@rtrvr-ai/site-intel
v0.2.9
Published
Runtime-agnostic core for turning captured web traffic into ranked API endpoints, replayable recipes, and trust-scored agent capabilities. Single source of truth shared by rtrvr-relay, extension-use, rtrvr-web-agent, rtrvr-cloud-backend, and rover-scout-w
Maintainers
Readme
@rtrvr-ai/site-intel
Runtime-agnostic core for turning captured web traffic into ranked API endpoints, replayable recipes, and trust-scored agent capabilities.
This is the shared foundation described in the Rover Scout ground-up revamp. It exists so that the network-tool "capture → evidence → cluster → classify → provenance → recipe → verify → trust" pipeline lives in exactly one place instead of the four divergent copies that exist today (rtrvr-relay original, rtrvr-cloud-backend prompt side, extension-use snapshot, rtrvr-web-agent backend half, plus rover-scout-worker's own lineage).
Layout & runtime scoping
Entrypoints are runtime-scoped so a browser bundle never pulls Node/CDP code:
@rtrvr-ai/site-intel/core— isomorphic. No capture, no LLM calls, no CDP. Pure functions over theNetworkExchangemodel: templating, clustering, scoring (rank-only, no silent drop-gate), response-shape hints, JSON-Schema inference, secret redaction, the acorn AST runner validator, and the Unbrowse-style composite trust score.node:cryptois used for stable hashing; browser consumers polyfill it (a P10 concern when relay/ extension-use migrate).@rtrvr-ai/site-intel/capture— capture adapters behind aCaptureAdapterseam (HAR today; webRequest / main-world / CDP adapters land as the other repos migrate)..(root) re-exports both.
Where this package lives
It is a workspace package inside the rover-scout-worker monorepo so that Scout — its
first consumer — can iterate against it with a pnpm workspace link. It publishes to npm
(like @rtrvr-ai/rover-scout-core) so rtrvr-relay / rtrvr-web-agent / extension-use can
depend on the published version later. Nothing about the public API assumes this location;
it can be extracted to its own repo without an API change.
Core principle: static tables are EVIDENCE, never GATES
The cheap provider/keyword/URL heuristics survive here as advisory signals attached to endpoint clusters. They rank and tag; they never silently drop a candidate. "Nothing found" is a quality-gate signal for the caller to surface, not an empty success.
