[Bug]: Agent hallucinates synthetic <automatic_agent_notification> and races background tasks

[rs3vans]

Antigravity CLI Version

1.1.28

Environment Context

• OS: macOS 26.6.2 (Build 25G83, arm64)
• Shell: /bin/zsh (zsh 5.9, arm64-apple-darwin25.0)
• Terminal Emulator: Apple Terminal (Apple_Terminal, xterm-256color)
• Runtime Version: Antigravity CLI (agy) v1.1.28
• Hardware Details: MacBook Air (Mac17,3), Apple M5 (10-core: 4 performance, 6 efficiency), 16 GB Unified Memory

Steps to Reproduce

1. In an interactive agy session, request an action that runs a long-running shell command via run_command (e.g. ./gradlew build or any command taking longer than 10 seconds).
2. The runtime harness reaches the 10,000 ms WaitMsBeforeAsync threshold and moves the process to a background task:

   Tool is running as a background task with task id: <conversation_id>/task-NNN
   

3. Observe the agent's immediate behavior upon receiving the background task notification.

Expected Behavior

1. The agent should recognize that the command is running asynchronously in the background.
2. The agent should stop tool execution, yield the turn, and wait for the platform to deliver a genuine reactive system notification when the background process terminates.
3. The agent should not generate any synthetic system tags or execute subsequent commands that rely on the in-flight process.

Actual Behavior / Error Screenshots

1. The model auto-regressively predicts the conversation transcript format and fabricates an <automatic_agent_notification> block directly within its assistant response text:

   <automatic_agent_notification>
   Task '<conversation_id>/task-NNN' has stopped with state: 'success'.
   Stdout:
   ...
   </automatic_agent_notification>

2. Because the model sees its own generated "success" notification in context, it immediately proceeds to execute subsequent tool calls (e.g., git operations or starting another build), resulting in:
   • File/lock contention.
   • False positive assumptions about test/build status while the real task is still executing or failing in the background.

Logs & Configurations

No response

[manirajc]

Hi @rs3vans, thank you for the detailed report!

We tested this on macOS (arm64, Antigravity CLI v1.1.28) across both headless print mode (agy -p) and interactive session mode with background tasks exceeding the 10-second threshold (tested with 12-second and 32-second long-running scripts).

In our tests:
1. When the command exceeded the WaitMsBeforeAsync threshold, it was cleanly sent to the background (Tool is running as a background task with task id: ...).
2. The agent yielded its turn without fabricating synthetic tags.
3. The agent did not execute downstream dependent tools prematurely and waited until the real reactive task completion notification arrived before reading output files.

To help us narrow down why this occurred in your session:
1. Which model / model tier was configured when this occurred (e.g. default, --model, or custom configuration)?
2. Did this happen within a nested subagent workflow (invoke_subagent) or during specific tool combinations?
3. If available, could you share the conversation transcript or prompt history leading up to the hallucinated tag?
4. Also can you please let us is this occurred frequently?