You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 798591c
Browse filesBrowse the repository at this point in the historyBrowse files
feat(dotnet): add subagent events and attribution to AGUI.Abstractions
Part of PNI-196.
Brings the .NET wire format to parity with TypeScript: SubagentStartedEvent,
SubagentFinishedEvent and SubagentErrorEvent, plus `subagentId` on every
event type the design makes attributable, wired through the polymorphic
event converter, the source-generated serializer context, and the message
model.
Deliberate choices:
- The three required lifecycle properties are declared nullable and
rejected only when MISSING, not when empty. An earlier pass rejected
empty strings too, which made .NET stricter than the zod schemas and
would have had the two SDKs disagree on whether a stream is valid.
Where the schemas differ, the schemas win.
- `subagentId` travels on ChatMessage.AdditionalProperties under
"agui.subagentId" so attribution survives a round trip through
Microsoft.Extensions.AI types, which have no field for it.
- PublicAPI.Unshipped.txt is appended to in place rather than re-sorted.
The analyzer accepts either, but re-sorting turned 83 new entries into
a 611-line diff that buried them.
Known limitation, with a test pinning it: parallel tool calls owned by
DIFFERENT subagents keep only the first owner, because AsChatMessages
merges a parallel-call run into one ChatMessage and one message holds one
attribution. Splitting the run by owner was implemented and reverted — it
split without reordering the results and so produced the exact shape
OpenAI rejects (microsoft/agent-framework#2699). The provider constraint
is adjacency, not grouping, so interleaving would satisfy it AND keep both
owners; that changes reconstruction for non-subagent parallel calls too,
so it is a deliberate decision rather than a drive-by fix here. PNI-293
carries the analysis; the pinning test should be replaced if it lands.
0 commit comments