trust / peer auth
NIMI_DB_SSLMODE	unset (e.g. require for managed Postgres)
NIMI_DB_SKILL_USER	nimi_skill_reader
NIMI_DB_SKILL_PASSWORD or NIMI_DB_SKILL_PASSWORD_FILE	unset

Which Claude account NIMI uses

Triage and RCA runs start claude with a clean environment: variables such as ANTHROPIC_API_KEY or ANTHROPIC_BASE_URL exported in your shell for other work are not passed through. Only these settings decide:

Setting	Effect
NIMI_CLAUDE_CONFIG_DIR	A separate Claude login and user config, e.g. ~/.claude-nimi (log in once with CLAUDE_CONFIG_DIR=~/.claude-nimi claude). Unset = your default ~/.claude. Keep it outside the repo: it holds login tokens.
NIMI_ANTHROPIC_API_KEY_FILE	Use an API key read from this file (API billing) instead of the stored login.
NIMI_CLAUDE_SETTING_SOURCES	Settings layers RCA runs load: default user,project. project ignores your user-level settings, skills and plugins. Triage never loads any.
NIMI_CLAUDE_PASS_ENV	Extra variables to pass through, e.g. HTTPS_PROXY,NO_PROXY.
NIMI_CLAUDE_BIN	The claude executable (default: claude on PATH).

With nothing set, everything uses your normal login and ~/.claude config. Your organization's managed Claude Code policy always applies on top.

Settings you are most likely to change

Setting	Default	Meaning
NIMI_POST_ENABLED	false	Post replies to chat. Off = dry run (results are stored, not posted).
NIMI_POST_TARGET / NIMI_POST_WEBHOOK	thread / unset	Reply in the alert's thread, in a separate space via an incoming webhook, or both (Where replies go).
NIMI_ADMIN_WEBHOOK	unset	Incoming webhook of the admin's chat space for red flags.
NIMI_ALERT_SENDERS	any bot	Comma-separated sender IDs allowed to trigger RCAs.
NIMI_STALE_HOURS	6	Alerts older than this when NIMI can process them get no RCA.
NIMI_MAX_PER_DAY / NIMI_MAX_PARALLEL	50 / 2	Daily cap on RCA runs (UTC) / concurrent runs.
NIMI_TRIAGE_MODEL	haiku	Model for triage.
NIMI_SKILL_HISTORY	true	Let RCA runs query past messages and RCAs.
NIMI_REPLY_DISCLAIMER	built-in	The disclaimer on the first line of every RCA message (👁️ NIMI — <disclaimer>), e.g. to say where NIMI runs.

Secrets

Secret values (NIMI_DB_PASSWORD, NIMI_DB_SKILL_PASSWORD, NIMI_ADMIN_WEBHOOK, NIMI_POST_WEBHOOK) are masked by config and never logged. Prefer the *_FILE variants with a secrets manager. Never put real values in env.example: it is committed.