Repository navigation
codeSplitter: rebuilds can still wait for an AgentSandbox container behind feature runs #1242
Description
#1241 puts fix tasks on a FIXES queue of their own, so they no longer queue behind feature runs. But AGENT_SANDBOX is maxInstances: 2 (apps/codesplitter/rustybuns.config.ts). If two feature runs hold both containers, a rebuild that has been picked up still waits for one.
Possible fix: keep one container for agent_runs rows opened by heal.ts. For example, execute could count the running hosted feature runs and hold a feature run back while it would take the last container. Or raise maxInstances by one and cap feature runs at the old number.
Metadata
Metadata
Assignees
Labels
No labels
Activity
QuinnsCode commentedon Oct 13, 2026
Parked for later (owner's call, 2026-10-13). #1252 did this by raising AgentSandbox
max_instancesfrom 2 to 3, which moves the 24/7 worst case from ~$41 to ~$62 a month. Today the fix and rebuild work runs on the owner's Mac (desktop codeSplitter and Agent Office workers), so a hosted fix lane only matters when codeSplitter fixes things while the Mac is off.Options when this comes back:
office/widget-e595is kept.codeSplitter: hold one hosted container for fix tasks; feature runs s…