Comparing schemas
What the other efforts actually support
Turnclaude_code.interaction
one per user message sent — turn #1, #2, #3… in the session, plus its text
Who & Where
which user, which session, terminal type (e.g. VS Code) — on every span
none capture this
Requestchat <model>
one per call to the model — what we sent it + what it replied
Tool Callclaude_code.tool
tool name + args, nested execution
Token Accounting
input/output token counts per call
Full Message Body
the actual prompt sent & the model's reply, verbatim; redacted at save time
Request Params
temperature, max_tokens as sent to the model per call
Prefix / Cache Structure
shared-prefix or cache-hit signal across requests
Timing & Nesting
span start/end, parent_span_id; subagents nest under caller
Tool Executionclaude_code.tool.execution
the tool actually ran — success / error
User Gated Decisiontool.blocked_on_user
permission/clarification prompt + the user's decision
none capture this
Tool Definitions
the full set of tools available to the model on a call, not just which ran
Each box lists the datasets that record that field. A dashed, faded, italic name means partial support (hover it for details). "None capture this" means no dataset here records that field. Two caveats: Baseten has no tool fields because it replays /v1/completions, which has no tool calling to record. SWE-smith is shown for its tool split, the richest one; its xml and ticks splits turn tool calls into prose. Click a dataset in the legend to highlight its fields, and click again to clear.