BREAKER DELTA
浪尖レーダー
OverviewGetting StartedReading the SiteCLI ReferenceConfigurationSkill & MCPExtensionDeploy Your Own

日本語版は準備中です。このページは英語版を表示しています。

Configuration

All tuning happens in exactly one place: config.yaml in your data dir (copy the engine's config.example.yaml and edit). Load order is config.yaml > config.yml > config.json > the engine's example file. YAML is read by a built-in subset parser — no dependencies to install — and JSON stays compatible. Never hardcode thresholds anywhere else: changing industry or taste means editing this file only.

The knobs people actually turn

  • keywords.mute: a candidate whose title + summary contains any of these (substring, case-insensitive) is dropped outright. Put what you never want to see again here, e.g. giveaway, airdrop.
  • keywords.boost: matches are flagged; when grading, a single-source echo with a boost hit can be promoted to Watch. Put terms you follow long-term here, e.g. a product name, mcp.
  • keywords.alert: a match that also has a second source's resonance becomes a Must candidate directly. Keep this list very short, or Must stops meaning anything.
  • topics (weights 1–5): relevance is scored against these weights instead of a yes/no topical check. Switching industries is mostly editing this section (defaults: AI Agents 5, independent products 4, developer tools 4, content creation 3, social observation 2).
  • sources.<name>.enabled / max_items: any source can be turned off or capped individually. A disabled source is reported as disabled, never as missing, and the agent will not try to fill it in.
  • filters.max_age_hours: freshness window; only applies to candidates that carry a timestamp. Null means no filtering.
  • vocab.window_days (new-term vocabulary, default 60): on entity sources (GitHub / Hugging Face / OpenRouter / Civitai / Product Hunt / directories / App Store), a name unseen within the window is flagged novel — the highest-priority echo signal. The first run only builds the baseline and flags nothing; that is expected.
  • grading: min_sources_for_hotspot (default 2 — independent sources needed for a hotspot), max_items_per_report (default 8), max_echoes (default 3).
  • review.after_days (default 14): when a reported topic comes due for its track-record review.
  • editions / timezone / language: editions are just labels (name, time, per-edition focus). Scan frequency is yours to define — nothing about it is baked into the program or the copy. Default timezone is Asia/Tokyo.

Changing sources

All 26 fetcher groups live under sources in the config, each with its own parameters (the official-blogs feed list, GitHub's min_stars_per_day, and so on). Adding a feed, raising a limit, or switching a source off all happen here. The full catalog and each source's rules are in the repo's SOURCES.md; the site's Sources page also shows each source's effective config verbatim, so you can verify a change there.

Note: X (Twitter) is the only source that needs a token. Without one it is reported as missing and the agent fills it in by scanning the web; every other source runs keyless.