Context graph

The context graph is bobby's memory of page structure. It exists so an agent can ask "where is the control described as X" and get a bound target with a confidence score instead of pulling a whole accessibility tree into its context.

Two layers:

• Session-hot — observations from the current session, invalidated on any

command that may have changed the page. Always available, never persisted.

• Persisted — per-profile, per-site structural memory promoted from

verified intent outcomes. Only runtimes whose engine selection carries a durable profile identity (a Firefox companion enrollment) write or read this layer. Chromium sessions have disposable profiles and no durable identity, so they read nothing and write nothing — by design, not by accident.

What persists

Per site (keyed by scheme + registrable domain, never a full URL), per page pattern (query/fragment stripped, numeric path segments templated), per form, per control:

• role, accessible_name, ordinal, form membership
• Per intent kind: success/failure counters, the day of the last verified

success, and how the record entered the graph (observed or vision-promoted)

Never persisted: typed values, credentials, page text, screenshots, journal ids, exact timestamps. Timestamps are day-precision by construction. The CI privacy canary (context_privacy) fills a form with a canary value through the live harness and scans every byte of the store for it.

Provenance

Every context_ask answer says where it came from: observedAt is a live page generation or persisted, and remembered answers carry their source. A remembered answer never claims to be a live observation.

Retention and erasure

• Records not verified within [context].ttl_days (default 90) are swept at

store open.

• bobby context list --profile <id> shows remembered sites.
• bobby context forget <site-key> --profile <id> erases one site

immediately and totally, and verifies the erasure before reporting.

• bobby doctor reports the store path, site count, bytes, and lock health.
• The store is single-writer: only the runtime process holds it. CLI and

doctor access is read-only or refused while the runtime runs.

Reading

• MCP context_ask (requires page:read) — live first, persisted fallback.
• MCP context_neighbors (requires context:read) — the remembered form

structure around a located control.

• HTTP GET /v1/context/ask and GET /v1/context/site/{key} (both require

context:read).

On a known site, ask before you snapshot: context_ask answers before the first accessibility observation of a session.