Observe changes

The poller makes read-only REST requests and compares a complete collection with its previous snapshot. It is useful for a scheduled study assistant, not a complete Canvas audit log. Changes that occur and disappear between polls can be missed. Polling uses the same catalog exclusions as direct reads. Canvas can update access logs and derived caches even during GET; polling is not evidence that the upstream account remains unchanged.

folava-mcp poll assignments \
  --path '{"course_id":"101"}' \
  --state ./course-101.sqlite

The first successful run establishes a baseline and prints nothing. Later runs print new events as JSON Lines. The SQLite database atomically stores hashes and events. It contains private metadata, even though coursework bodies are not saved. Do not commit it. The requested folder is created if needed; the database must be owner-only on POSIX. On Windows protect the folder with your user ACL.

Each event carries a durable cursor, event ID, query-scope hash, operation, Canvas entity ID, observation time and kind (created, changed, removed). removed means no longer visible in this successful query, not proven deletion at Canvas. The scope is bound to origin, Canvas identity, operation, path, query and catalog version. Changing filters establishes a new baseline. Only recognized volatile signature/expiry query parameters are excluded from item hashes. Meaningful query parameters and fragments remain part of the comparison.

Incomplete pagination, errors, missing IDs and duplicate IDs never replace the baseline. Overlapping polls of one state database are rejected before fetching so an old request cannot roll the baseline back. The limit is 20 pages per poll; large collections should be divided into smaller stable queries.

Replay and start an agent

folava-mcp events read --state ./course-101.sqlite --after 0 --limit 100

Schedule poll using your OS scheduler, then let a trusted automation consumer read events after its last acknowledged cursor. Persist that cursor only after the consumer has completed its work; deduplicate by event ID. Delivery is at least once. A crash after commit but before stdout is recoverable through replay. The feed is retained until you delete the database; no silent retention cutoff invalidates a consumer's cursor.

The consumer can read the changed object by ID and start its chosen agent through that host's documented interface. Keep arguments fixed and feed structured data through stdin; never interpolate course text into a shell command. This release does not execute shell hooks or send coursework to a webhook destination.

MCP does not guarantee waking an idle or closed agent. Canvas Live Events needs institutional administrator configuration and is not included. Polling plus your scheduler is the supported event mechanism. Start with a gentle cadence such as 15 minutes, watch rate-limit responses and back off when necessary.