[benbrandt]

Summary

The subagents RFD reports a message between sessions as two independent views: the sender's outgoing session_message and the recipient's incoming one. Message IDs are local to each transcript, and Clients must not link views by equal IDs, so a Client has no way to go from one view to the other. This revision lets each view name the other, adds the fields to the unstable schema, and keeps a message's identity off its chunks.

RFD (docs/rfds/subagents.mdx)

• senderMessageId on the incoming view names the messageId of the outgoing view in the sender's transcript.
• recipientMessageId on the outgoing view names the messageId of the incoming view in the recipient's transcript.
• Both are optional and nullable and follow the participant-ID rules: omitted or null means not supplied and keeps a known value, and a known value doesn't change. Either can arrive in a later chunk or update, so an Agent can start streaming a view before it knows the other one's ID.
• When both views carry a reference they must name each other; a Client that sees conflicting references treats the views as unlinked. A reference doesn't mean the recipient finished processing the message, and it doesn't establish ownership.
• Message IDs stay transcript-local, and the rule against linking by equal IDs stays.
• Chunks carry only content. The participant IDs move off session_message_chunk, and the references are only on session_message, as message-level fields are for agent_message, whose chunks also carry only content. An Agent that learns a message's identity while streaming sends a metadata-only upsert. Before this, the participant IDs had two sources that could disagree, and a chunk patched message state even though its _meta doesn't. I found no recorded reason for putting them on chunks (feat(unstable): Intial RFD and schema for subagents #1992 and its review), and no implementation sends them there: claude-agent-acp only sends upserts. A new FAQ entry, "Why do chunks carry only content?", records this.
• Both fields exist so that either side can find its counterpart directly, including when the other transcript isn't loaded or rendered.
• The FAQ records the rejected alternatives: identical IDs (the ID is fixed at the first update, so an Agent would have to hold a view back until it knows the link), and references to the delegating tool call (session messages exist so delegations don't have to be modelled as tool calls).
• Adds a validation scenario and a revision-history entry.

Schema

• sender_message_id and recipient_message_id on SessionMessage, in v1 and v2, behind unstable_subagents, with builders.
• SessionMessageChunk loses sender_session_id and recipient_session_id (and their builders) in v1 and v2. This is a breaking change to unstable types: identity fields sent on a chunk are now ignored. A malformed value is treated as not supplied (DefaultOnError), as for the participant IDs.
• Regenerated schema.unstable.json for v1 and v2 and the draft schema docs. The draft prompt-turn (v1) and prompt-lifecycle (v2) pages mention the references, and their example now names the participants with an upsert before streaming chunks.

Users

• claude-agent-acp sends both references when it reports a delegation as session messages: feat: report RFD delegations as linked session messages claude-agent-acp#1405.
• Zed will use them to link message rows and to build one subagent card per delegation (acp_thread: Show ACP session messages in the transcript zed-industries/zed#66028 is the first step; it works without the fields).

Testing

• New round-trip test in v1 and v2: both references serialize and decode, and null or malformed references mean not supplied.
• The session-message tests now check that identity fields on a chunk aren't part of it, and supply late participants and references with upserts.
• npm run generate and npm run check (clippy, fmt, prettier, spellcheck, cargo test --all-features) pass.