[boogieLing]

Summary

• Pass the parent agent's configured max_iterations into SubagentManager
• Use that value for subagent AgentRunSpec instead of the old hardcoded 15
• Default directly constructed SubagentManager instances to AgentDefaults().max_tool_iterations
• Keep subagent limits synchronized when AgentLoop.max_iterations changes at runtime, including updates made through my(action="set", key="max_iterations", ...)
• Add regression tests for direct construction, AgentLoop wiring, runtime synchronization, and MyTool updates

Root Cause

The main agent loop already reads agents.defaults.maxToolIterations and stores it as self.max_iterations, but subagents did not use that setting. SubagentManager._run_subagent() always passed max_iterations=15 to AgentRunSpec, so long-running subagent work could stop early even when users configured a higher iteration budget.

Fix

This changes subagent iteration budgeting to follow the same runtime limit as the parent agent:

• AgentLoop passes its configured max_iterations into SubagentManager
• SubagentManager stores the value and passes it through to AgentRunSpec
• direct SubagentManager construction falls back to AgentDefaults().max_tool_iterations
• AgentLoop._sync_subagent_runtime_limits() keeps the subagent manager aligned before each agent loop run
• MyTool calls the same sync helper after changing max_iterations

Impact

Users who increase maxToolIterations now get the same budget for spawned subagents. This avoids premature subagent termination for longer research, polling, or multi-step background tasks while keeping the change scoped to the existing runtime limit.

Fixes #970.

Validation

• uv run --extra dev pytest tests/agent/tools/test_subagent_tools.py tests/agent/tools/test_self_tool_runtime_sync.py tests/agent/test_runner.py::test_subagent_max_iterations_announces_existing_fallback -q
• uv run --extra dev ruff check nanobot/agent/subagent.py nanobot/agent/loop.py nanobot/agent/tools/self.py tests/agent/tools/test_subagent_tools.py tests/agent/tools/test_self_tool_runtime_sync.py

Note: the targeted pytest run passes with one existing async queue warning from tests/agent/test_runner.py::test_subagent_max_iterations_announces_existing_fallback.

[Re-bin]

This is the right fix.

The bug was that the parent agent already respected agents.defaults.maxToolIterations, but subagents still used a hardcoded max_iterations=15. This PR wires the configured limit into SubagentManager, uses it for subagent AgentRunSpec, and keeps it synced when max_iterations changes at runtime through the my tool.

I merged latest origin/main locally and reran the relevant tests:

• python -m pytest tests/agent/tools/test_subagent_tools.py tests/agent/tools/test_self_tool_runtime_sync.py tests/agent/test_runner.py::test_subagent_max_iterations_announces_existing_fallback -q → 10 passed
• python -m ruff check nanobot/agent/subagent.py nanobot/agent/loop.py nanobot/agent/tools/self.py tests/agent/tools/test_subagent_tools.py tests/agent/tools/test_self_tool_runtime_sync.py → passed

The scope is narrow, the tests cover direct construction, AgentLoop wiring, runtime sync, and MyTool updates.

From my side, this is ready to merge.