Weekly suggestions from your interest list

What is sent: to make the suggestions, the jobs you marked (title, company, JD text and your status for each) and your preferences are sent to the AI provider(s) you have switched on. Switch the weekly run off with suggest.enabled: false in schedule.yaml. See privacy.md.

Once a week, the AI looks at the jobs you marked Interested / Preparing / Applied / Interviewing / Offer (and the ones you Ignored, for contrast) and makes three kinds of suggestion: new signal words (T2 / T3 / T4 / benefit tags), companies to watch, and filters (exclude keywords, add cities, exclude companies). Suggestions are only proposals, shown in the “Learned suggestions” block of /interests; only when you click approve are they written to data/config/interests.yaml or data/profile/preferences.yaml, and if you do not approve nothing changes.

How it runs

• The runner, per suggest.cron in data/config/schedule.yaml (default Monday 08:00, in timezone), leaves a run_suggest request (control request), which is executed by the process that runs AI (the runner, or the ai-worker on the host). If at runner start the most recent one was missed and catch_up is on, it adds a request to catch up.
• The “Find suggestions now” button on /interests can trigger it manually at any time; the result can be seen in the recent requests list of /settings.
• Command-line check: run pnpm suggest:smoke --confirm-write <db name> against a copy of the database (it writes interest_suggestions, changes no config file, and sends no Telegram message).
• With fewer than min_positive_jobs (default 3) liked jobs the AI is not called, and /interests shows “Too few interesting jobs”.

Every suggestion comes with evidence

• Signal words: must appear in at least min_evidence.signal_word jobs you liked, each with a verbatim JD quotation; the quotation must be a fragment that literally exists in that job’s JD (whitespace ignored), and the word itself must be in the quotation; a suggestion whose quotation cannot be found is dropped entirely. Words already on the list (or already covered by some pattern) are not suggested again; words that are as common in ignored jobs as in liked jobs are not suggested either.
• Watched companies: attached are the jobs of that company that you liked (company names are matched after removing GmbH and applying aliases), and what is stored is the company name in the database.
• Filters: an exclude keyword must appear in jobs you ignored and no liked job may contain the word (a hard filter would filter those out too); an added city must be the city of a liked job; an excluded company must not be a company you have liked.
• Suggestions where the AI’s confidence is below min_confidence are dropped; at most max_suggestions are stored each time (high confidence first).

No repeats

Every suggestion has a dedup key (words are folded by case and whitespace, regardless of tier; companies by normalised name). A suggestion that is pending, approved or rejected is never proposed again. Rejected = never shown again.

After approval

• It is written back with writeConfig (YAML comments preserved, atomic replace, zod validation). Words go into the keywords of the matching tier (benefit tags into benefit_tags), companies into watched_companies, filters into preferences.yaml (a snapshot of the profile is saved, and the related scores are marked stale and re-scored).
• Ranking values are recomputed immediately (reconcileBonuses, no AI call); new words only start to score once the next cycle’s signals step rescans the JDs with the new word list (the page says so).
• When the config file cannot be read or written, the suggestion goes back to pending and nothing is lost.
• If the value is already on the list (you added it manually): it is marked approved and the file is not changed.

Telegram

One message per week when there are new suggestions (both suggest.notify and telegram.enabled in notify.yaml are on; during quiet hours it is not sent, and the suggestions are still on the page). The message lists the first few, with a link to /interests (needs public_base_url in notify.yaml).

Settings (suggest in schedule.yaml)

enabled, cron, lookback_weeks, positive_statuses, ignored_statuses, min_positive_jobs, max_positive_jobs, max_ignored_jobs, jd_max_chars, max_suggestions, min_confidence, min_evidence, max_term_chars, notify. These have no settings-page form; edit the YAML directly (the runner hot-reloads).