[kimlj]

Summary

• Claude Code writes one JSONL line per content block, so a message that calls several tools in parallel spans several lines with the same message.id. dedupeStreamingMessageIds keeps only the last line (the MCP tool calls and tool_use blocks missing from dashboard due to streaming write deduplication #110 fix), which has the right usage but only the last tool_use. Every earlier parallel tool call was dropped from the tool, MCP and shell breakdowns and from turn classification.
• The last line is still kept. The tool_use blocks of the earlier lines are put back in front of its content, matched by block id, so the older shape from MCP tool calls and tool_use blocks missing from dashboard due to streaming write deduplication #110 (every block restated in the final write) is not counted twice. A message whose last line already holds every block is returned unchanged.
• Calls, tokens and cost do not change. The claude parse tag gets streamed-tool-blocks-v1 so cached sessions re-parse. DAILY_CACHE_VERSION is not bumped: the v77 note says the next bump has to raise the Node floor first, so days the daily cache has already settled keep their current figures until then.

Found while building the session comparison proposed in #1790: the tool counts for two of my sessions did not match the transcripts.

Testing / Verification

• I have tested this locally against real data (not just unit tests)
• npm test passes: no new failures, but not clean on Windows (see Tests below)
• npm run build succeeds
• Not a UI change
• UI change: before and after screenshots attached below

Real data, one day of my Claude Code history (2026-09-11), codeburn report --format json, main vs this branch:

	main	this branch	Counted from the raw JSONL
API calls	482	482	482 assistant message ids
Cost	$170.77	$170.77
Tool calls (tools[])	417	1,589	1,589 unique non-MCP tool_use ids
Shell commands (shellCommands[])	367	945
Exploration turns	24	20
Coding turns	34	38
Delegation turns	0	2

On main, 1,172 of 1,589 tool calls that day (74%) were missing. The ground-truth column is a script that reads every tool_use block in that day's transcripts and counts unique block ids. One session alone, 49 tool calls in its transcript, showed 18.

The classification changes are the knock-on effect: a turn whose Agent or Edit call sat on an earlier line was classified without it, so delegation turns showed as zero.

Tests:

• tests/parser-streamed-tool-blocks.test.ts (new, 5 cases): parallel tool calls on separate lines are all kept; an MCP call in parallel with a built-in tool is kept; usage still counts once; the MCP tool calls and tool_use blocks missing from dashboard due to streaming write deduplication #110 restated shape does not double count; an already-complete last line is returned as-is. The two tool cases fail on main with expected [ [ 'Bash' ] ] to deeply equal [ [ 'Read', 'Read', 'Bash' ] ].
• Full suite (npm test), this branch vs stock main, Windows 11, Node 24: no new failures. main has 43 failures on this machine before any change (symlink, keychain and timing tests); the branch fails the same set. Six files that differed between full runs (sync-attribution, cli-plan, dashboard-process-exit and three more) give the same result on both branches when rerun on their own, so those are load flakes on Windows.

Environment: Windows 11, Node 24.18, Claude Code transcripts.