Bug Description

Delegation observability was misclassifying successful tool calls as errors in tool_trace.

Root Cause

tools/delegate_tool.py used a naive string heuristic that marked a tool result as an error whenever the returned content contained the word error near the start. That incorrectly flagged successful terminal JSON payloads like {\"error\": null, \"exit_code\": 0, ...}.

Fix

• parse tool result JSON before classifying status
• treat only real error states / non-empty errors / failed success flags / non-zero exit codes as errors
• keep a conservative text fallback for non-JSON tool payloads
• add regression tests covering error: null success payloads and non-zero exit codes

How to Verify

1. Run source venv/bin/activate && python -m pytest tests/tools/test_delegate.py -q
2. Call delegate_task with toolsets=[\"terminal\"] for a read-only command like hostname && pwd
3. Confirm the returned tool_trace[0].status is ok instead of error

Test Plan

• Added regression test for this bug
• Existing tests still pass
• Manual verification of the fix

Risk Assessment

Low — the change is narrowly scoped to delegation observability/status classification and is covered by targeted regression tests.