* fix(agent-launch): an AI button plans its agent inside its own pane's spawn, with the host's planner and no launch record
The window queues the launch on the tab; the pane's reserved spawn plans it with the planner every
host launch uses and starts it at the pane's size, with its launch token. Every way the spawn ends
finishes the launch, and the window pastes the prompt only once this pane says the agent started.
An ordinary spawn reaches its provider without waiting on a plan. An OpenCode tab with a model
keeps main's host hop untouched (TEMPORARY, until a local producer of model options exists).
* fix(agent-launch): an AI button's prompt waits for its pane to start before the spawn budget runs, so a pane that settles slowly still gets its prompt; stale comments
* test(agent-launch): the late pane-start test's store double is typed without an assertion
* test(agent-launch): the late pane-start test pins the prompt written once and Enter once, and nothing when the pane didn't start
* fix(agent-launch): an AI button's prompt goes only to the PTY its pane started, never a split's shell that bound first; the pane's start answer names that PTY
* fix(agent-launch): an AI button's launch watches each split's readiness as it binds, so the launch pane's own ready output is seen even when a split bound first and it drained at once
* fix(agent-launch): an AI button's pane plans with the host's current planner and runs its runtime fact's shell, so its command is quoted for the shell it is typed into, as a host launch is; the tab's inferred shell is not sent as a request
Brennan Benson
a0de43aa1f