Subagent spend is not invisible — you are reading the wrong meter

A turn delegated four subagents. They came back with tidy summaries, /context showed your window barely moved, and twenty minutes later you hit a rate limit you did not see coming. The folk explanation for this is that subagent tokens are invisible to Claude Code's own instruments. That explanation is close enough to be repeated a lot and wrong in the part that matters: the numbers exist, they are just not on the screen you looked at.

What each meter is actually scoped to

The confusion has a clean origin. The docs say, in one sentence, "It covers the main conversation only, not subagents." That sentence is about the Prompt cache (main) line — requests, hit ratio, misses, warm or cold. It has been generalized into a claim about all cost telemetry, which the same page contradicts two sections later when it lists subagents as an attribution category in /usage.

What to do instead

Keep the habit, drop the reasoning behind it. Delegation genuinely can cost several times what the same work costs inline — each subagent pays for its own system prompt, its own CLAUDE.md load, and its own MCP tool definitions, and none of that rides your warm cache. The fix is to look at the meter that is scoped to the thing you are worried about:

• Before and after a fan-out turn, check /usage, not /context. Press d or w to switch between 24 hours and 7 days. The Attribution list will name the subagents.
• During the turn, the agent panel under your prompt already shows a token count per running subagent. Watching four rows climb in parallel is the most direct signal there is.
• If you want it in your status line, subagentStatusLine receives every visible row as JSON with tokenCount, contextWindowSize and tokenSamples per task, so you can render a running total yourself. contextWindowSize — the field that turns a raw count into a percentage — arrived late: it and the per-task model field "require Claude Code v2.1.205 or later and are omitted for a task whose model isn't resolved yet."
• If you need real accounting, export OpenTelemetry and group by query_source and agent_id. That is the only path that gives you per-subagent token and cost metrics you can trust rather than estimate.

What is still unverified

Nobody has run a controlled measurement for this entry. Specifically:

• The cost total. No doc sentence says the session Total cost figure includes subagent API calls. It says it accumulates all API calls in the session, and a subagent's requests are made in that session — but that is an inference from wording, not a stated guarantee. This is the single load-bearing gap. If the total in fact excludes subagents, the original "invisible" claim is right about /cost and wrong only about /usage.
• How fast attribution updates. /usage is described as showing recent usage; whether a subagent that finished sixty seconds ago is already in the percentages is not documented.

What would settle it: two headless runs of comparable work, one that delegates and one that does not, comparing the reported total_cost_usd against the token counts the subagents themselves reported. If the delegating run's total is roughly the inline run's total plus the subagents' own consumption, the cost line is inclusive and the strong claim is dead.

# Same task, one delegating, one not. Compare the totals.
claude -p "Use a subagent to summarize src/, then report" --output-format json \
  | jq '{total_cost_usd, usage}'
claude -p "Summarize src/ yourself, then report" --output-format json \
  | jq '{total_cost_usd, usage}'

Run it twice in each direction — cache warmth alone moves these numbers enough to fool a single comparison.