Repository navigation
v0.1.3 — one panel per app
Fix — a workspace up opened two panels. The TUI picker dispatched every rundev command twice:
session.command resolves to void, so the guard res !== undefined never returned and the CLI
fallback always fired as well. One picker confirmation ran the verb twice — two flutter run splits
typed into the focused terminal, 3s apart (the server event log shows two up: app runs from a single
select event). The guard now checks that the client exposes the method, and the CLI path stays as
the fallback for an unreachable server.
Fix — interactive services were never idempotent. statusOf promised to report a panel as
running when a record existed, but nothing ever wrote one (the pidfile is only written for process),
so every up reported unknown and relaunched the panel. rundev now records the panel it opens and
down forgets it:
- a second
upreportsalreadyinstead of opening a duplicate panel; up <svc> --forcerelaunches on purpose (also exposed as aforcein the agent tool), and an
explicitapp@iosover an android panel retargets instead of answeringalready;check— ortargets.<t>.check, for a service that switches device — makes a panel verifiable:
a closed app is detected anduprelaunches by itself;- the marker lives in its own file (
<name>.panel.json):stopProcesskills by pidfile, andpid: 0
would have addressed the whole process group. A defensive pid guard covers that path too; logsfor an interactive service no longer claims "no log yet": the panel owns its output.
12 new smoke assertions cover the marker, --force, the retarget, the .env guard and the per-target
check. Prefer a verifiable setup when you write a manifest:
Upgrade
opencode service restart # re-resolves opencode-rundev@latest