feat: report RFD delegations as linked session messages (#1405)

## Summary

For a client on the subagents RFD (every ACP v2 client, and v1 clients
that declare `subagents`), session messages now replace the Agent/Task
call that launches a child. The parent no longer sees that tool call.
AIR's form is unchanged.

- **Prompt.** The launch's prompt is a `session_message` in both
transcripts: the parent's outgoing view and the child's incoming one.
Both use the launch's tool use id, and each names the other with
`recipientMessageId` / `senderMessageId` from
agentclientprotocol/agent-client-protocol#2465.
- **Answer.** A foreground child's answer, from the launch's result, is
the child's `session_message` to the parent (`<launch id>:reply`). It
uses the result as `AgentReporter` renders it, so the model-facing
`agentId: …` line and `<usage>` block are gone. A failure or
cancellation comes from the child's `idle` stop reason, which it already
reported. A background launch's result gets no reply.
- **Hand-back mode.** Claude Code (seen with 2.1.303) can have the child
hand its report back with a `SubagentHandback` call; the Agent result
then only says that the report "was delivered to you as a message". That
call is reported as the child's message to the parent, in both
transcripts and linked the same way, and the launch's reply is skipped
for that run. A failed hand-back delivered nothing, so it stays a tool
call of the child.
- **Unexposed work.** A launch whose child is never exposed is shown as
a tool call after all, with its final status
(`unexposedControlFallback`, the former `failedControlFallback`).
- **Hidden launches by id.** The adapter's own report that closes an
unfinished call (for example at a forced cancel) carries no `toolName`,
so the runtime tracks hidden launches by id to keep that report from
reaching the client for a call it never saw.
- **Replay** sends the same messages from the child's transcript: the
launch's prompt, the reply, and hand-backs. A launch without Claude
Code's metadata file keeps its tool call and the missing-history notice.
Replay can't tell a child's runs apart, so any hand-back in its
transcript suppresses its launch's reply.
- **SendMessage** calls stay tool calls of their sender, and the child
gets the incoming view without references, as before.
- README's subagent section describes the new behavior.

Do not merge before agentclientprotocol/agent-client-protocol#2465
settles the field names: a merge to `main` publishes a preview right
away. Clients on the published schema ignore the two fields (no
`deny_unknown_fields` in the Rust schema; the scenario validation with
the SDK's zod schema passes).

## Testing

- `npm run build`, `npm run check` and `npm run test:run` pass (57
files, 4 skipped as before).
- The six v2 `subagent-*.jsonl` scenario snapshots changed: the launch's
tool-call updates are gone, the parent gets the outgoing view, the
child's view carries `senderMessageId`, and foreground results became
`:reply` messages. AIR snapshots are unchanged.
- New unit tests cover hidden launches, nested launches, replies, failed
and background launches, unexposed launches with both final statuses,
the close report without a tool name, and hand-back success and failure,
live and on replay.
- A real run: the built adapter driven over v2 stdio with Claude Code,
one foreground subagent asked to answer PONG, then `session/resume` with
replay in a fresh process. Live and replay both showed the linked prompt
(parent → child), the linked hand-back "PONG" (child → parent), and no
Agent tool call.
This commit is contained in:
Ben Brandt
2026-10-23 13:12:33 +02:00
committed by GitHub
parent 50cd76e78d
commit 59e292e29f
13 changed files with 767 additions and 126 deletions

View File

@@ -47,9 +47,18 @@ Task subagent as its own child session, following the [subagents RFD](https://gi
`end_turn` or `cancelled` when the SDK gives the reason. A failure is `idle` with the `error`
stop reason and the SDK's error text as its `error`. The child keeps its session for every
delegation to it.
- The prompt of the Agent or Task call that launches the child, and the message of each
SendMessage call that resumes it, is a `session_message` in the child's transcript. The tool
call stays a tool call of the session that made it.
- Session messages replace the Agent or Task call that launches the child, which its parent is
not shown:
- The launch's prompt is a `session_message` in both transcripts: the parent's outgoing view and
the child's incoming one. Both use the launch's tool use id, and each names the other with the
draft RFD's `recipientMessageId` and `senderMessageId`.
- A foreground child's answer, from the launch's result, is the child's `session_message` to
the parent. In Claude Code's hand-back mode, the child's `SubagentHandback` call is its
message to the parent instead, in both transcripts and linked the same way.
- A launch whose child is never exposed is shown as a tool call after all, with its final
status.
- The message of a SendMessage call that resumes the child is a `session_message` in the child's
transcript. The SendMessage call stays a tool call of the session that made it.
- `session/cancel` on the parent interrupts Claude Code, which stops the foreground children that
the turn waits on. Background children keep running, as the RFD prefers; a client stops one with
its own `cancel`. Each stopped child reports `idle` with `cancelled` from Claude Code's own
@@ -65,8 +74,8 @@ Task subagent as its own child session, following the [subagents RFD](https://gi
exits is `idle` with no stop reason.
- A child whose parent is not yet known when it ends is not exposed, since the RFD forbids
guessing its parent.
- A replay restores each child under its live session id, from its own transcript: its prompt and
each SendMessage message are `session_message`s, and its end is an `idle` before the response.
- A replay restores each child under its live session id, from its own transcript, with the same
session messages as live, and its end is an `idle` before the response.
A launch without Claude Code's metadata file for the child stays an ordinary tool call of the
parent, and a notice, for a client that takes notices, says that subagent history is missing.
- After a load or resume, with or without replay, a SendMessage that resumes an agent from before