3.17.3 - 2026-09-20

Fixed

• MCP exec no longer kills a command that outlives the 30-second budget. Killing the CLI mid-operation left the work it had dispatched running inside the Editor while the agent was told the operation failed, so the natural response — retrying — duplicated it. Every command is now detached on timeout and harvested by a later exec call, not just the five project lifecycle commands (/open, /o, /new, /clone, /recent). Non-lifecycle commands report status："running" with elapsedSeconds and the in-flight command list; lifecycle commands keep reporting status："booting". (#221) • While a detached run is in flight, every exec call now reports its progress instead of spawning a competing unifocl process against the same daemon. /close still aborts it. MCP exec calls are serialized so that invariant holds under concurrent tool calls — they already shared one session seed, one remembered project, and one pending-run slot. • MCP command classification matched on a bare prefix, so /opened-elsewhere was read as /open. Triggers now have to be followed by end of input or whitespace.

Added

• get_agent_workflow_guide(section='long_running') — what booting and running mean, how to poll a detached command, and why polling cannot duplicate the work.