Show live token usage on backgrounded sub-agent placeholder
Problem
When a sub-agent is launched with run_in_background: true, the parent thread renders a placeholder row like:
Frontend Developer(Ship W29 S-BPB pipeline board)
⎿ Backgrounded agent (↓ to manage · ctrl+o to expand)
There is no visibility into the agent's token consumption while it runs. Token totals only appear in the final completion payload (<usage><total_tokens>...</total_tokens></usage>), which surfaces after the agent finishes — sometimes 10+ minutes later.
For workflows that fan out 3+ parallel worktree-isolated agents (compound-engineering pattern), this matters because:
- A runaway agent (stuck in retry loop, broken tool call) cannot be detected until it completes — by which time it has burned significant tokens
- No way to budget across parallel agents in real time
- The orchestrator has to guess whether agent is making progress or hung
Proposed UX
Render the live cumulative token count on the backgrounded-agent row:
Frontend Developer(Ship W29 S-BPB pipeline board)
⎿ Backgrounded agent (↓ to manage · ctrl+o to expand · 47.2k tok)
Updated each time the sub-agent emits a turn. Could optionally show tool-use count too (47.2k tok · 23 tools).
Why it matters
Per-agent JSONL transcripts (<temp>/tasks/<agentId>.output) already contain message.usage blocks per turn — the data exists, just isn't bubbled up to the parent UI. Running tail -f <file> | jq '.message.usage' from a separate terminal works as a workaround, but requires the user to leave Claude Code and chase paths manually.
Workaround today
- Read the final completion payload
<usage>block - Tail the JSONL transcript externally (parent thread is forbidden from reading it — would overflow context)
- Approximate via duration (highly imprecise — varies 50–500 tok/s based on tool mix)
None of these give the user a live view inside the parent thread.
Environment
- Claude Code on Windows 11 Pro (10.0.26200), bash + PowerShell
- Sub-agents launched with
Agenttool,run_in_background: true,isolation: "worktree" - Worktree-isolated parallel-agent pattern (3 simultaneous agents typical)
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗