latitude-dev/latitude-llm claude-code-telemetry-0.0.2

Claude Code Telemetry v0.0.2

on GitHub

Added

• Workspace and git context on every span — the hook now reads the Claude Code session's cwd and derives workspace.name / workspace.path, git.branch / git.commit / git.repo, claude_code.version, host.user, and hook.event. The workspace name is attached as a span tag (latitude.tags); the full set is attached as shared trace metadata (latitude.metadata) on every emitted span so traces can be sliced by repo, branch, or CLI version in the Latitude UI.
• Full conversation history on llm_request input messages — gen_ai.input.messages on the first LLM call of each turn now contains every prior user/assistant turn plus the current user prompt, matching the context actually sent to the model. Subagent turns accumulate their own isolated history from prior turns within the same Agent invocation.
• Tool calls embedded in the conversation — assistant messages now carry {type: "tool_call"} parts inline alongside text, and tool results are emitted as {role: "tool"} messages with tool_call_response parts. Matches the OpenTelemetry GenAI semantic conventions, so the Latitude UI can render tool invocations in-context instead of only as disconnected tool spans.

Changed

• One llm_request span per model call (tool-loop aware) — a single user turn that triggers a tool loop now emits N llm_request spans as siblings under the interaction span, one per distinct assistant message.id. Tool executions are siblings of the llm_request spans (also parented to the interaction), reflecting that generation and tool execution are sequential, not nested. Previously a tool loop collapsed into a single llm_request with tools as children, which hid the request/response structure and double-counted input tokens across calls.
• Full accumulated conversation on every llm_request input — each llm_request now carries the full conversation that reached the model on that step (session history + current user prompt + every prior call's assistant message with tool_calls + tool responses), not just the delta since the previous call. Matches the billed input_tokens and lets the UI read each span standalone.
• Proper waterfall timing per call and per tool — each llm_request span now runs from the prior phase boundary (user prompt or preceding tool result) to its last-flushed transcript row, and each tool span runs from its emitting call's end to the tool_result row. Single-row calls (typical for the final assistant text) are floored to a 1ms minimum duration so they always render.
• Per-call token attribution — input_tokens / output_tokens / cache counters are now reported per llm_request span instead of summed across all calls in a turn. Fixes inflated totals for multi-call tool loops (e.g. a 3-call turn no longer triple-counts the base context).
• Canonical tool-span attributes — tool spans now emit gen_ai.operation.name="execute_tool", gen_ai.tool.name, gen_ai.tool.call.id, gen_ai.tool.call.arguments, and gen_ai.tool.call.result, replacing the non-standard tool.name / tool.input / tool.output keys. Tool failures additionally set error.type="tool_error" and OTel status code 2.
• Tool span timestamps are per-call — tool spans now run from the emitting assistant row's timestamp to the tool_result row's timestamp, instead of inheriting the whole turn's bounds. The waterfall now reflects actual tool latency.