chat: true mid-turn steering for cloud sessions

[osortega]

Problem

When steering a Copilot cloud chat session mid-turn, the message was stranded: it showed a "STEERING" pending bubble that only sent after the turn finished (queue behavior), rather than injecting into the running turn the way github.com/copilot/agents does.

Scope

This PR now delivers true mid-turn steering for cloud sessions, in three layers:

1. Queue-flush fix — a message queued during a streamed (activeResponseCallback) turn is flushed when the turn completes (was previously stranded forever).
2. Immediate mid-turn dispatch (core) — a Steering message is dispatched to the participant immediately instead of waiting for completion, so the provider can POST the steer server-side while the turn runs.
3. Provider injection + continuous rendering — the cloud provider POSTs /tasks/{id}/steer and the running stream renders the injected result, matching github.com/copilot/agents.

Core (chatServiceImpl)

For non-server-managed streamed sessions:

• Dispatch a queued Steering message immediately, but only when the in-flight pending request is the synthetic streamed-turn tracker (identified via an explicit WeakSet, since real requests are inserted into _pendingRequests before their requestId is assigned) — never a real in-flight request.
• Restore in-progress tracking deterministically when the dispatched request settles (via responseCompletePromise), so requestInProgress stays true for the still-active stream.
• Non-steering Queued messages are unaffected and still flush on completion.

Extension (copilotCloudSessionsProvider)

• Track live TaskTurnStreamer task ids; when a stream is already active, handleTaskFollowUp only POSTs the steer (no second streamer).
• The follow-up stream is yield-aware (mirrors the Copilot CLI provider): it polls context.yieldRequested and cancels only its local streamer on yield, so the chat service flushes the next steer immediately.
• Render mode is chosen from the pre-steer state: a steer into a genuinely active turn (both task and latest turn active) renders mode:'current' (the injection produces no new task.sessions[] row, so mode:'next' would time out); a steer onto a settled task renders mode:'next'.

Testing

Unit tests in chatService.test.ts:

• Immediate dispatch — a steering message on a streamed session is dispatched exactly once, mid-turn (stream not complete), preserves in-progress tracking (a plain send is rejected), and is not re-dispatched on completion.
• Completion flush — a non-steering Queued message waits during the turn and is flushed when it completes.

Verified end-to-end against a live cloud task: the steer POSTs within ~100ms and the injected message + continued output render inline. typecheck-client and the copilot extension typecheck are clean.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[Copilot]

Pull request overview

Fixes queued steering messages becoming stranded after streamed contributed-session turns complete.

Changes:

• Flushes pending requests after streamed-response completion.
• Adds regression coverage for queued steering messages.

Show a summary per file

File	Description
chatServiceImpl.ts	Processes queued requests after stream completion.
chatService.test.ts	Tests steering-message delivery after completion.

Review details

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

[osortega]

Added a second commit that builds on the flush fix to implement true mid-turn steering for cloud sessions (previously the steer only sent after the turn settled — queue behavior — rather than injecting into the running turn like github.com/copilot/agents does).

Core (chatServiceImpl) — for non-server-managed streamed (activeResponseCallback) sessions, dispatch a queued Steering message immediately, but only when the in-flight pending request is the synthetic streamed-turn tracker (requestId === undefined), never over a real in-flight request (which it would dispose). Tracking is re-established on the next progress tick so the turn stays in-progress.

Extension (copilotCloudSessionsProvider):

• Track live TaskTurnStreamer task ids; when a stream is already active, handleTaskFollowUp only POSTs /steer and lets the running stream render the injection (no second streamer).
• Make the follow-up stream yield-aware, mirroring the Copilot CLI provider: poll context.yieldRequested and return on yield so the chat service flushes the next steer immediately (near-instant chaining instead of waiting ~60s for the turn to settle).
• Choose the render mode from the pre-steer task state: a steer into an active turn renders mode:'current' (the injection produces no new task.sessions[] row, so mode:'next' would time out); a steer onto a settled task renders mode:'next'.

Verified end-to-end against a live cloud task: the steer POSTs within ~100ms of submitting, and the injected message + continued output render inline. typecheck-client and the copilot extension typecheck are clean.

[Copilot]

Review details

• Files reviewed: 3/3 changed files
• Comments generated: 12
• Review effort level: Medium