Isolating Claude Code Subagent Events from Parent-Step Tracking
A Vercel AI SDK contribution that prevents Claude Code subagent messages from contaminating parent-step usage and tool-event tracking. Merged upstream in PR #18740 with explicit credit to @patrickswedish.
PR #18740
merged upstream
PR #18631
credited source PR
69
harness tests
APPROVED
Felix Arntz
01 · Engineering problem
The failure mode
Claude Code Task-tool subagents emit messages with a non-null parent_tool_use_id. The harness processed those messages as belonging to parent agent steps, corrupting token accounting.
02 · Technical response
What changed
- 01Filtered subagent-originated messages before parent-step normalization.
- 02Preserved parent-agent token usage and tool handling without breaking public APIs.
03 · Verified outcome
Evidence-backed result
- ✓Vercel engineer Felix Arntz approved and shipped the fix upstream in PR #18740 with full credit.
- ✓Prevented nested token leakage across agent workflows.
04 · Public record
Source evidence
Independent open-source engineering by Patrick Ribbsaeter. Company and project names identify the public repository and maintainer context only. No employment, client, vendor, or partnership relationship is implied unless explicitly stated.