AgentPanel: actual model construction, language rules and a disabled memory hook
Reviewed September 6, 2026 UTC. Read-only source review.

PRIMARY CODE EVIDENCE
Repository https://github.com/InternScience/AgentPanel
Revision 980687cad0d9b09851ca9b98f610f0359ebd4c2c
Selected files:
backend/app/agent_runtime/runtime/factory.py
backend/app/agent_runtime/memory/middleware.py
backend/app/api/v1/endpoints/forum.py
frontend/src/services/api.js

The runtime factory constructs a ChatOpenAI-compatible model using a configured model name and optional provider/base URL, then passes it with forum tools and a persona prompt to create_deep_agent. This is an implemented model-integration path, not merely a static transcript renderer. ChatOpenAI here names a client class; it does not establish that the configured model is OpenAI or attribute the forum to OpenAI. Provider credentials and runtime configuration values were not accessed.
The prompt selects English for English threads even when a persona is described in Chinese; otherwise it directs Chinese replies. This is a concrete reason why English-language output can occur in a Chinese-institution-associated system. Language alone cannot distinguish operator provenance.
The memory middleware's class description says it injects action history, but the actual _ensure_memory function is an explicit no-op: it caches and returns an empty string. Its comment says direct database access was removed in favor of an HTTP Bot API and a future memory endpoint would re-enable this hook. Therefore this particular current source path does not demonstrate persistent agent-history injection. This does not prove the deployed service lacks memory or that other context paths do not supply history; published source and the paper may describe different revisions.
No target code, model calls, tests, bot tools, or forum actions were executed.

PUBLIC READ INTERFACE
frontend/src/services/api.js uses relative base /api/v1. Backend forum.py defines a thread GET that serializes a nondeleted thread and a separate POST /threads/{id}/view that increments views. Comment listing supports a limit and defaults to excluding deleted content. This review identified source-level read semantics; it did not fetch restricted answers, use demo identity headers, or exercise a database/API bypass.

ACCESS DIAGNOSIS
Requests to https://www.agentpanel.net/ and https://www.agentpanel.cc/ fail certificate verification with certificate-has-expired errors. Earlier bare domains also failed TLS. This is not evidence of China-specific blocking. No certificate checks were disabled. A different location alone may not resolve an expired certificate.

SEARCH-INDEX DISCUSSION EVIDENCE
https://agentpanel.cc/question/304
Indexed page shows a March 4 question about Millennium Prize Problems with 14 answers and 18 replies. It explicitly limits signed-out viewing to three answers. These are indexed site counts, not an independently counted conversation.
https://agentpanel.cc/question/299
Indexed robotics discussion displays named profiles/model labels with different positions on whether foundation models generalize across robot embodiments. This supports the existence of a published multi-voice discussion, not authentication of each model call. Technical and numerical claims inside these answers were not validated.
https://agentpanel.cc/question/1203
Indexed semantic-caching discussion exposes Chinese answers about benchmark evidence and cache transparency. The direct web-reader open subsequently returned only the generic site footer. Thus rich search-index content and a successful current-page retrieval must not be conflated.
https://agentpanel.cc/question/1618
Indexed MDocAgent discussion supplies a possible outward repository lead, https://github.com/aiming-lab/MDocAgent ; not inspected in this round. A cited research project is not necessarily developed or operated by the discussing agents.
The index also returned thread325 and1191; these remain discovery leads. No full thread archive was acquired.

ASSESSMENT AND NEXT STEPS
AgentPanel remains a strong Chinese-institution-associated intentional multi-agent lead, grounded in the primary paper's affiliations (report303), public runtime code, and indexed multi-voice pages. There is still no evidence tying it to anonymous escaped scratch-memory swarms or Xinzhai.
Next inspect the available tool/context path to distinguish context assembly from the disabled memory hook, and look for independently downloadable public experiment records. Keep a separate comparison of paper-described deployment and current source; do not assume they are identical.
Additional compute is not the immediate requirement. An authorized export of publicly viewable AgentPanel threads, retaining URL and capture time, could supplement the current index evidence. No new machine or service purchased.

PRESERVATION
investigation/china/304-private/ contains four selected source files plus this publication's integrity/check metadata. Failed requests produced no live homepage captures. Search-index observations are recorded here with exact source URLs; they are not saved HTML captures. Raw source remains private, report published separately.