fix(slack): Stop an empty Seer mention before it runs an agent turn

[sehr-m]

Problem

_handle_seer_prompt already rejects empty text, but it checks before the bot mention is stripped — that happens later, in process_mention_for_slack via extract_prompt. A bare @Seer has truthy text "<@U0BOT>", so it passes the guard and reaches Seer as query="".

Seer then runs a full turn on nothing — an ExplorerIndex query, four context engine stages, three Vertex embedding calls and an LLM call — only to answer "It looks like your message came through empty!".

The embedding calls 400 with The text content is empty.:

path	error	events / 30d
direct google-genai	ClientError: Empty instances.	SEER-8GR — 393 (296 prod, 97 de)
llm-proxy /embeddings	Vertex 400: "The text content is empty."	LLM-PROXY-S

All 393 are explorer.category_key:slack_thread and explorer_main_task. Slack is the only entrypoint affected — the Explorer chat endpoint already enforces allow_blank=False, and every other caller templates its prompt, so query is structurally non-blank there.

Fix

Guard at the webhook, before set_thread_status("is thinking...") so no spinner is left running.

The check is a new has_prompt_content — is there anything besides mentions and whitespace — rather than reusing extract_prompt. extract_prompt needs the bot's user id, which comes from the event's authorizations; Slack omits that in some cases, and test_app_mention_dispatches_task_no_authorizations already covers it. With no id the regex degrades to <@> and strips nothing, so a bare mention would have slipped through as the non-empty prompt "<@U0BOT>".

Another user's mention still counts as context whenever real text accompanies it; only a mention-only message is empty.

Halting silently would have been a regression — the user used to get an answer, however wasteful. send_empty_prompt_message sends a fixed ephemeral nudge, mirroring send_not_org_member_message: no link button, and send failures are logged rather than raised so a Slack outage can't turn the webhook into a 500. No model, no agent run, no Vertex call.

Uses a distinct EMPTY_PROMPT halt reason rather than MISSING_EVENT_DATA, which is documented as meaning the event was malformed — a bare mention is a normal user action, and conflating them would make that metric misleading.

Test

tests/sentry/integrations/slack/webhooks/events/ + tests/sentry/seer/entrypoints/slack/ — 201 passed. mypy and pre-commit clean.

New coverage: bare mention halts and sends the nudge; halts with no authorizations; mention-only-with-other-users halts; a mention plus real text still dispatches; a failing nudge still returns 200 without dispatching.

Notes for review

• This guard is the only empty check on the Slack path. process_mention_for_slack re-derives the prompt and trigger_agent forwards it, neither re-checks — so invoking the task directly still bypasses it.
• Not fixed here: when authorizations is missing, the <@U0BOT> token still rides along into the query as prompt noise. Storing the bot user id on the integration would close that; it's a prompt-quality issue, not correctness, now that this guard doesn't depend on the id.

🤖 Generated with Claude Code

[alexsohn1126]

looks good! thank you 🙏