Intermittent: subagents hang on their first tool call (0 tool uses, no error, parent waits forever)

Environment: Claude Code 2.1.207, Linux VM (Hyper-V, kernel 7.0.0-27), Opus 4.8 at high/max effort, Agent tool (general-purpose subagents), local execution.

Symptom: intermittently, a spawned subagent emits its opening text ("I'll read the file…") and hangs at its first tool call — the transcript shows zero tool_use blocks, no error is surfaced, and the parent session waits indefinitely until the subagent is killed manually. When the condition is present it affects every subagent spawned from that session — parallel fan-outs AND a single clean diagnostic subagent with a trivial read-one-file prompt — which suggests session-level rather than prompt-level state.

Occurrences: overnight 2026-07-15→16, a 24-batch subagent fan-out lost every subagent this way and fell back to serial main-loop execution (hours lost, measurably noisier output). A fresh session the same day mostly succeeded but hit two more first-tool-call hangs in one batch. The same night we also saw misrouted-content cases: a subagent transcript containing content apparently belonging to a different concurrent subagent.

Repro: no deterministic repro; intermittent and clustered in time. Local transcripts of the affected parent/subagent sessions are retained and available.

Asks: (1) is this a known issue in 2.1.x? (2) is there client-side debug logging we can enable to capture the hang state on next occurrence?