- search_runs.json: append-only history of every searcher run
(started_at, finished, added, seen, platforms, lookback_days)
- search_progress_last.json: snapshot of final progress state after
each completed run — answers 'what keywords/tracks were searched?'
- filter_runs.json: append-only history of every filter batch
(batch_id, submitted/collected timestamps, model, passed/filtered/errors)
Fixes the 'did the 90-day run complete?' ambiguity going forward
- Batch API = 50% cost savings vs synchronous calls
- Prompt caching on system prompt (profile + criteria shared across all jobs)
- One request per job with custom_id = job ID for result matching
- Two-phase state machine: submit → poll/collect (hourly cron safe)
- filter_state.json tracks pending batch ID between runs
- Model configurable via settings.filter.model (default: claude-sonnet-4-6)
- Telegram notifications on submit + collect
- Errors pass through — never block applications due to filter failure
- --stats flag for queue overview
- lib/filter.mjs: batch scoring engine (10 jobs/call, Claude Haiku)
- job_filter.mjs: standalone CLI with --dry-run and --stats flags
- Threshold configurable globally + per-search in search_config.json (filter_min_score, default 5)
- Job profiles (gtm/ae) passed as context via settings.filter.job_profiles
- Filtered jobs get status='filtered' with filter_score + filter_reason
- Filter errors pass jobs through (never block applications)
- status.mjs: added 'AI filtered' line to report
