Fix logging for parallel tool calls in Langfuse hook

[junwai7159]

Problem

When a turn fires several tools in parallel (e.g. 3 WebFetch calls at once or multiple subagents launched together via Agent), only the last tool span shows up in Langfuse. The rest silently vanish.

Root cause

Claude Code writes one logical assistant message (single message.id) as several transcript rows: the thinking/text rows first, then one row per parallel tool_use block, and these rows can be interleaved with tool_result rows. build_turns() de-dupes assistant rows by message.id with a "latest row wins" assignment:

mid = get_message_id(msg) or f"noid:{len(assistant_order)}"
if mid not in assistant_latest:
    assistant_order.append(mid)
assistant_latest[mid] = msg  # overwrites; earlier tool_use rows are lost

So every parallel tool call but the last gets overwritten before emit.

Raw transcript evidence

A real turn that fired 4 parallel Read calls. These are 4 separate JSONL rows (distinct uuid + timestamp), but all share one message.id and each carries a single tool_use block (fields trimmed to the relevant ones):

// row 1
{
  "message": {
    "id": "msg_bdrk_…je4ivmq",
    "role": "assistant",
    "content": [
      {
        "type": "tool_use",
        "id": "toolu_…HV2",
        "name": "Read",
        "input": {
          "file_path": ".../subagent_one.md"
        }
      }
    ]
  },
  "uuid": "334452ca-…",
  "timestamp": "2026-06-19T08:42:41.885Z"
}

// row 2
{
  "message": {
    "id": "msg_bdrk_…je4ivmq",
    "role": "assistant",
    "content": [
      {
        "type": "tool_use",
        "id": "toolu_…8q4",
        "name": "Read",
        "input": {
          "file_path": ".../subagent_two.md"
        }
      }
    ]
  },
  "uuid": "d3cfa674-…",
  "timestamp": "2026-06-19T08:42:42.885Z"
}

// row 3
{
  "message": {
    "id": "msg_bdrk_…je4ivmq",
    "role": "assistant",
    "content": [
      {
        "type": "tool_use",
        "id": "toolu_…y4J",
        "name": "Read",
        "input": {
          "file_path": ".../subagent_three.md"
        }
      }
    ]
  },
  "uuid": "bd31bb32-…",
  "timestamp": "2026-06-19T08:42:42.921Z"
}

// row 4
{
  "message": {
    "id": "msg_bdrk_…je4ivmq",
    "role": "assistant",
    "content": [
      {
        "type": "tool_use",
        "id": "toolu_…PFY",
        "name": "Read",
        "input": {
          "file_path": ".../subagent_four.md"
        }
      }
    ]
  },
  "uuid": "7e37ddc2-…",
  "timestamp": "2026-06-19T08:42:43.139Z"
}

All four rows hash to the same key in assistant_latest, so the dict assignment keeps only row 4 (subagent_four.md) — rows 1–3 are silently overwritten and never emitted to Langfuse.

Note the rows aren't contiguous in the file either: tool_result rows for the earlier calls interleave between them, so a positional fix wouldn't be safe — merging by message.id is.

Fix

Merge the content blocks across rows that share a message.id instead of overwriting, so all tool_use blocks survive.

Metadata (model, usage, timestamp) is identical across the split rows, so the first row's values are kept.

Notes

• No behavior change for single-tool turns.
• Only affects turns traced after the fix; turns already recorded in ~/.claude/state/langfuse_state.json are not rewritten.

[CLAassistant]

All committers have signed the CLA.