Feature request: a supported way to disable the agent_summary spinner (22.7% of requests, ~18% of weighted tokens in a subagent-heavy session)

Request

Provide a supported way to disable the agent_summary timer (the 3-5 word subagent spinner text) — an env var, a settings.json key, or coverage under an existing flag.

Version: 2.1.233 · Platform: darwin 25.6.0 · Entrypoint: cli

Why

In a subagent-heavy session it is a large, invisible share of consumption. Measured over one 4-hour session (27 subagents, 1,496 requests mapped from raw payloads captured with OTEL_LOG_RAW_API_BODIES):

| kind | calls | cache read | cache write | output | raw tokens | raw % | weighted % |
|---|---:|---:|---:|---:|---:|---:|---:|
| real agent turns | 1,076 | 193,219,620 | 5,209,321 | 825,201 | 199,321,158 | 70.4% | 72.7% |
| agent_summary | 339 | 61,560,341 | 254,270 | 21,966 | 62,757,361 | 22.2% | 18.0% |
| suggestion | 37 | 10,117,116 | 9,370 | 229 | 10,145,363 | 3.6% | 2.5% |
| task-notification | 36 | 9,583,910 | 205,636 | 45,758 | 9,835,376 | 3.5% | 3.8% |

Weighted at Opus 5 list rates (reads 0.1×, 5m writes 1.25×, 1h writes 2×, output 5×).

22.7% of all API requests in the session, producing 21,966 output tokens total (~65 per call) of cosmetic status text.

Each call re-sends the subagent's entire conversation history plus the full tool schema, on claude-opus-5, every 30 seconds per active subagent. Cost therefore scales with fan-out and wall-clock duration, not with useful work: a session with three concurrent long-running subagents pays six of these per minute.

Why gating it should be safe

The call is already architecturally decoupled from the agent it describes — from the 2.1.233 bundle:

await eW({
  promptMessages:[Sn({content:mXS(p)})],
  querySource:"agent_summary", forkLabel:"agent_summary",
  maxTurns:1, skipTranscript:true, skipCacheWrite:true,
  canUseTool: async () => ({behavior:"deny", message:"No tools needed for summary"}),
})

It forks context (forkContextMessages) rather than mutating it, denies all tools, caps at one turn, writes no cache, and is excluded from the transcript. The summary never re-enters the agent's context. Suppressing it cannot change any real turn — with it disabled, every other request in the session is byte-identical.

There is already a conditional at the call site (let Z = l ? (…) => { KEf(…) } : void 0), so a gate exists structurally; it just isn't reachable by configuration.

Suggested options, in preference order

1. A dedicated flag — CLAUDE_CODE_DISABLE_AGENT_SUMMARY=1, or a settings.json key. I checked the binary's CLAUDE_CODE_* string table for summ|spin|agent: no matches, so nothing exists today.
2. Cover it under CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC. This call fits that category well — purely cosmetic, no effect on agent behavior. If it is already covered, documenting that would be enough and I'll close this.
3. Make the 30s interval configurable. var fXS = 30000 is read via i.intervalMs ?? fXS, and no caller passes intervalMs — plumbing that through to config would let users trade update latency against tokens.

Secondary observation

These calls carry skipTranscript: true, so they appear in no transcript. Any transcript-based cost tooling — including users' own — silently understates real token consumption by this amount. Even were the calls kept, emitting them to the transcript (or to OTel usage metrics) would make the spend visible.

Related: #86876 (subagent resume drops injected role:"system" messages), from the same investigation.

"env": {
    "CLAUDE_CODE_ENABLE_NARRATION": "0",
    "CLAUDE_CODE_FORK_SUBAGENT": "0",
	"DISABLE_AUTOUPDATER": "1"
}