RagLeap
ragleap-core · release notes

ragleap-core v0.13.4

ragleap-core v0.13.4

Security fixes from an external review, and groundwork for sub-agents (off by default). Upgrading switches nothing new on.

Security fixes (an outside assessment, each finding checked by us in the code)

  • An approval runs only once (#1444). Two answers to the same pending action at the same moment (dashboard, API or chat) could both run it, or an approve and a reject could both win. The pending action is now claimed in one atomic step; the second caller gets "already processed". (F-CORE-005)
  • File actions always need approval (#1445). In full autonomy mode, POST /autonomy/execute took the action label and the channel separately, so a harmless label with a file channel skipped the forced approval. Forced approval now covers both, and a label and channel that do not belong together are refused with 400. Agents were not affected: their channel comes from the tool table. (F-CORE-004)
  • Approvals are accepted only from a private chat (#1447). In a Telegram group or a Discord server channel, everyone in it looked like the owner. Now an approval typed there is ignored, and saving a Telegram group id as the approval target is refused. See docs/approval-chats.md. (F-CORE-006)
  • Update backups stay out of the image build, and Redis credentials stay out of the log (#1446). .dockerignore lists backups/; the Redis URL is logged without its user and password. (INFRA-F-001, F-CORE-001)

Sub-agents (off by default)

  • Run tree and cancel (#1418), a delegate tool with limits (#1424) and a Runs tab that shows sub-agent runs and a Cancel run button (#1427). It needs SUBAGENTS_ENABLED=true. Limits: 3 children per run, 2 at the same time, one level deep, shared token caps. So far it is tested only with scripted models, not with a real model.

Upgrade notes

  • One additive schema change: db/schema.sql adds four columns and one allowed status (cancelled) to agent_runs. It is safe to run twice. If you run your own database, apply the new block (it starts at the comment "A run can have child runs") before you restart the app.
  • No new setting you must set. File writing, deleting and sub-agents stay off.
  • Not tested yet: a real Telegram or Discord group, separate app processes, and sub-agents with a real model. The outside security review of file writing (#800) has not happened. Nothing here makes an installation secure, compliant or certified; it helps an operator keep approvals with one person and with one run.

Thanks

Thanks to Aleksei Musienko, Chief Scientist & Founder of AMQIC (MSEC), QIAIMP Scientist, Q-Architect & Q-Developer (assessment powered by CERNOTIS), for the outside assessment, which we checked finding by finding in the code, and to everyone who tested and reported. Contributors are credited by name in the pull requests.

View on GitHub Package page All releases
← v0.13.3v0.13.5 →

Release notes mirrored from GitHub Releases.