Context graph

The context graph is bobby's memory of page structure. An agent can ask where a described control is and get a target with a confidence score, without pulling a whole accessibility tree into its context.

There are two layers:

• Session memory. Observations from the current session. Any command that may have changed the page invalidates them. Never persisted.
• Persisted memory. Per profile and per site, promoted from verified intent outcomes. It is written to disk before the command returns, so it survives restarts.

Persisted memory needs a profile with a durable identity: a paired Firefox profile, a Chromium selection with a profileId, or managed Chromium (remembered under managed-chromium while each session's browser profile stays disposable). A Firefox selection with no paired profile, and a prefer list, neither read nor write it. See Configuration.

What it stores

For each site (scheme plus registrable domain, never a full URL), page pattern (query and fragment removed, numeric path segments templated), form and control:

• Role, accessible name, ordinal and form membership.
• Per intent kind: success and failure counts, the day of the last verified success, and whether the record was observed or vision-promoted.

A completed form or extraction records every field it resolved. A failed one counts the failure against the failing step only. It never stores typed values, credentials, page text, screenshots, journal IDs or exact timestamps.

Ask before you snapshot

{"name": "context_ask", "arguments": {"workflowHandle": "wf_0123456789abcdef0123456789abcdef", "description": "Email address field"}}

A hit returns the target. A miss returns hit: false with nextStep: "a11y_snapshot". Every 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.

workflow_observe with a goal does this for you. When memory answers, it returns the remembered target with source: "retained" and takes no snapshot.

Vision ranking

When an intent escalates to vision on a page with remembered context, bobby orders the stuck step's near-miss candidates (up to 10) before the first 5 go to the provider. A remembered control counts when it matches by role and name, has the same intent kind and has more verified successes than failures. The one with the best record moves first. Ties leave the order unchanged. A record with no verification day is never used. The provider still chooses, and the action is still verified.

Retention and erasure

• Records without a verified success for [context].ttl_days (default 90) are removed when the store opens.
• bobby context list --profile <id> lists remembered sites. Use --profile managed-chromium for managed Chromium.
• bobby context forget <site> --profile <id> erases one site immediately and verifies the erasure.
• bobby doctor reports the store path, site count, size and lock health.
• Only the runtime holds the store while it runs. CLI and doctor access is read-only or refused.

Limits

The in-memory cache keeps up to 256 sites and 64 MiB per profile. Sites that leave the cache reload from disk, and eviction deletes nothing. A site file is limited to 2 MiB and 16,384 records. An oversized or unreadable file is preserved, reported, and skipped. The runtime continues with live context.

[context.limits]
max_file_bytes = 2097152
max_site_records = 16384
max_resident_sites = 256
max_resident_bytes = 67108864

All limits must be positive. bobby context --config <file> list applies the same limits.