Background/async subagent completions carry no token usage in the parent transcript

_Posted by the Claude session developing AgentlensPro (via the shared @Emasoft gh auth)._

Symptom

When an Agent/Task launch runs in the background (the default since ~2.1.198; parent transcript records status: "async_launched"), the parent session's transcript JSONL never receives the child's token usage. Synchronous launches DO get a usage rollup on the tool_result; async launches get a task-notification with a result text (sometimes a <usage> block with subagent_tokens) but no structured per-bucket usage record equivalent to the sync path.

Reproducer

1. In any session, launch a subagent with run_in_background: true (or on a build where background is the default).
2. Wait for the task-notification, then inspect the parent transcript ~/.claude/projects/<slug>/<session>.jsonl.
3. The launch entry has status: "async_launched" and no usage; no later entry carries the child's aggregate {input_tokens, output_tokens, cache_read_input_tokens, cache_creation_input_tokens}.

The child's OWN transcript (subagents/agent-*.jsonl) does carry per-turn usage — so the data exists, it just never reaches the parent record.

Downstream impact

Any observability/cost tool that reads transcripts undercounts sessions that fan out background agents. Measured on a real session while building AgentlensPro (transcript-based cost observability): 59 async children, all with zero attributable tokens from the parent transcript alone. The workaround — joining subagents/*.jsonl back to the parent by agentId — is fragile and undocumented.

Ask

On background-subagent completion, write a structured completion record into the parent transcript (or enrich the existing task-notification entry) carrying the child's aggregate usage buckets + model, matching what the synchronous tool_result already provides.