[BUG] Background subagent final response delivered as thousands of duplicated <acp> message-ref tags instead of report text #43746

Description

A background subagent (task tool, background=true, general agent) completed its work correctly on disk, but the final response delivered to the parent session was corrupted: instead of the actual report text, the parent received ~11K tokens consisting of the literal string <acp tokens=N type=text>mNNNNN</acp> repeated hundreds of times, referencing the child session's own internal message IDs (e.g. m00001–m00086). Zero actual report text was delivered.

This looks like the message-reference annotation pass (ACP/bili-style context-management tags rendered as <acp ...>mNNNNN</acp> prefixes) recursing on or echoing its own output while serializing the child→parent completion payload.

Environment

• OpenCode harness with ACP context-management (bili) tooling
• Platform: Windows 11, PowerShell
• Subagent type: general, launched with background=true
• Session context: long-running session (~10 prior subagent completions, heavy bili_compress usage)

Steps to reproduce

Not reliably reproducible yet. Observed twice-shaped failure modes in one session (~10 background subagent runs):

1. This bug: child finishes real work (verified independently on disk), final message arrives as thousands of duplicated <acp> tag lines instead of text.
2. Related: another run same day ended state=completed with no text at all and no work performed — matches [BUG] Subagent terminates prematurely mid-execution but reports completed #32132.

Observed behavior

• Parent received a completion event with an ~11K-token message containing only repeated <acp tokens=... type=text>mNNNNN</acp> fragments
• The referenced mNNNNN ids belong to the child session, not the parent
• No error surfaced anywhere; state was completed

Expected behavior

• Final assistant text is delivered verbatim exactly once, with at most one annotation prefix per part
• If serialization fails or final text is empty, the completion should fail loudly (or retry) rather than deliver garbage/silence

Impact

Parent agents silently lose subagent reports and must independently verify all work (which is what happened here — the work itself was fine). At ~20% of background completions misbehaving in one session, this materially erodes trust in fire-and-forget delegation.

Suggested fix

Sanitize/validate the child→parent completion payload: render annotations exactly once, strip nested/self-referential tag output, and treat an empty final text as an error state rather than a successful completion.