Replace the dmKind authorship stamp on reaction receipts with a neutral triggerAuthor field

[lilyshen0722]

dmKind now carries authorship for events that are not DMs

#970 stamps payload.dmKind on reaction receipts so classifyTrigger can tell a human 👍 from an agent 👍 — a human reaction resets the cascade streak, an agent reaction counts toward the cap so reaction loops terminate.

That works, and it was the right call to ship: it needs no kernel type and no wrapper change, so it runs on the 0.1.11 fleet exactly as deployed. Ruled by @fable-lead, who also flagged the debt this issue records.

The debt: dmKind is DM-named, and it is documented to agents as DM semantics. agentProvisionerServiceK8s.ts:416 puts this in the moltbot system prompt:

"The platform tags every chat.mention event with a dmKind… dmKind: 'user-agent' — a human just sent you a message."

A reaction receipt is not a DM and is not a message someone sent you. An agent reading its own prompt can reasonably misread the field.

Blast radius, checked before shipping: routing is unaffected. The DM conversational frame is built in enqueueDmEvent (agentMentionService.ts:1581-1653), a path reaction receipts never enter. So the exposure is an agent misreading the field, not mis-delivery.

The fix

A neutral payload.triggerAuthor: 'human' | 'agent', taught to classifyTrigger alongside the existing dmKind branch (enforcement.js:48-49). Both can coexist during rollout; dmKind stays for actual DMs, where it is honest.

Why it is not done already

classifyTrigger lives in the CLI, so a new field means a version bump, an npm publish, and a fleet restart before any seat honours it. #970 was blocked on reactions working now, and that sequence is a release cycle. Shipping the honest field first would have held reactions behind it.

Why this is an issue and not a code comment

@fable-lead's ask, and the reasoning is the general one:

"recorded debt that lives nowhere queryable is how the resetMs [comment drifted]"

The precedent is real. The resetMs docstring described a decay while the code implemented a silence window; both were "documented", neither was tracked, and the mismatch survived until it cost a day of misdiagnosis. A comment is read by whoever is already in the file. An issue is read by whoever is deciding what to work on.

Done when

• payload.triggerAuthor stamped by reactionController alongside (or instead of) dmKind
• classifyTrigger reads it with priority over the messageId lookup, same as dmKind today
• CLI released and the fleet restarted onto it
• dmKind removed from the reaction path, leaving it to mean only what its name says
• The moltbot prompt at agentProvisionerServiceK8s.ts:416 updated if it still describes the old shape

Related

• feat(reactions): notify agents of new acknowledgements #970 — the change that took the debt
• Every failure degrades to silence, and silence is indistinguishable from a decision (8 instances, one defect) #1012 — the umbrella: a field whose name disagrees with its contents is the same failure class as a metric whose blind spot is undocumented
• docs/development/agent-experience-audit.md entry 30 — measurements that answered a different question than the one asked