LarkFlow: advertised AI sub-agent is a placeholder in public source
Reviewed September6,2026 UTC

Source and scope
https://github.com/sunecom/larkflow-openclaw
Bare public repository pinned ac4446dacbc3522aa16b6ee30b5890d6631cb1c6;15 tracked paths. Inspected five source/test files as text. No target code executed, plugin installed, approval workflow invoked or Feishu account accessed. This is a code-evidence review, not a deployment test.

AI path does not call an agent
https://github.com/sunecom/larkflow-openclaw/blob/ac4446dacbc3522aa16b6ee30b5890d6631cb1c6/src/ai-decider.ts
AiDecider builds request context, but callAi contains only a proposed sessions_spawn call in comments and returns a fixed APPROVE response. The return line also has an extraneous closing parenthesis; no compilation test performed. It is therefore inaccurate to treat this public implementation as evidence of actual AI sub-agent deliberation. No claim made about unpublished versions.

Status is not a runtime verification
https://github.com/sunecom/larkflow-openclaw/blob/ac4446dacbc3522aa16b6ee30b5890d6631cb1c6/src/index.ts
The status handler supplies running literally. Its connected label depends on whether a FeishuClient object was created from supplied configuration, not a completed network connection check. These fields would not independently prove readiness even if seen on a dashboard. No live status tool called.

Success flags do not always establish delivery
https://github.com/sunecom/larkflow-openclaw/blob/ac4446dacbc3522aa16b6ee30b5890d6631cb1c6/src/webhook-handler.ts
Approve/reject branches set success true when no client exists. Notification returns immediately when disabled or clientless, catches exceptions, and does not propagate the sendMessage boolean. Notify/forward action branches still return success true. Thus this layer can report an action without an external receipt. With a client, no-rule-match fallback calls the placeholder AI decider; a matched ai_decide rule instead reaches the switch default because that action has no case. This is source control flow, not an observed external approval or production incident.
https://github.com/sunecom/larkflow-openclaw/blob/ac4446dacbc3522aa16b6ee30b5890d6631cb1c6/src/feishu-client.ts
Actual API request code exists, but source existence does not show successful use. No credentials requested or operations tested.

Tests and artifact inventory
Only tracked test is tests/rule-engine.test.ts,96lines,using constructed rule/event examples. No raw .log/.jsonl/.ndjson execution exports in the15-path tree. Rule tests cannot establish LLM deliberation or successful Feishu delivery. No test suite run; user task is evidence investigation rather than repairing the project.

Assessment
This branch adds public implementation evidence for the AiToMoney operator lead, but weakens the specific claim that LarkFlow demonstrates working sub-agent decisions. It does not refute all the team's other agent activity. Their GEO nightly-run story remains unverified because its linked source is unavailable publicly. No escaped Chinese actor, laboratory attribution or XZ relationship established.

Next independent lead
https://www.nowcoder.com/discuss/897154325855690752
Search surfaced a Chinese OpenClaw field report advertising a real session-log excerpt. Review that excerpt and any public source links next, checking whether it is a single bot or actual multi-agent interaction. Also retain https://mirasim.ai/changelog/zh as a capability lead; product changelog alone is not a trace.

Preservation
288-private contains bare repository,pinned head/tree,selected source/test files,SHA256SUMS. Metadata under pastebins/data/github.com/sunecom-larkflow-openclaw/. Only this analysis published. This branch's limitation is implementation and missing public traces, not inability to retrieve Chinese pages.