Instrument tool calling to track tool invocations, arguments, and results

[psschwei]

Description:
Instrument tool calling to track tool invocations, arguments, and results.

Detailed Requirements:

1. Add spans for tool operations:
   • tool.call - Tool invocation
2. Add attributes:
   • tool.name - Name of the tool being called
   • tool.arguments_hash - Hash of arguments (for privacy)
   • tool.status - success/error
3. Add span events for:
   • Tool call start with argument summary (redacted)
   • Tool call completion with result summary
   • Tool call error with exception details
4. Handle tool call extraction from model responses
5. Track multiple tool calls in single response

Files to Modify:

• mellea/stdlib/tools/interpreter.py - Tool execution instrumentation
• mellea/backends/tools.py - Tool call extraction
• mellea/helpers/openai_compatible_helpers.py - OpenAI tool handling

Span Hierarchy Example:

backend.request
├── [model returns tool calls]
└── tool.call (tool_1)
    └── [tool execution]
└── tool.call (tool_2)
    └── [tool execution]

Acceptance Criteria:

• Each tool invocation has a span
• Tool name and status captured
• Arguments handled with privacy in mind
• Multiple tool calls tracked separately

[ajbozarth]

Closing as part of the tracing epic (#444) re-plan. This sub-issue predates the plugin/hooks pattern (PR #653) and the GenerationMetadata partition (PR #908); its spec has drifted.

Replacement: #1049 — tool call spans via the tool_post_invoke hook.

See #444 for the full refreshed plan.