Three issues in CodexSessionRuntime found reviewing #136, all upstream's code
rather than Pylon adaptation, and none verifiable without a real Codex 0.150
app-server. Separate from #146, which is about decode tolerance.
A transient timeout permanently loses a child's model. lookupStarted is
latched before the request and never reset on failure or timeout, so one 5s
timeout leaves that sub-agent without model or effort for the rest of the
session. The closed + lookupStarted pair likewise blocks re-lookup for a
re-registered child.
Cleared effort keeps a stale value. updateCollabChildMetadata only assigns
truthy values, so an effort cleared by thread/settings/updated retains the old
one even with overwriteKnown: true.
collabChildMetadataRef grows without bound. Entries are only flagged
closed, never deleted, and the intercept inserts one for any non-root thread
id — unlike the sibling maps, which clean up on thread/closed.
Related: #136, and #146 for the decode half.
Three issues in
CodexSessionRuntimefound reviewing #136, all upstream's coderather than Pylon adaptation, and none verifiable without a real Codex 0.150
app-server. Separate from #146, which is about decode tolerance.
A transient timeout permanently loses a child's model.
lookupStartedislatched before the request and never reset on failure or timeout, so one 5s
timeout leaves that sub-agent without model or effort for the rest of the
session. The
closed+lookupStartedpair likewise blocks re-lookup for are-registered child.
Cleared effort keeps a stale value.
updateCollabChildMetadataonly assignstruthy values, so an effort cleared by
thread/settings/updatedretains the oldone even with
overwriteKnown: true.collabChildMetadataRefgrows without bound. Entries are only flaggedclosed, never deleted, and the intercept inserts one for any non-root threadid — unlike the sibling maps, which clean up on
thread/closed.Related: #136, and #146 for the decode half.