`socket.on('message.created', ...)` dispatched with
`accountId: agentId` (the openclaw agent id, e.g. 'analyst2') instead
of the binding's `match.accountId` (the fabric account slot label, e.g.
'interviewee'). For most agents the binding is `{agentId: X, accountId: X}`
so the two coincide and the call works by accident. For shared-placeholder
slots (the recruitment `interviewee` apikey reused across pre-onboard
agents) the slot label is `interviewee` not the agent_id, so the lookup
returns bindings=0 and openclaw core silently falls back to the `main`
agent — which then handles the sub-discussion turn under main's
workspace identity. Symptom: every sub-discussion interview reply
masquerades as the human user's IDENTITY.md text.
Walk cfg.bindings for the entry that ties this agentId to a fabric
account; use its accountId. Fall back to agentId when the agent has
no explicit fabric binding declared (preserves prior behavior for
agents wired before the binding format was uniform).
Verified on prod-t2 recruitment retest 2026-05-31:
Before: routing log `accountId=analyst2 ... bindings=0`, main session
ran instead of analyst2.
After: `accountId=interviewee ... bindings=1`, analyst2 session ran
(0 main sessions in sub channel).
