Summary

The subagent-announce flow's "expectsCompletionMessage" branch (Branch 2) lacks the SILENT_REPLY_TOKEN guard that Branches 1 and 3 have. This causes the parent agent to be re-prompted to deliver child results even when the parent has already delivered them in a prior assistant message — producing user-visible duplicate messages.

Repro

1. Parent agent spawns a subagent for a verification task.
2. Parent receives intermediate progress, decides to deliver the result inline rather than waiting for the child's completion event.
3. Parent posts user-facing message with the result.
4. Child eventually completes; runtime fires the announce flow.
5. Parent receives synthetic user message: "A completed subagent task is ready for user delivery. Convert the result above into your normal assistant voice and send that user-facing update now."
6. Parent re-sends the same content. User sees duplicate.

Root cause

In subagent-announce-4mg0K78I.js around line 32, three branches build the prompt for the parent. Branches 1 and 3 include language like "Reply ONLY: ${SILENT_REPLY_TOKEN} if you have already delivered...". Branch 2 (the expectsCompletionMessage path) does not.

Result: the LLM has no canonical signal to short-circuit. Even with NO_REPLY discipline in agent prompts, the absence of the runtime-injected silent-token instruction makes duplicate suppression unreliable.

Proposed fix — minimal (Branch 2 SILENT guard)

Change Branch 2 from:

if (params.expectsCompletionMessage)
  return `A completed ${params.announceType} is ready for user delivery. Convert the result above into your normal assistant voice and send that user-facing update now. Keep this internal context private (...).`;

to:

if (params.expectsCompletionMessage)
  return `A completed ${params.announceType} is ready for user delivery. Convert the result above into your normal assistant voice and send that user-facing update now. Keep this internal context private (don't mention system/log/stats/session details or announce type), and do not copy the internal event text verbatim. Reply ONLY: ${SILENT_REPLY_TOKEN} if you have already delivered the substance of this result to the user in your most recent assistant message (look at the last assistant turn in your transcript above; if its content already conveys the same finding, this is a duplicate).`;

5-line change. Brings Branch 2 into parity with the other two paths.

Proposed fix — structural (recommended follow-up)

The above relies on the LLM correctly self-detecting duplicates. A more robust fix: in runSubagentAnnounceFlow, before building the announce prompt, check whether the parent session's most recent assistant message has a timestamp after params.endedAt AND fingerprint-matches the child's findings. If so, skip the announce entirely.

Pseudocode:

const recentAssistant = await getMostRecentAssistantMessage(targetRequesterSessionKey);
if (
  recentAssistant &&
  params.endedAt &&
  recentAssistant.ts >= params.endedAt &&
  fingerprintMatches(recentAssistant.text, findings)
) {
  return true;  // already delivered, no announce needed
}

fingerprintMatches: normalise both to lowercase + collapse whitespace, check for any contiguous ~60-char window match. Conservative — false negative (still announce) is far better than false positive (skip when actually duplicate).

This is the structural fix: runtime stops re-prompting when it has direct evidence the parent already covered the result.

Optional: spawn-side cooperative hint

Add a new spawn parameter finalAnswerWillBeFromParent: true. When set, force the announce branch to the SILENT-guarded path regardless of expectsCompletionMessage. Lets orchestrator agents declare "I will write the final answer myself" at spawn time.

subagent-spawn-Btdu5pK2.js line 652 area.

Effort

• Branch 2 SILENT guard: ~5 lines, trivial
• Structural short-circuit: ~30 lines + tests
• Cooperative hint: ~10 lines + schema entry

Context

Discovered during a series of orchestration tasks where Ghost (parent) delivered results inline and child completion events triggered visible duplicates. Branch 2 is the most-exercised path and has the highest user impact.