Vorno 0.22.0-beta.8

Features

Added opt-in decision-model controls for Guarded permissions, turn outcomes, adaptive thinking, mid-turn message handling, titles, suggestions, large results and automation conditions. They remain off by default. (from upstream v0.14.0)
Added decision-assisted task verdicts and repair selection, semantic label rules, and decision usage reporting. (from upstream v0.14.0)

Remote decision settings follow the executing host. Settings and keys apply across that host's workspaces. Authenticated host administration is not workspace-isolated. See ADR-0034. #232, abba04a4.
Large Claude MCP results now pass through a pre-model guard. Removed redundant summarization of the displayed result copy. (from upstream v0.14.0)

Tightened read-only command detection, remembered permission scopes and spawned-session permission ceilings. (from upstream v0.14.0)
Queued messages retain crash recovery. Decision-driven steering and merged continuations preserve durable replay markers. Automation history distinguishes skipped conditions from executions. #232, abba04a4, 9b1e4669.
Guarded writes check real paths. Symlink escapes prompt before outside-project writes. Disabling the feature during a risk check restores Ask permissions. Failure-handler sessions retain matcher identity and chain depth, and late title/label decisions discard results after shutdown. #232, a0ff3f51.
Semantic labels trigger automations. New semantic labels notify the existing metadata differ. The full-server Docker install also retains the workspace manifests required by its frozen lockfile. #232, 5e3d26c3.

Headroom compression and retrieval still apply to the host-side tool-result copy. Claude model-input Headroom compression is separate follow-up work; displayed compression statistics do not establish model-input or billing savings.