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。

Getting Started

First, keep two things apart: the engine is code (fetching, wave-height scoring, site generation); the data belongs to your instance (editions, the topic ledger, the track record, your config). The engine repo stores no instance data; everything you produce lives in your own data directory. Reading the public site needs no install at all — this page is for people who want to run their own.

Requirements

  • Bun ≥ 1.2. No other runtime dependencies, no database, no server.

Step 1: get the engine and bundle it

git clone https://github.com/daiyanze/breaker-delta.git
cd breaker-delta
bun install
bun run bundle        # builds apps/cli/dist/breaker-delta.mjs — a single file

Every command runs through that single file; below it is called BD:

BD="bun apps/cli/dist/breaker-delta.mjs"

Step 2: prepare a data directory

The fastest route is to copy the shape of the working data repo daiyanze/breaker-delta-data (or clone it and replace the data with yours). A data directory looks like this:

your-data/
  config.yaml        # your config (copy the engine's config.example.yaml and edit)
  editions/          # one result file per edition
  state/             # topic ledger, vocabulary and sitemap baselines (maintained by the program)
  track-record.md    # the pick ledger and its 14-day review results

Step 3: check, scan one edition, build the site

$BD --data-dir /path/to/your-data doctor                    # offline self-check
$BD --data-dir /path/to/your-data scan --edition morning    # fetch one edition's candidates
$BD --data-dir /path/to/your-data publish --out site        # build the static site
  • Run doctor first: it checks Node, the config, the data dir and the engine root offline, names any problem, and exits 1 on failure.
  • scan is deterministic fetching only: it writes candidates to state/candidates-<date>-<edition>.json and reports per-source status. Grading candidates into Must / Watch / Echo is the agent's job under the methodology (see Skill & MCP); graded picks are written back with record.
  • publish turns the editions and ledger in the data dir into a complete static site, ready to deploy as-is.
  • Every command accepts --json for machine-readable output; running with no command is the same as scan.

Next: tune sources and keywords to your field in Configuration, and put the site on GitHub Pages with Deploy Your Own.