A durable operation record is shared by concurrent retries. The design challenge is preventing a later timeout observation from downgrading an already committed success, while still allowing a timeout to be upgraded if the operation eventually completes. The key unknown is the exact storage and transaction mechanism, so the guidance should remain implementation-neutral and emphasize atomic monotonic settlement.

An agent plugin caches task context before injecting it into a system-message transform. Multiple root sessions and child sessions share one process. The design problem is how to prevent cross-session context leakage and bound memory without inspecting local configuration.

A durable operation record is shared by retries, but completion signals can arrive in either order: one retry may learn and persist success while another later reports a timeout or unknown outcome. The central question is how to define settlement so uncertainty cannot overwrite a committed result, regardless of completion order. Known evidence is the shared record and the conflicting success and timeout observations; unknowns include the exact storage guarantees, state vocabulary, and whether outcomes carry versions or fencing tokens. The reusable reasoning should distinguish authoritative execution evidence from weaker observations and explain the required transition ordering.

An agent plugin caches task context before injecting it into a system-message transform. Multiple root sessions and child sessions share one process. Question: how to key and isolate that cache so context never leaks across sessions, and how to bound memory under concurrent sessions. Known: process-global caches are the failure mode; transforms run per message turn. Unknown: whether child sessions should inherit a parent snapshot or only an explicit subset, and what eviction policy fits long-lived hosts. Seeking durable design guidance, not repo-specific steps.