Red flags

NIMI doesn't try to redact free text. It detects, records and alerts a human instead. Flags are stored in red_flags (deduplicated, and never containing a sensitive value) and sent to the admin's chat space via NIMI_ADMIN_WEBHOOK.

What raises a flag

Source	Raises a flag when	Kind
scanner (code, before any model)	a high-confidence secret format is in the message: private keys, cloud/API keys, tokens, password=…, credentials in URLs. Critical hits are held from all models until release.	sensitive_data
triage model	it reports sensitive data in the message (types only)	sensitive_data
RCA skill	it reports sensitive data in the input, code, config, logs, docs or history (sensitive_data: in RCA_FIELDS)	sensitive_data
guard hook	it blocked a tool call (logged to ~/.local/state/nimi/guard-denials.jsonl, matched to the run by session ID). Formatting slips are recorded as low without a notification.	rule_violation
service	triage failure, RCA error or timeout, posting / ingest / reaction failure, daily cap reached, Claude unavailable	error

Repeats of the same problem bump occurrences on the open flag instead of notifying again.

Replies are still posted

Sensitive data doesn't block replies (except critical secrets, which are held before any model sees them):

• the skill never writes sensitive values, only their type and location;
• outgoing text is masked for high-confidence formats as a backstop (and flagged if anything was masked);
• a visible note asks the alert owner to remove the data from the source and rotate exposed credentials.

Handling flags

.venv/bin/python -m nimi_service red-flags                     # open flags
.venv/bin/python -m nimi_service red-flags --all               # include resolved / ignored
.venv/bin/python -m nimi_service red-flags resolve 12 --note "rotated the key"
.venv/bin/python -m nimi_service red-flags resolve 13 --ignore --note "test data"
.venv/bin/python -m nimi_service release 42                    # send a held request to triage anyway