211 — Chinese group-chat failure reports: shared context corroborated in source Recorded September 6, 2026 UTC. Read-only public-source review. RESULT A Chinese-language first-person report describes a two-agent review failing because unrelated earlier errors remained in their shared room history. Pinned Hermes Studio source corroborates the shared-history mechanism. No raw execution transcript, external publishing receipt, or Chinese-lab attribution was recovered. This is a useful operational lead, not confirmation of an escaped swarm. STRONGEST REPORT https://github.com/heypandax/cc-pocket/issues/232 Created August 10, 2026 at 10:51:13 UTC; closed at capture. The author requests independent agent sessions and compares existing products. They report testing two Hermes Studio members: A produced an answer, but B, despite receiving that answer in its input, interpreted two older unrelated failure messages as evidence that A had not answered. B then produced its own answer instead of reviewing A. This is a reporter's account, with no raw input/output transcript attached in the inspected body. Model names used as examples do not verify actual model identities. RELATED SOURCE-LEVEL PROPOSAL https://github.com/EKKOLearnAI/hermes-studio/issues/2456 Created August 10 at 02:52:33 UTC; open with zero comments at capture. Proposes isolating each agent's context and using explicit handoff summaries. References master@b8c03b0. We resolved that reference to b8c03b013a0e830c4263ec63f93d3290bcf613a4, fetched its non-truncated Git tree and verified two source files against their Git blob hashes. Verified source behavior: - group-chat/index.ts getMessagesForContext(roomId, cutoff) reads recent room messages with a message-window/cutoff, excluding workspace_diff. There is no recipient-agent argument at this method boundary. This is bounded recent context, not necessarily every historical message. - context-projection.ts maps the agent's own messages to assistant role and other speakers to user role, retaining textual speaker attribution. Tool results become user-role text with tool labels. The projection filters workspace_diff tool messages. - A supplied summary is inserted with a canned assistant acknowledgment. Such an acknowledgment in a serialized history can be application-generated, not a model's actual response. - Ordinary projected content has @mentions stripped by a regular expression. These mechanisms make the reported cross-agent history confusion plausible; they do not reproduce or prove the reported model behavior. No investigated code was executed. Pinned source: https://github.com/EKKOLearnAI/hermes-studio/blob/b8c03b013a0e830c4263ec63f93d3290bcf613a4/packages/server/src/services/hermes/group-chat/context-projection.ts https://github.com/EKKOLearnAI/hermes-studio/blob/b8c03b013a0e830c4263ec63f93d3290bcf613a4/packages/server/src/services/hermes/group-chat/index.ts OLDER MUTUAL-MENTION REPORT https://github.com/EKKOLearnAI/hermes-studio/issues/1385 Created June 7 at 06:16:45 UTC; closed with six comments at capture. A user supplies a manual patch tutorial for bot-to-bot mentions and following up with the previously addressed agent. Its mentionDepth > 10 guard shows concern about recursive mentions. The A/B examples are illustrative, not timestamped logs proving an actual runaway exchange. June 14 and June 16 comments redirect remaining requirements to issue 1547; closure does not itself establish a shipped fix. That follow-up issue was not inspected in this pass. SINGLE-AGENT CONTROL CASE https://github.com/agentscope-ai/QwenPaw/issues/6241 Created July 18 at 07:21:47 UTC; closed at capture. Reports repeated responses and repeated memory_search calls, claiming QwenPaw 1.1.12.post2, Windows/Feishu and mimo-v2.5. The body uses schematic turn labels rather than a raw timestamped trace. This is a single-agent repetition report, not multiple agents coordinating. Proposed compression/guard explanations remain reporter hypotheses. NAGA FOLLOW-UP: BOUNDED NEGATIVE Three GitHub issue searches in RTGS2017/NagaAgent, using 探索, 论坛 and travel separately, each returned total_count 0. This does not exclude differently worded, unindexed, deleted or externally hosted reports. Combined with report 210's exact-marker web search, it supplies no public exploration run/post join. Keep the distinctive Naga markers for later corpus searches, but reduce priority until a public receipt or transcript appears. NEXT SEARCHES Follow explicitly linked public reproduction artifacts and issue 1547, particularly raw shared-room transcripts with stable message IDs and model requests. Search Chinese descriptions of 串台, 上下文污染, 互相@, 自言自语, 接力 and 复核, as well as English phrases; translation of agent alone is not a useful detector. Distinguish designed mention routing and host-generated dialogue from independently observed autonomous behavior. A secondary search surfaced a Hermes issue number 32791 through an automated Chinese ecosystem digest, claiming Discord bots looping. Follow-up report 212 verifies the primary Hermes issue and corrects the initial OpenClaw repository attribution. INFRASTRUCTURE Current GitHub access is working. An owned mainland broadband machine with SSH, a dedicated browser, 2 CPU cores, 4 GB RAM and 40–70 GB disk is a reasonable starting configuration for comparing challenged public Chinese pages. No GPU needed. A Hong Kong machine is a comparison location, not guaranteed mainland access. Authorized exports of public posts and voluntarily shared redacted agent run logs would be especially valuable. Missing hosts, private records and API authentication requirements are not solved by another IP. No new infrastructure was purchased. CAPTURES investigation/china/211-private/: three Naga issue-search responses; four issue bodies; six issue-1385 comments; pinned Studio tree and two Git-verified source files. SHA256SUMS records capture integrity. These API snapshots establish what was returned during this review; they are not independent historical archives.