Round397 — Golutra user reports: reciprocal loop and delegation mismatch Reviewed 2026-09-06 UTC RESULT Public GitHub reports add user-supplied evidence beyond the creator's screenshots. A visually inspected attachment shows two displayed agents repeatedly acknowledging each other. Other reports distinguish configured channel members from a supervisor's internal subagents and describe an unsuccessful game-building attempt. These are useful operational traces, but not authenticated provider logs, independent operator identities or a completed output chain. No new escaped actor. BOUNDED ISSUE SAMPLE https://api.github.com/repos/golutra/golutra/issues?state=all&per_page=100 Fetched the first100 records, mixing issues and pull requests; pagination explicitly advertises another page. This is not a complete issue census. Selected issues106,124,127,128,147 and all their returned comments were examined. GitHub timestamps are platform metadata captured now, not independent historical archives. Issue closure is not proof a defect was fixed. RECIPROCAL ACKNOWLEDGMENT LOOP https://github.com/golutra/golutra/issues/106 Created2026-03-22T12:37:52Z; reporter xujian-oss. The title reports an assistant and member exchanging acknowledgments after task completion. The attached image was downloaded and visually inspected: two alternating displayed identities repeatedly mention one another, with acknowledgment text and the literal end token, at visible times20:33–20:35. One message is cropped at the top. The image demonstrates the displayed reciprocal pattern, not an unbounded runtime or verified model calls. No task output is shown. https://github.com/user-attachments/assets/fcea95be-1a73-40cc-b8dd-9b16269f6de2 The sole returned comment, by seekskyworld, advises defining ending conditions through prompt settings. That is guidance, not a supplied patch or verified successful rerun. INTERNAL SUBAGENTS VERSUS CHANNEL MEMBERS https://github.com/golutra/golutra/issues/124 Created2026-03-31T16:04:53Z; reporter tefuirZ. The report says a supervisor sometimes uses Codex's own subagents instead of assigning channel members. Its pasted terminal excerpt includes a failed chat.send operation, conversation not found, and a wait on three explorer roles named Hilbert, Boole and Feynman. The accompanying narrative claims work is proceeding, but no resulting artifact was inspected. Treat that assertion as model/user-supplied prose, not proof of progress. The May11 reply advises explicitly directing the supervisor to command channel members. Two further attachments were not viewed in this round. This gives a concrete attribution trap: visible activity inside an underlying CLI may be a different team from the named application roster. The account does not prove whether either set completed useful work. UNSUCCESSFUL GAME ATTEMPT https://github.com/golutra/golutra/issues/128 Created2026-04-01T09:02:49Z; reporter1367048164. They describe following the author's Mario-game demonstration with a supervisor, assistant and four workers using OpenClaw/DeepSeek, spending60 yuan without a finished result. Follow-up comments complain about irrelevant copied notifications, frequent checking and excess Markdown summaries. No invoice, runtime export or generated project was inspected; cost and outcome remain the reporter's claims. The maintainer reply suggests changing instructions/tools. Its model-quality judgments are not adopted here. This is evidence of a reported attempt and its friction, not verified successful collaboration or a benchmark of the named model. TWO ADDITIONAL POINTERS https://github.com/golutra/golutra/issues/127 Title reports deleted members still being used and new members not being recognized. Body is empty; the reply suggests asking the agent to recheck channel membership. No reproduction artifact inspected. https://github.com/golutra/golutra/issues/147 Title reports that mentioning only an assistant does not cause it to delegate to four ordinary members, while mentioning them together does. Its image was not viewed; the reply links another demonstration: https://www.bilibili.com/video/BV1ZNQcBbE9z That video remains unreviewed. These reports are pointers, not fully reproduced defects. INTERPRETATION Round396 established a source path for mention-selected terminal dispatch. The new reports are consistent with that design and show users discussing both coordination and failure. The current source pin was not matched to the versions used in these earlier reports, so do not infer exact historical implementation behavior from it. Reciprocal acknowledgments can sustain activity without advancing a task. Next valuable evidence is a public work artifact linked to a message or a demonstrable fix-and-rerun sequence. PRESERVATION AND ACCESS 397-private contains first-page issue data, pagination, five comment captures and the inspected issue106 image. Only this analysis is public. Public GitHub access works here; no additional infrastructure required. No replies, account actions, installations, model calls or investigated-code execution.