Repository navigation
Architecture
This page is generated from
docs/ARCHITECTURE.md. Edit the source in the repository and the change will be synced here.
ReboundOps is one Node.js process on your computer. It serves an Angular single-page app and a JSON API from 127.0.0.1, and it keeps everything in one SQLite database in a private folder. This guide describes the repository at v0.2.0; the package list was read from each package.json.
flowchart LR
subgraph Machine["Your computer"]
B["Browser tab<br/>http://127.0.0.1:4175"]
subgraph P["Node.js process (apps/api, Fastify)"]
G["Guards: loopback host, Origin,<br/>JSON-only writes, body limits"]
R["Routes + zod validation<br/>(packages/contracts)"]
L["Domain packages"]
S["packages/storage<br/>node:sqlite"]
end
DB[("~/.reboundops<br/>0700 folder, 0600 files")]
O["Ollama (optional,<br/>loopback only)"]
end
J["Public job boards<br/>(after consent)"]
C["Cloud AI provider<br/>(opt-in, per-request consent)"]
B -- "same-origin fetch /api" --> G --> R --> L --> S --> DB
L -. "packages/ai-providers" .-> O
L -. "packages/job-sources" .-> J
L -. "packages/ai-providers" .-> C
Dotted lines are optional, and each one needs the user's action or consent. With AI off and Discover unused, no request leaves the machine after installation.
flowchart TD
web["apps/web<br/>Angular UI"] --> contracts
api["apps/api<br/>Fastify server"] --> contracts & core & ji & docs & apps & importer & sources & ai & storage
core["packages/core"] --> contracts
ji["packages/job-intelligence"] --> contracts
docs["packages/documents"] --> ji
apps["packages/applications"] --> ji
importer["packages/importer"] --> ji
sources["packages/job-sources"] --> ji
ai["packages/ai-providers"] --> contracts
storage["packages/storage"] --> contracts
contracts["packages/contracts<br/>zod schemas"]
apps/web depends only on packages/contracts, and reaches everything else through the HTTP API. Domain packages never import Fastify, Angular or SQLite, so they can be tested on their own.
| Package | Responsibility | Notable third-party runtime dependency |
|---|---|---|
apps/api |
Server start-up and configuration, guards, routes, static serving with CSP, restore, the discovery orchestration |
fastify, zod
|
apps/web |
Angular UI: standalone components, signals, zoneless | Angular, rxjs
|
packages/contracts |
Shared zod schemas: the API contract and the import format | zod |
packages/core |
Workflow rules (status transitions, Applied needs confirmation), dashboard, analytics, follow-up ranking | none |
packages/job-intelligence |
Field extraction with evidence, deduplication, grounding checks, the fit rubric and eligibility | none |
packages/documents |
Resume import (PDF, DOCX, text), tailoring, cover letters, PDF and DOCX writers, export read-back |
pdfjs-dist (PDF text reading) |
packages/applications |
Screening answers from saved facts, the checklist, follow-up drafts | none |
packages/job-sources |
Greenhouse, Lever, Ashby and Workable adapters, and a polite HTTP client (allow-list, rate limit, cache) | none |
packages/importer |
Import planning: validation, deduplication, no-overwrite profile merge | none |
packages/ai-providers |
Optional Ollama, Anthropic, OpenAI, Gemini and OpenRouter adapters, grounded assist tasks, consent checks | @anthropic-ai/sdk |
packages/storage |
SQLite persistence, append-only migrations, backups (optionally encrypted), crash-safe restore, data-folder lock | none (node:sqlite is built in) |
Job intake → assessment. The API validates the pasted text against a contract schema. job-intelligence extracts fields together with the source text that supports each one; anything unsupported stays empty. The posting is stored as unverified. Assessments are computed on demand from the stored posting and the verified profile, and are never stored, so a rubric change (fit-2026.1 → fit-2026.2) takes effect immediately.
Tailoring → export. documents selects and orders only highlights the user has marked verified, and uses [[placeholders]] for gaps. Exports are written to PDF and DOCX, then read back to check that employers, dates, education and links survived.
Discover. After consent, job-sources calls one board's documented public endpoint. Requests go only to allow-listed HTTPS hosts, without cookies, at no more than one request per second per host, with a 15-second timeout, a 25 MB cap and a 30-minute cache. Results are labelled with the time the board last listed them.
AI assist. Each task sends only the fields it needs, and only verified profile entries. Its output is compared with the user's records, and any suggestion that introduces new facts is withheld. Nothing is saved without the user's action. Ollama must be on a loopback address.
Backup and restore. A backup is a full database snapshot, optionally encrypted with scrypt and AES-256-GCM. Restore stages and validates the backup, checks free disk space, writes a recovery copy, journals the swap, and rolls back automatically if it is interrupted.
- TypeScript runs natively on Node 24 or newer (type stripping); only the Angular app has a build step.
- Vitest covers packages, the API and integration tests; Angular tests include axe accessibility checks; Playwright runs the end-to-end journey.
-
scripts/holds the smoke test, export verification, the release audit (paths, secrets, licences, SBOM,npm audit, CSP), and the reproducible packaging and verification scripts. - CI (
.github/workflows/ci.yml) runs theverifyande2ejobs on Ubuntu 24.04 with Node 24.