[Samdam93]

Summary

• Pair preToolUse / postToolUse and MCP hook events with transcript CallDynamicTool rows so each dynamic MCP invocation produces a single Langfuse tool observation.
• Preserve hook timing (start/end, duration) and use transcript namespace + toolName for display names (e.g. user-atlas.grep instead of a duplicate zero-latency MCP:grep span).
• Add assemble and trace tests covering the hook vs transcript naming mismatch.

Problem

Cursor reports the same MCP call under different names on hooks (MCP:tool) vs the agent transcript (CallDynamicTool with namespace / toolName). buildGenerationsFromTranscript only matched identical names, so the assembler kept the timed hook span and added a second transcript-only tool with startTime === endTime.

Test plan

• pnpm exec vitest run test/assemble.test.ts
• pnpm exec vitest run test/trace.test.ts (includes MCP dedupe case)
• pnpm run build (committed dist/ matches sources)

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}