Retest record (2026-10-07, on current main 62210ee8 + KUMA 0.3.4, model glm-5.3-flash):

Blocking — must be fixed before merge. The PR contains only 8 files: agent/ was excluded by the /resources/agents/*/agent/ rule in .gitignore, and agent.toml does not declare Git acquisition. On main, load_registry validates every entry, so merging this would make agentbench run / evaluate fail for all Agents, not only kiroku. Either fix works:

• add method = "git" under [source] in agent.toml (same as Add 50-zhisaotong-agent (bamboo-moon/zhisaotong-Agent) #288 and 12-minimax-code; ABB then restores 4b09df8 into agent/ before evaluation — this is what I used for the retest), or
• commit the source with git add -f resources/agents/50-kiroku/agent.

Non-blocking, but worth addressing:

1. KirokuBridge.invoke ignores the config ABB passes in (bindings/kiroku_bridge.py:312-330), so no callbacks/tags reach the graph, and each invoke builds a new DocumentWriter + MemorySaver, ignoring ABB's thread_id. Multi-step Cases therefore have no memory.
2. The document specification is fixed (title, sections, max_revisions=1; :117-137), so the requested topic is not carried into the spec — this is what the KUMA Judge flagged. This is a binding design choice, not an upstream kiroku defect.
3. The binding clears interrupt_before / interrupt_after on the compiled graph (:221-222) and resumes with an empty instruction. No feedback is fabricated, but the review stages are effectively skipped; worth stating in requirement.md.
4. The 50- prefix collides with Add 50-zhisaotong-agent (bamboo-moon/zhisaotong-Agent) #288 / onboarding: add M-Cube (yycyyv/M-Cube) draft workflow agent (issue #238) #289.

@YaoAnthony the integration runs once the source acquisition is fixed; please don't merge before that, since it breaks registry loading for everyone.

Retest of 8f6d8a88 (2026-10-07, main 62210ee8).

The blocker from my previous comment is still present. The new commit adds only ground_truth/ files, and load_registry still fails on main:
FileNotFoundError: Agent source directory does not exist: resources/agents/50-kiroku/agent.
Merging this would stop run / evaluate for every Agent. Add method = "git" under [source] in agent.toml (or git add -f the source).

Ground truth defect-1. ground_truth/manifest.json passes ABB's load_manifest (hashes and structure are valid), but I don't think the defect belongs to kiroku:

• Upstream keeps one DocumentWriter per session; each further user instruction calls update(instruction) on the same thread and draft (agent/kiroku_app.py around :299-330).
• The binding instead builds a new DocumentWriter and checkpoint thread for every input on purpose (bindings/kiroku_bridge.py, docstring around :59 and :191-261).

So "ignores the conversation history" is produced by the integration, not by upstream behaviour. As written, the manifest would record a binding artefact as an Agent defect. I'd suggest either making the binding keep one writer per ABB thread_id (then re-checking whether the behaviour still occurs), or dropping defect-1.

@YaoAnthony please hold this one until the source acquisition is fixed and the ground-truth attribution is resolved.

感谢复测，问题全部属实，已逐条修复（commit 18e03b3）：

Blocking — 源码获取：已按方案 a 处理，[source] 下加了 method = "git"。本地做了合并态验证：把 agent/ 移走后跑 evaluate，ABB 自动按 4b09df8 恢复了源码，Docker local smoke 1/1 通过（judge: pass）。不会再拖累 registry 加载。

前缀冲突：50-kiroku → 51-kiroku（registry.toml 同步），避开 #288/#289。

1. config/thread_id 被忽略：重写了这块。现在 invoke 会透传传入的 callbacks/tags/metadata，thread_id 也按 ABB 给的来；writer 和 checkpoint 按 thread 缓存复用，不再每次新建。DocumentWriter.invoke 透不出 config（还会把 resume 的 None 返回值搞崩），所以改为直接驱动 writer.graph.invoke。效果可验证：修复后同一 smoke 的 step-3 能基于上文作答了（judge: pass, medium，评语 'responsive to the conversation'，之前有轻微表述偏差但已响应上文）。

2. 题目没带进规格：title 不再写死，取输入文本首句（截 80 字符），让题目真正进入原生写作任务。你说得对，这是绑定的设计问题，不是上游缺陷——ground_truth 的 defect-1 记录的是修复前绑定下的行为（step-3 无记忆那条在旧代码上两次复现；题目跑偏那条归因已按你的意见修正）。

3. 人工审查被跳过：requirement.md 的 Known Limitations 里已写明——人工审查阶段在本部署中是被自动批准跳过的（空指令恢复，不伪造反馈，但也没有人真的审），并同步更新了 config 与记忆相关的过时表述。

再麻烦看一眼，还有问题随时说。

Retest of 5b41e24f (2026-10-07, main 62210ee8, runtime model qwen3.8-max).

The blocker is fixed. Thanks!

Remaining points:

1. Ground truth defect-1 should go. It records the "ignores the conversation history" behaviour of the old binding. With this binding the same smoke Case now passes, so the defect no longer reproduces on the code being merged, and ground truth is meant to describe Agent behaviour. I'd remove ground_truth/manifest.json + observations/defect-1/ (keep .gitkeep).
2. Numbering: 52-kiroku now collides with Add local-rag-researcher-deepseek onboarding unit (#223) #297 (52-local-rag-researcher-deepseek). Whichever merges first keeps it.

@YaoAnthony the integration is good now; point 1 before merging, please.

两点都处理完了（commit b1759e6）：

1. defect-1 已删：认同归因判断——旧绑定每个输入新建 writer 才导致看不到历史，不是上游行为（上游有 update(instruction) 同会话续写）。修复后同一 Case 复查已通过，缺陷在待合并代码上不复现，ground_truth 不该记接入产物。已删 manifest.json + observations/defect-1/，保留 .gitkeep。

2. 编号：52 和 Add local-rag-researcher-deepseek onboarding unit (#223) #297 撞了，已改 53-kiroku（查过 main 和所有开放 PR：50=Add 50-zhisaotong-agent (bamboo-moon/zhisaotong-Agent) #288/onboarding: add M-Cube (yycyyv/M-Cube) draft workflow agent (issue #238) #289、51=Add 51-cra-agent (kulkarnirohit123/cra-agent) #314、52=Add local-rag-researcher-deepseek onboarding unit (#223) #297、54=Add 54-beacon (123-qw-as/Beacon) LangGraph unit #319，53 没人占）。

辛苦再跑一眼，应该齐了。

按建议完成两项调整并重跑通过（commit f3574a0）：

• step = 1：单轮文档工作流不再被多轮客服题误伤；
• 策略组 CAND-009@1：研究类题目，KUMA 本轮出了远程办公×城市经济的研究写作题；
• 顺手修了标题在 80 字符处截断丢词的问题（你复测时也看到了 'impact of remote work on ' 这个残题），现在按词边界截断、上限 120。

结果：official evaluation 1/1，judge pass（high confidence，零 issues），run_0a19046831814960bd580b2e92ba5b43，trace complete。前一轮 judge 指出的 Abstract 多段/标题截断/句数超标三项，本轮均未再出现。

Retest of f3574a03 (2026-10-08, merged onto current main 06afd79f, KUMA 0.3.4, runtime model qwen3.8-max).

I pushed one small commit on top of your branch (e170c5a7), since maintainerCanModify is on:

1. ENV GRADIO_ANALYTICS_ENABLED=False in the Dockerfile. kiroku_app.py imports gradio, which calls api.gradio.app on import. That host is not routed, so the egress proxy denied it on every run. With the variable set, the retest above shows no egress denials.
2. enabled = false in registry.toml. Every onboarded unit is registered disabled while status = "adapting".

Note: the trace capture status is partial with reason trace_value_truncated. That is the KUMA SDK capping one long span attribute (the generated document). No span was dropped (dropped_count = 0), so nothing to change here.

From my side this is ready to merge.

cc @YaoAnthony

谢谢！
辛苦 @YiWang24 两轮复测还顺手代改代码，也谢谢 @YaoAnthony 一路批 credit、指路和最后把关。合并后如果还有要补的（认证、后续 ground truth 之类），随时叫我。