agentHost: Don't split the final response at the reconnect boundary

[roblourens]

Reconnecting to an agent host session while a turn is still streaming could split the final assistant response in two — in the observed case mid-word, with the prefix folded into the collapsed tool/thinking section and the remainder rendered as a separate markdown block outside it, making it look like a tool call had interrupted the answer.

Root cause

Reconnect runs two independent converters over the same response parts:

	what it is	when it runs
activeTurnToProgress	one-shot imperative snapshot	reconnect only
_observeTurn	always-on reactive observable graph	live, reconnect, server-initiated

They de-duplicate through adoptInvocations, which is keyed on live ChatToolInvocation instances. A tool call that had already settled before reconnect is rendered by the snapshot as a toolInvocationSerialized part, which cannot be adopted — so the live graph emitted a second, live invocation for it.

That duplicate lands between the restored markdown prefix and the markdown still streaming into the same response part. ChatResponseModel only merges a markdown update into an immediately preceding markdown part, so the two fragments stayed separate:

markdown prefix
tool invocation   <-- spurious
markdown continuation

On completion, disclosure grouped the prefix into the preceding activity section and left the continuation outside it, exposing the reconnect boundary as if it were a semantic boundary in the model's answer.

The underlying data was never wrong: the model produced one continuous response, the agent host persisted one continuous markdown part, and AHP delivered the snapshot and continuation correctly. The split was introduced entirely in the workbench.

Fix

Record what the snapshot emitted per tool call rather than only the adoptable subset, so per-tool setup can skip a settled tool call that the snapshot already rendered in full. Subagent tools are excluded from the skip because their setup is what streams the child session's inner tool calls into the parent response.

This is deliberately the minimal, low-risk fix. The deeper problem is that two converters exist at all and must be kept in sync by hand; a follow-up PR removes the snapshot path entirely so reconnect and live turns share one code path. That one is riskier and wants more bake time in Insiders.

Testing

• New regression test in the reconnection to active turn suite, verified failing before the change:

  + expected - actual
    [
      "toolInvocationSerialized"
      "markdown:that fallback fails to"
  -   "toolInvocation"
      "markdown:ward destroying history."
    ]
  

• Full AgentHost suite: 2206 passing.
• npm run typecheck-client clean.

(Written by Copilot)

[Copilot]

Pull request overview

Prevents settled tool calls from being re-emitted during agent-host reconnection.

Changes:

• Tracks snapshot tool calls, including serialized invocations.
• Skips duplicate settled tools and adds regression coverage.

Show a summary per file

File	Description
agentHostSessionHandler.ts	Adds snapshot-aware tool-call handling.
agentHostChatContribution.test.ts	Tests settled-tool reconnection behavior.

Review details

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

Suppressed comments (1)

src/vs/workbench/contrib/chat/browser/agentSessions/agentHost/agentHostSessionHandler.ts:3507

• This exception leaves the reconnect split in place for completed subagent calls. The snapshot has already emitted the serialized parent invocation, but this path continues into _setupServerToolCall, which creates and sinks a second live parent invocation before observing the child session; that second part can still land between the restored markdown and its continuation. Please separate child-session observation from emitting the parent invocation, and cover the completed-subagent reconnect case.

			&& !shouldObserveSubagentChat(initial)) {

• Files reviewed: 2/2 changed files
• Comments generated: 3
• Review effort level: Balanced

[roblourens]

Thanks — the suppressed comment about the subagent exception was a real bug, and a good catch. Fixed in 274bd59.

I confirmed it before fixing. With a completed subagent tool in the active turn, reconnect produced exactly the split this PR is meant to prevent:

toolInvocationSerialized
markdown:"that fallback fails to"
toolInvocation              <-- spurious, splits the answer
markdown:"ward destroying history."

The guard declined to skip subagent tools so their child session would still be observed, but that fell through to _setupServerToolCall, which sinks a second live parent invocation. Observation and emission were coupled.

Fix is exactly the suggested separation: still build the invocation so subagent observation has something to drive, but don't emit it when the snapshot already rendered the tool as a serialized part. Added replays a settled subagent tool call once, keeping the streamed markdown in one part to cover the case; verified failing before the change.

The three comment-verbosity notes are addressed in the same commit.

Full AgentHost suite: 2207 passing. typecheck-client clean.

One related gap I am not fixing here, to keep this PR small and safe for today: a completed subagent's inner tool calls also arrive from observation and are appended after the markdown prefix, so they can split the response the same way. That is pre-existing rather than introduced here, and it goes away entirely in the follow-up (#331047), which removes the snapshot so everything replays in response-part order through one producer.

(Written by Copilot)