Background subagent's first completion was another conversation's content (crosstalk); ignored its own tasking prompt

Bug: background subagent received/echoed another conversation's content

Summary

A freshly-spawned background Agent (subagent_type: general-purpose) returned, as its FIRST assistant message, the verbatim text of a user-style question that was never part of its prompt, followed by the harness's small-request guidance template ("The user's exact request is small. Focus on completing it, and resist the urge to gold-plate…"). It made 0 tool calls and ignored its actual tasking prompt entirely.

Environment

• Claude Code CLI on macOS (Darwin 25.5.0), model claude-fable-5, session 2026-07-05 ~09:11 SGT.
• The agent was spawned via the Agent tool with a long task prompt (web-data collection). Remote Control was enabled for the session.
• Parallel context: 3 other background agents had just been spawned in the same session; a queued user message was delivered to the MAIN loop at approximately the same time.

Evidence (from the subagent transcript JSONL)

• entry 1 (user role): the correct tasking prompt — delivered intact.
• entries 2–3: harness attachments (deferred_tools_delta, skill_listing) — normal.
• entry 4 (assistant role, FIRST completion): verbatim — "When was the last time each of these skills was used? For assess, drill, teach - check my notion." followed by the harness anti-gold-plating reminder template. 0 tool calls, ~5s duration.
• The echoed question does NOT appear in the user's typed-prompt history (~/.claude/history.jsonl, 6,215 entries, zero matches) and was not sent by the orchestrator (all SendMessage calls accounted for).
• Downstream effect: when the orchestrator re-tasked the agent via SendMessage, the agent treated the echoed question as the "real user request" and refused the actual task as a suspected injection — i.e. the crosstalk also poisoned its trust model.

Hypothesis

Completion or message routing cross-wired between concurrent requests (possibly another session/Remote-Control message or another sampling request's prompt-side payload returned as this agent's completion). The presence of the harness's own reminder TEMPLATE inside an assistant completion suggests the content originated harness-side, not model confabulation.

Repro likelihood

Observed once (2026-07-05 ~09:11 SGT) under: 4 near-simultaneous background agent spawns + a queued main-loop user message + Remote Control enabled. Not yet reproduced.