[olzhik11]

Problem

I tested the plugin end to end against a local Laminar, with headless claude -p sessions and interactive sessions driven through expect. Every turn that used a tool lost its final answer:

• the root span's output was "";
• the last LLM Call N span was missing;
• the late transcript row was then dropped for good, because SessionEnd logged "Processed 0 turns".

The cause is that Claude Code fires Stop before it writes the turn's final assistant row. A probe hook measured this:

Mode	Final row in the transcript when Stop fires?	When it lands
claude -p (4/4 runs)	no	61–82 ms later
interactive (2/3 runs)	no, and the file doesn't grow at all while the hook runs	only after the hook exits

The hook saw an assistant row, assumed the turn was complete, and emitted it. The late final row then had no user prompt in front of it, so it was ignored.

Replaying the finished transcript through the same bundle produced correct traces, so the parser was fine and timing was the only problem. Because interactive sessions write the row only after the hook exits, waiting inside the hook can't fix this; I tried a 2 s wait and it timed out.

Fix

splitTrailingIncompleteTurn now also holds the trailing turn (in pendingTurnRows, as it already does for a prompt with no reply) when either:

• its last chat row is a tool_result: the reply to it hasn't been written yet; or
• none of its assistant rows carries the end of the Stop payload's last_assistant_message. That field is the turn's final text, so this also covers a final message whose thinking row landed but whose text row didn't.

The held turn is emitted whole on the next hook run: the next Stop, or SessionEnd, which never holds. When the row is late, the trace arrives one hook later; in -p, SessionEnd runs right after Stop, so the delay is about a second. Nothing is lost, and the hook adds no wait.

Also fixed: empty output on background subagents

A background subagent ends with a SubagentHandback tool call instead of a final text message, so the Subagent: … span's output was "". Its output now falls back to the call's message.

Testing

• npm run typecheck, and npm test with 61/61 passing. That includes 3 new tests: a turn ending in a tool_result, a turn whose payload text isn't written yet, and the SubagentHandback report.
• End to end against a local Laminar, checked in ClickHouse. Headless and interactive sessions with a plain turn, a turn with Read and Bash calls, and a turn that runs a subagent (both sync and background). Every turn now has its final answer, its last LLM call, and its subagent output.

Session	Before	After
headless, Read + Bash	4 spans, root output "", no LLM Call 2	5 spans, root output "notes.txt has 2 lines…", LLM Call 2
interactive, Read + Bash	4 spans, root output "", no LLM Call 2	5 spans, root output "notes.txt has 2 lines…", LLM Call 2
interactive, background subagent	Subagent: output ""	Subagent: output "14…"

CLAUDE.md now documents the hold rules. dist/hook.cjs is rebuilt. The version bump is in #9.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes transcript→turn gating and export timing for every Stop hook; wrong hold logic could delay or duplicate traces, but SessionEnd flush and existing offset/export retry limits blast radius.

Overview
Fixes incomplete Laminar traces when Claude Code’s Stop hook runs before the transcript has the turn’s final assistant row (common after tool use).

Trailing-turn hold (pendingTurnRows): splitTrailingIncompleteTurn no longer treats “any assistant row” as complete. It also holds when the tail ends on a tool_result, or when assistant rows don’t yet include the suffix of the Stop payload’s last_assistant_message (wired from hook.ts as finalAssistantText). Held rows replay on the next hook run; SessionEnd still flushes everything.

Subagent spans: getSubagentReport sets subagent output from normal assistant text or from a SubagentHandback tool’s message when background subagents don’t end with text.

Docs (CLAUDE.md), tests (tool-result race, payload text race, handback), and dist/hook.cjs are updated.

^{Reviewed by Cursor Bugbot for commit 561dc75. Bugbot is set up for automated code reviews on this repo. Configure here.}