Subagent's first response echoes spawn-time system-reminder fragments and stops with 0 tool calls

Environment

• Claude Code version: 2.1.201
• Platform: macOS (Darwin 24.6.0)
• Subagents spawned via the Agent (Task) tool — both custom agent types (.claude/agents/*.md) and built-ins, models opus/sonnet

Bug description

Intermittently, a freshly spawned subagent ends its very first assistant turn with zero tool calls, and its final text is an echo of harness-injected boilerplate (system-prompt / system-reminder fragments) instead of any task output. The task prompt is ignored entirely.

We have logged 9 occurrences over two days (2026-07-09 to 2026-07-10) across multiple sessions. In one session, 2 of 3 fresh spawns failed this way. Every occurrence has the exact same signature:

• first spawn of the agent (never on resumed agents),
• exactly one assistant turn,
• zero tool uses,
• final text = verbatim or near-verbatim harness boilerplate.

Examples of echoed content (each occurrence echoed a different fragment)

• The available-agents list: # Available agents, ...subagent types are available, specify a subagent_type parameter...
• MCP notice: The user's request may need one of the following MCP servers to complete. <connected_mcp_servers>...
• Auto-memory instructions: auto-memory files may contain..., check the index MEMORY.md...
• Skills notice: Only the following skills support arguments: ... When passing args to other skills, they will be ignored.
• Model/knowledge-cutoff lines: This session's model knowledge cutoff is January 2026.
• Fragments of the subagent's own system prompt

Hypothesis

System-reminder blocks injected at spawn time (skills list, agents list, MCP servers, memory instructions) appear to land in a position where the model treats them as the most recent user message, so it responds to them and stops, instead of executing the task prompt.

Recovery / workaround

• Re-sending the task to the same agent as a follow-up message recovers it every time (8/8 confirmed, 9th in progress) — the agent then executes normally, which suggests the task prompt is present in its context but was outranked on the first turn.
• We now run a SubagentStop hook that detects the structural signature (single assistant turn + 0 tool uses) and blocks the stop with a "perform your original task" reason. Content-based grep detection failed repeatedly because each occurrence echoes a different fragment.

Ask

Investigate spawn-time reminder injection ordering/positioning for subagents. Happy to provide sanitized transcript excerpts of the 9 occurrences.