A team keeps skills in a store and edits a SKILL.md to fix a wrong step. Running locally, the fix appears on the very next run. On their LangSmith Deployment, new threads get the fixed instructions, but users continuing existing threads keep receiving the old skill list and wording. What explains the difference, and what is the documented fix?
- A.
LangSmith Deployment disables the discovery step of
SkillsMiddleware, so the skill list is frozen to whatever was bundled into the image - B.
Skills load once per thread and stay in state; the deployment's persistent checkpointer keeps that list for the thread, whereas the local runs had none. Reset
skills_metadatatoNoneto reload - C.
The deployment snapshots the store when a revision is built, so a new revision must be deployed before any thread can see the edited skill
- D.
Provider-side prompt caching is replaying the previously cached system prompt, so removing the prompt caching middleware from the agent's stack would let the existing threads see the edited skill at once
Show answer
Answer: B
Skill metadata is cached in thread state; with a persistent checkpointer it survives for the thread's life until skills_metadata is reset to None.
- A. Discovery clearly runs on the deployment, because new threads pick up the fixed skill; it is not disabled there.
- B. Correct. Thread state holds the skill list; the deployment's checkpointer preserves it, so a reset to `None` is needed to reload.
- C. New threads already see the edit, which rules out a store snapshot taken at build time; no redeploy is needed for a store change.
- D. Prompt caching only reuses a prefix that is actually sent; the stale wording comes from skill metadata held in the thread's state.