latitude-dev/latitude-llm claude-code-telemetry-0.0.10

Claude Code Telemetry v0.0.10

on GitHub

Fixed

• Parallel subagents now each get their own trace subtree. When one turn spawned several Agent subagents at once, only one tool:Agent span received a nested interaction — the others showed no children. The stitcher matched subagents to their parent Agent call by promptId, which parallel calls share, so they collapsed onto a single call. Subagents are now matched by the toolUseId recorded in each subagent's .meta.json (unique per invocation), with promptId kept as a fallback for older transcripts.
• A subagent's final llm_request is no longer dropped. A subagent's transcript usually finishes flushing after its parent turn was already shipped (the final synthesis lands last), and the one-shot incremental read froze each subagent's offset before that row arrived, with no way to re-attach it. Subagents now emit as a standalone pass keyed off the parent Agent call's persisted span link: each subagent file is re-read every Stop and its spans are emitted incrementally, exactly once — a call once a later call closes it, and the trailing call plus interaction span once the file stops growing. Emitting each span once (rather than re-sending the whole subtree) keeps the additive traces rollup correct, since that view has no per-span dedup unlike the raw spans table.

Added

• subagent.name attribute (the agent type, e.g. Explore) on subagent spans, alongside the existing subagent.id and subagent.type.