
CliGate simplifies approvals with task-scoped trust
CliGate's new approval model reduces repetitive permission prompts during multi-step AI-assistant jobs by introducing a task-scoped trust flag, as reported on DevTo
The CliGate project, which powers local assistants such as Claude Code, Codex CLI, and Gemini CLI, introduced a new approval model on June 12, 2026, that flips a conversation flag after the first tool call is approved, allowing subsequent steps to run without additional prompts [DevTo]. This change, part of CliGate v0.9.1, also fixed a language-mix bug where system-generated continuation messages forced later prompts into English even when the user’s original request was in Chinese [DevTo].
Previously, each tool invocation—checking for Python, launching a PDF converter, focusing a window—triggered a separate confirmation dialog. The new logic records the intent of the first approval, treats it as a task-scoped grant, and only re-asks when the user explicitly returns to safe mode (via the /safe command) or a new boundary type appears.
By collapsing approvals into a single task-level grant, the workflow cuts approval overhead by an estimated 60% for typical multi-step jobs [DevTo]. This reduction in approval fatigue is significant, as users reported spending up to 70% of interaction time clicking “yes” on trivial prompts. The task-scoped model preserves the safety barrier while avoiding the habituation trap, and the continuation-message fix ensures that the assistant respects the user’s original language throughout a session, preventing mixed-language dialogs that can confuse non-English speakers.
Subscribe to the broadcast.
Daily digest of the day's most important tech news. No fluff. Engineering signal only.
// delivered via substack · double-opt-in confirmation


