fix: clean citation leaks and duplicate agent output in thinking process

[Ayaz-Microsoft]

Purpose

Fix citation leaks and duplicated agent output in the AI Thinking Process

Summary

Fixes four related issues in the multi-agent streaming/orchestration layer: citation markers leaking into responses, agent names/responses duplicating in the "AI Thinking Process", the same agent repeating across consecutive rounds, and a lack of diagnostics for the Magentic manager's round loop.

Background

The Magentic orchestrator re-invokes agents across rounds by design, and streaming output arrives in partial deltas. The previous handling assumed append-only rendering and per-chunk cleaning, which broke down for multi-round runs and split markers.

Changes

Backend —  src/backend/callbacks/response_handlers.py 

• Cross-chunk citation cleaning: buffer a trailing, not-yet-closed  【…  marker until its closing  】  arrives in a later delta, then strip it. Prevents markers like  【5:0†source】  from leaking when split across streaming chunks.
• Per-agent grouped "thinking" snapshot: accumulate output into an ordered  {agent → latest text}  map and emit the full grouped snapshot each delta, so each agent appears once with its latest response.
• Turn-boundary reset:  mark_streaming_turn_complete()  resets the current section at each agent turn boundary, so consecutive same-agent rounds don't concatenate.

Backend —  src/backend/orchestration/orchestration_manager.py 

• Apply  clean_citations()  to the Group Chat Manager's final answer, which previously bypassed citation cleaning entirely.
• Call  mark_streaming_turn_complete()  at each  executor_completed  turn boundary.
• Add a per-round diagnostic log:  [ROUND N] next_speaker=… satisfied=… in_loop=… progress=…  from the manager's progress ledger.

Frontend —  src/App/src/hooks/usePlanWebSocket.tsx 

• Replace the streaming buffer with the latest backend snapshot instead of appending, matching the new backend grouping.

Issues fixed:

• Citation markers ( 【5:0†source】 ) leaking in agent responses — fixed by cross-chunk citation buffering.
• Same markers leaking in the Group Chat Manager final answer — fixed by applying  clean_citations()  to  final_text .
• Agent names/responses duplicated in the AI Thinking Process — fixed by the backend grouped snapshot + frontend replace.
• Same agent repeated ×N in consecutive rounds (e.g. Triage ×7) — fixed by the turn-boundary section reset.
• No visibility into the manager round loop — fixed by the per-round  [ROUND N]  log.

Does this introduce a breaking change?

• Yes
• No

How to Test

• Get the code

git clone [repo-address]
cd [repo-name]
git checkout [branch-name]
npm install

• Test the code

•  py_compile  clean on both backend files;  tsc --noEmit  clean on the frontend.
• Verified  clean_citations()  strips the exact marker  【5:0†source】 .
• Deployed to  rg-macae-az0  and confirmed on live runs: clean 6-round convergence with real output + image, no duplicated/repeated agents in the thinking process, and per-round logs present. Group Chat Manager final answer no longer shows the citation marker.

What to Check

Verify that the following are valid

• ...

Other Information

[Copilot]

Pull request overview

Fixes streaming/orchestration “AI Thinking Process” rendering issues by cleaning citation markers across chunk boundaries, emitting a grouped per-agent snapshot (instead of append-only deltas), and adding per-round diagnostics in the Magentic round loop.

Changes:

• Backend: add cross-chunk citation stripping + grouped per-agent thinking snapshot state (with turn-boundary resets).
• Backend: clean citations from the Group Chat Manager’s final answer and add per-round [ROUND N] ... diagnostic logging.
• Frontend: treat streamed messages as full snapshots and replace (not append) the streaming buffer.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
src/backend/orchestration/orchestration_manager.py	Resets thinking state per run, cleans citations on final manager output, adds per-round logs, and marks turn boundaries for streaming snapshot behavior.
src/backend/callbacks/response_handlers.py	Implements citation buffering across streamed chunks and builds/emits grouped “thinking” snapshots per user.
src/App/src/hooks/usePlanWebSocket.tsx	Updates client buffering to replace with the latest backend snapshot instead of concatenating.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.