Skip to content

v3.33.3

Choose a tag to compare

View all tags
@goruck goruck released this 29 Aug 12:16
· 9 commits to main since this release
v3.33.3
b536214
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Fixed

  • The agent no longer learns to stop controlling your home. When a command was handled by another assistant sharing the conversation — Home Assistant's own built-in one, for instance — the agent read that turn back as if the assistant had simply said "Turned on the light" without doing anything. That is a perfect example of the wrong behaviour, and the model copied it: the next command got a confident spoken reply and no action, which was then read back as another example, and so on. Nothing in the log looked wrong, because from the agent's side nothing failed. The whole of a turn that used tools is now kept out of the history rather than just the part that names the tool, so a reply that no longer shows its actions can't be mistaken for one that never needed any (#588).

    This is more likely to bite smaller local models, which follow the shape of a conversation more readily than the large cloud ones. Conversations that already went wrong stay wrong — the old messages are saved with the conversation. Start a new one and it will be clean.

Verified against the live model with history as the only variable: 0/6 tool calls with the affected history, 6/6 with the history this release produces. Field-validated on Home Assistant 2026.8.3 with qwen3.8 via Ollama.

Full Changelog: v3.33.2...v3.33.3

Assets 3
Loading