Release graphsense-lib 2.13.3 · graphsense/graphsense-lib

[2.13.3] - 2026-05-18

tagpack sync keeps a persistent git repo cache instead of re-cloning every run. Previously each repo in the sync list was cloned afresh into a throwaway temp directory. Synced repos are now kept under a stable cache directory and refreshed with a git fetch (_sync_repo/_repo_workdir in src/graphsenselib/tagpack/cli.py); only changed objects are downloaded. The refresh is authoritative — fetch + checkout + reset --hard + clean -fdx — so a reused checkout is identical to a fresh clone, and a missing/corrupt/wrong-remote cache is transparently re-cloned. New --repo-cache-dir option overrides the location (defaults to a graphsense_tagstore_sync_repos folder in the system temp directory). Non-breaking: no existing option changed and sync results are unaffected.

Delta-update hardened against Cassandra-outage corruption mid-flush. DbWriterMixin.apply_changes now also retries Unavailable and NoHostAvailable (previously only WriteTimeout/OperationTimedOut), so a recoverable database outage is waited out — bounded by stop_after_attempt — instead of aborting a half-flushed batch and leaving the keyspace inconsistent; statements are bound once with literal values, so retries are idempotent. Additionally, pure auto-resume runs (no explicit --start-block/--end-block) now refuse to continue when summary_statistics points at a block with no matching delta_updater_history row — a sign the last bookkeeping write was torn — turning a silent double-count/skip into an actionable stop. New TransformedDb.delta_updater_history_has_block point-read backs the guard, which can be bypassed with --disable-safety-checks. Changes in src/graphsenselib/db/analytics.py and src/graphsenselib/deltaupdate/deltaupdater.py.

API docs now describe the MCP (Model Context Protocol) interface. A non-technical "AI assistant access (MCP)" section was added to the API description shown at the top of the Swagger UI / ReDoc pages (API_DESCRIPTION in src/graphsenselib/web/app.py). It explains, for a general audience, that the same deployment exposes an MCP endpoint at the /mcp path which lets AI assistants query GraphSense in natural language, and how to connect one. Docs-only and additive — no endpoint, schema, or generated-client behaviour changed.

Full Changelog: v2.13.2...v2.13.3