Model emits only 1 Task tool call per message when intending multiple parallel calls

Description

When the model spawns multiple Task (subagent) tool calls in a single message for parallel execution, there are two issues:

1. Ctrl+O task viewer only shows one task — even when 3 agents are running concurrently, the task viewer (ctrl+o) only displays one of them.
2. On subsequent attempts, only 1 of N intended Task calls may actually be emitted — the model writes a response intending 3 parallel Task calls, but only 1 is actually executed. The other 2 "results" are hallucinated (plausible-looking but fabricated). This happened on the 2nd and 3rd attempts in the same conversation. It's unclear whether this is a model-side issue or a client-side issue where the calls are made but dropped.

Reproduction

1. Ask Claude Code to spawn 3+ agents in parallel (e.g., "spawn 3 agents to do different random tasks so I can see what parallel looks like")
2. On the first attempt, all 3 may run but ctrl+o only shows 1
3. On subsequent attempts, only 1 Task call may be emitted, with the other results hallucinated

Expected behavior

• All Task tool calls in a single message should be emitted and executed in parallel
• Ctrl+O should list all currently running/completed tasks

Related

See #29182 for a separate TUI rendering corruption issue when agents run in parallel.

Environment

• Claude Code (CLI)
• Model: claude-opus-4-6
• macOS 15 (Darwin 25.3.0)
• iTerm2 + tmux