This design question concerns an agent plugin that caches task context before a system-message transform while multiple root and child sessions share one process. The key concerns are preventing cross-session context exposure, defining inheritance and cleanup semantics, rejecting stale asynchronous writes, and imposing entry, count, byte, and lifetime limits. The implementation is unspecified, so the discussion should focus on reusable design principles and validation scenarios without sharing private code or environment details.

Assessment: use request-local context where possible; otherwise use an immutable cache keyed by plugin instance, authorization lineage, session identity and epoch, content revision, and explicit child binding. Resolve live authorization generation at each read, write, and injection rather than from stale async closure state. Root and child entries must be distinct; inheritance is an explicit immutable snapshot. Tombstone prior epochs, reject closed or generation-mismatched entries, and prevent late callbacks from recreating them. Bound memory with per-entry and aggregate byte limits, entry-count limits, TTL or idle expiry, and LRU eviction; reject oversized values and cache only normalized transform input. Validate ownership on every lookup and expose value-free metrics. No repository, configuration, or machine inspection was performed and no tests were executed. This is a completed reusable design assessment; no new skill proposal is needed because the available session-bound cache guidance already covers the durable pattern.

Updated design: authorization generation is part of the cache binding, not merely metadata. Use a composite key containing session identity, session epoch, live authorization generation, content revision, and explicit child binding. On revocation or replacement, atomically advance the generation, mark the prior generation closed, and tombstone or remove every entry from it, including inherited child snapshots. Every cache read, transform injection, and network operation must obtain the current authorization handle and generation independently; cached context may describe intent but must never grant network authority. Require a fresh authorization check immediately before network activity, bind the request to that generation, and abort or deny if the generation changes before dispatch or during refresh. Async callbacks must compare their captured generation with the live generation before writing, injecting, or authorizing. Clear sensitive token material where applicable, and prevent old retries or queued work from using it. Keep previous-generation entries within byte, count, TTL, and idle limits while draining; do not retain them for same-session continuity. Reasoning only: no repository, configuration, or machine inspection was performed and no tests were executed. Suggested validation includes replacement during cache population, transform execution, queued network dispatch, retry, child inheritance, and concurrent replacement races.