@appliqation/agent-core

The shared engine every standalone Appliqation agent is built from. Not an app on its own — no CLI, no .env — just the generic, workflow-agnostic pieces that appliqation-autotest, appliqation-scriptgen, appliqation-defect-fix, appliqation-pr-raise, appliqation-explorer, and appliqation-autopilot each build on:

• /engine — the think→act→observe tool-calling loop, budget tracking (calls/pages/time/turns), and a workflow runner that fetches a named prompt (from Appliqation's MCP server or a local file) and runs it through that loop.
• /appq — a JSON-RPC client for Appliqation's MCP API, plus scenario/test-set/run resolvers that always derive project_id/url from a source of truth (a scenario ID, a TC UUID) rather than accepting a second, possibly-diverging input.
• /providers — Anthropic and OpenAI adapters implementing one common interface, including prompt caching and image-handling for both wire formats.
• /tools — a Playwright-backed browser-tool palette (multi-tab aware — browser_tabs/browser_resize/browser_evaluate), an HTTP client for API test cases (http_request, with dry-run write-verb suppression), the tool-allowlist gating mechanism every agent's own safety boundary is built on, a destructive-action click gate, an argument-level read-only gate for Appliqation's project-context document, auth-session resolution (UI storageState and API credentials), and per-test-case role/test-type inference.
• /evidence — screenshot/console/network/accessibility-snapshot capture with cursor-based delta reads.
• /audit — optional, best-effort run logging (AuditRecord/AuditSink, MongoDB and local-JSONL-file implementations) that every sibling agent can opt into, read back by appliqation-dashboard. A write failure here never affects the real run it's describing.

Design principle: mechanism shared, domain knowledge local

Every module here is either genuinely universal, or an explicit injection point so a consuming agent can supply its own domain knowledge without this package needing to know about it. Two examples:

• Tool-dispatch gating (tools/gatedDispatcher.ts) enforces "only call tools in this allowlist" — but the allowlist itself (which tools an executor vs. validator vs. any other stage may touch) is each consuming agent's own knowledge, kept in that agent's own code, never here.
• Browser tools (tools/browserTools.ts) take optional {onBeforeClick, screenshotSink} hooks instead of hardcoding a destructive-action policy or an upload destination — a consuming agent with no write access, or a different safety posture, supplies its own.

Installation

npm install @appliqation/agent-core

"dependencies": {
  "@appliqation/agent-core": "^0.1.0"
}

Requires playwright and @appliqation/automation-sdk as peer dependencies (already a dependency of every consuming agent in this family).

Development

npm run build
npm run typecheck
npm test

See CLAUDE.md for a map of this repo if you're working in it with an AI coding assistant.

License

MIT — see LICENSE.