ai.raw.anthropic.claude_code.spans
This dataset is restricted by default: no one can read its data until access is granted through a target group rule on its access policy. To grant access, see Access control.
Purpose
The ai.raw.anthropic.claude_code.spans dataset stores the spans that Claude Code sends when tracing is on. Each trace follows one prompt through the model requests and tool calls it triggers, so you can see where a turn spent its time: waiting on the model, waiting on a permission decision, or running a tool.
ai.raw.anthropic.claude_code.spans is a predefined dataset: Coralogix provisions it in the default dataspace, and your Claude Code telemetry fills it. Its entity type is rawAnthropicClaudeCodeSpans. Claude Code log events go to a separate dataset, ai.raw.anthropic.claude_code.logs.
Claude Code spans used to land in your spans dataset. They now go here, so they no longer appear next to your application traces.
How spans reach the dataset
Claude Code sends spans only when tracing is on, which requires CLAUDE_CODE_ENHANCED_TELEMETRY_BETA=1 in its env block. Coralogix routes a span to this dataset when its OpenTelemetry resource attributes identify Claude Code as the source:
| Resource attribute | Value |
|---|---|
cx.integration.source.type | claude_code_agent |
Spans that carry the newer name, cx.source.integration.type, also route here. Claude Code metrics use only cx.integration.source.type, so keep that name. Spans without either attribute go to your spans dataset as before.
What one span is
Each span is one step of a Claude Code turn. Every turn starts a claude_code.interaction root span, and the model requests and tool calls it triggers are its children:
claude_code.interaction
├── claude_code.llm_request
├── claude_code.hook
└── claude_code.tool
├── claude_code.tool.blocked_on_user
└── claude_code.tool.execution
When the agent starts a subagent, the subagent's model request and tool spans nest under the parent's claude_code.tool span.
| Span | What it covers | Notable attributes |
|---|---|---|
claude_code.interaction | One prompt and everything it triggers. | user_prompt (redacted unless OTEL_LOG_USER_PROMPTS=1), user_prompt_length, interaction.duration_ms |
claude_code.llm_request | One model request, including retries. | model, input_tokens, output_tokens, cache_read_tokens, cache_creation_tokens, ttft_ms, duration_ms, success, stop_reason |
claude_code.tool | One tool call, from the permission check to the result. | tool_name, tool_use_id, duration_ms, result_tokens |
claude_code.tool.blocked_on_user | Time spent waiting for a permission decision. | decision, source, duration_ms |
claude_code.tool.execution | Time spent running the tool. | success, error, duration_ms |
claude_code.hook | Hooks that ran for one hook event. Sent only with Claude Code's detailed beta tracing. | hook_event, hook_name, num_hooks, duration_ms |
Every span also carries a span.type attribute that matches its name, plus the same session and user attributes as the log events, such as session.id and user.email. Use session.id to join spans with log events. The tool_use_id on a claude_code.tool span matches the one on the claude_code.tool_result and claude_code.tool_decision log events for the same call.
For every span attribute, see the Claude Code monitoring reference.
Content capture
Spans redact prompt text, tool input details, and tool content by default. The same environment variables that add content to log events add it to spans: OTEL_LOG_USER_PROMPTS=1, OTEL_LOG_TOOL_DETAILS=1, and OTEL_LOG_TOOL_CONTENT=1.
Query the dataset
source ai.raw.anthropic.claude_code.spans
| filter $l.operationName == 'claude_code.llm_request'
Next steps
- Set up the source: Claude Code.
- Grant read access to more users: Access control.
- See the other datasets Coralogix provisions in Predefined datasets.