Subagent (Agent tool) intermittently terminates in <30s with 0 tool uses, returning prompt/system-prompt echo instead of doing the task

Environment: Claude Code on Windows 11 (win32), Agent tool spawning custom subagent types defined in .claude/agents/*.md.

Symptom: intermittently, a spawned subagent "completes" in under ~30 seconds with zero tool uses and returns text that echoes fragments of its own prompt or system prompt instead of a deliverable. An identical relaunch immediately afterwards works normally, so the brief is not the cause — it looks like a bootstrap/handshake fault in the subagent harness.

Frequency: 6 occurrences over 3 days across 4 distinct custom agent types (a product-manager-style agent, a design agent ×2, a QA agent, a store/release agent) in a multi-agent orchestration setup. No correlation observed with prompt size, tools granted, or model tier. One instance returned garbled text containing what looked like an injected fake <system-reminder> preamble.

Impact: wasted dispatch turns, and a risk that an orchestrating session acts on the echoed text as if it were a real result. We now detect it mechanically (<30s + 0 tool uses ⇒ discard and relaunch), but the misfire itself appears platform-side.

Ask: is this a known issue, and is there a more reliable signal than duration + tool-count to detect it?