AgentScope: public four-agent memory reproduction and merged broadcast fix
Reviewed September 6, 2026 UTC. Public read-only evidence.

FINDING
A Chinese user's AgentScope bug report provides reproduction code and printed cross-agent memory, followed by maintainer acknowledgment and a merged source fix. This is stronger evidence of a bounded multi-agent interaction than an architecture diagram. It remains a tutorial-derived local reproduction, not an unattended or escaped swarm.

PRIMARY REPORT
https://github.com/agentscope-ai/agentscope/issues/1327
Opened March14,2026 by toohandsome; environment lists AgentScope1.0.17, Python3.11, Windows11. Four ReActAgent instances, kimi2/Alice/Bob/Charlie, enter MsgHub and are called sequentially to introduce themselves. The configuration uses Chinese model labels through iflow/scnet endpoints; these are requested labels, not authenticated model weights or lab provenance. Credentials in the displayed reproduction are placeholders; none used.
The author's March16 follow-up includes Alice's printed memory: four text blocks attributed to the four agents and one Bob-labelled thinking block. A maintainer initially said the DeepSeek thinking block should remain private, then acknowledged the follow-up at06:45:58Z. These are user-published console records, not independently reproduced calls. They support shared message delivery and an unintended extra content block; no task completion or external posting is demonstrated.

SOURCE FIX
https://github.com/agentscope-ai/agentscope/pull/1332
Title: feat(agent): filter thinking content before broadcast.
GitHub metadata reports merged March26,2026 at09:01:49Z, merge revision346dce99a514278a212024d486845fe5c27af89f.
https://github.com/agentscope-ai/agentscope/blob/346dce99a514278a212024d486845fe5c27af89f/src/agentscope/agent/_agent_base.py
The retrieved PR diff modifies _broadcast_to_subscribers: it strips content blocks whose type equals thinking before passing messages to subscriber.observe. List input is handled per message; changed messages retain identifying metadata and the original ID. This confirms a concrete broadcast path matching the reported failure class. The patch does not, by itself, prove removal of inline think tags inside text blocks; typed thinking blocks and tagged text are different representations. No target code or tests were executed.

HICLAW NAME AND DEPLOYMENT FOLLOW-UP
https://api.github.com/repos/higress-group/hiclaw
HTTP200 after redirect to GitHub repository ID1163358617; returned full_name agentscope-ai/AgentTeams. This establishes the current GitHub repository identity behind the old address, not the exact historical transfer date.
https://github.com/agentscope-ai/AgentTeams/issues/422
Direct issue API read confirms a March24 Chinese remote-access report, including a400 error at /v1/chat/completions and the author's routing/TLS workaround narrative. It supplies no completed multi-agent conversation export. Its example addresses were not visited and its configuration advice was not executed or recommended. This is deployment troubleshooting evidence, separate from AgentScope issue1327.

ASSESSMENT
AgentScope1327 adds a user-authored, Chinese-context multi-agent reproduction with a corroborating code change. Shared internal memory is not public scratch-space escape. The actors were deliberately created and invoked in a fixed sequence. Model labels, forum language and API hostnames do not establish a Chinese lab as operator. No connection to XZ or the HiClaw stock-team user is established.

NEXT
Follow comparable public MsgHub/AgentTeams reports that include task handoffs or result artifacts beyond self-introductions. Prefer actual console records and matching code history. Avoid counting source safeguards against mention loops as proof that a particular loop occurred.

PRESERVATION
317-private: repo.json, issue422.json, memory-issue.json, memory-comments.json, selected memory-output-excerpt.txt, fix.json, fix-files.json, SHA256SUMS and publication-check.json. Raw reports remain private. No credentials used, external accounts contacted, model endpoints called or target code executed.