TL;DR
Every ~30 seconds, Claude Code re-sends each background agent's full transcript to the model to
fill that agent's status line in the agents panel. no-agent-summary.js is a Bun --preload
that stops that timer without modifying the Claude Code binary.
curl -fsSL https://gist.githubusercontent.com/flamerix-ai/c3905705cc4205e5de3011fa2e74efce/raw/no-agent-summary.js -o ~/.local/share/no-agent-summary.js
claude daemon stop --any
BUN_OPTIONS="${BUN_OPTIONS:+$BUN_OPTIONS }--preload=$HOME/.local/share/no-agent-summary.js" claude
claude daemon stop --any ends running background sessions. To make it permanent for every Claude
Code process, see Install; on Windows, Windows.
Contents
The problem
Every ~30 seconds, for every background agent, Claude Code sends that agent's entire
transcript to the agent's own model (usually Opus) and asks it to "describe your most recent
action in 3-5 words". The answer updates the agent's status line in the agents panel. It does not
reach the model on ordinary turns; after a compaction, up to 256 characters of it return to the
model as a Progress: line for a still-running agent (reported by
wtfsayo from local transcripts; we confirmed the route in the 2.1.301 binary, not in a
live capture). There is no setting that turns it off, and since 2.1.232 (August 2026) it has been
on by default.
The anomaly was detected by DRIM in Claude Code's OpenTelemetry export:
grouping claude_code.api_request events by query_source showed agent_summary costing more than the main conversation itself.
query_source = agent_summary |
week before the hook |
2 days after |
| share of all Claude Code spend |
16.4 % (main conversation: 15.7 %) |
0 % |
| share of all API requests |
19.7 %, about 1 in 5 |
0 % |
| part of that spend on Opus |
96 % |
— |
| input tokens per output token |
≈ 9,000 : 1 |
— |
Spend here means Claude Code's own per-request estimate (cost_usd), not an invoice: on a
subscription, what you lose is rate-limit headroom. We are not the only ones who measured it:
#86879 collects four independent
measurements of 18–26 %.
The fix (how it works)
no-agent-summary.js is a Bun preload (~70 lines of code). It wraps only the global
setTimeout, the one the summary timer is armed through. When a callback's source text contains
a summarizer anchor (e.g. the [AgentSummary] Timer fired log line), the wrapper returns an inert
handle instead of arming the timer. Every other timer passes through untouched.
Why a timer hook rather than something simpler: the module ships as precompiled Bun bytecode, so
editing its JS text inside the binary changes nothing, and the API client does not go through a
wrapped globalThis.fetch. But Function.prototype.toString still returns a callback's source
even when it runs from bytecode, so the timer can be recognised by a string inside it. The inert
timer uses exactly 2³¹−1 ms: Node and Bun treat a larger delay as 1 ms, so an "infinite" delay
would fire at once.
Which processes it touches
The hook installs only inside Claude Code: Bun.main must be the compiled entry
(/$bunfs/root/cli, or B:/~BUN/root/cli on Windows) and process.execPath must be a Claude Code
install. Any other bun or compiled Bun binary with another entry name loads the file and does nothing.
install channel (Linux rows measured by us with claude --version) |
process.execPath |
| nix |
…/bin/.claude-wrapped |
npm @anthropic-ai/claude-code |
…/@anthropic-ai/claude-code/bin/claude.exe (.exe on Linux too) |
| npm platform package |
…/@anthropic-ai/claude-code-linux-x64/claude |
| native installer |
~/.local/share/claude/versions/<version> |
| Windows, native and desktop-app bundled (reported by anhdoox, omgapnt) |
C:\Users\<user>\.local\bin\claude.exe, %APPDATA%\Claude\claude-code\<ver>\<hash>\claude.exe |
On any platform, Verify step 1 tells you whether the hook installed. One known pass-through: a foreign binary built with
bun build --compile --outfile cli and renamed claude; there, only timers whose callback
carries an anchor change.
Install
-
Set BUN_OPTIONS globally so every Claude Code process gets it. Write
--preload=<path>, with =: BUN_OPTIONS reaches every bun on the machine, and with a
space (--preload <path>) bun subcommands break (bun install → error: Script not found "install"). Claude Code reads both forms. One of the following (on Windows, see Windows):
# ~/.profile, ~/.bashrc, ~/.zshrc
export BUN_OPTIONS="${BUN_OPTIONS:+$BUN_OPTIONS }--preload=/abs/path/no-agent-summary.js"
# ~/.config/fish/config.fish
set -gx BUN_OPTIONS (string trim -- "$BUN_OPTIONS --preload=/abs/path/no-agent-summary.js")
# ~/.config/environment.d/50-no-agent-summary.conf (systemd user env; also reaches
# user services and timers; applies when the user manager next starts, usually next login)
BUN_OPTIONS=--preload=/abs/path/no-agent-summary.js
Or a wrapper named claude earlier on PATH (the case keeps it idempotent when one Claude
Code process starts another, and also recognises the v2 space form):
#!/bin/sh
HOOK=/abs/path/no-agent-summary.js
case " ${BUN_OPTIONS-} " in
*" --preload=$HOOK "*|*" --preload $HOOK "*) ;;
*) export BUN_OPTIONS="${BUN_OPTIONS:+$BUN_OPTIONS }--preload=$HOOK" ;;
esac
: "${CC_AGENT_SUMMARY_LOG:=${XDG_RUNTIME_DIR:-/tmp}/no-agent-summary.log}"
export CC_AGENT_SUMMARY_LOG
exec /path/to/the/real/claude "$@"
-
Stop the daemon with the command below. Restart every running session from an environment
that has BUN_OPTIONS.
claude daemon stop --any
Windows
Save no-agent-summary.js as .local\share\no-agent-summary.js in your user folder. In PowerShell:
$hook = "$env:USERPROFILE/.local/share/no-agent-summary.js".Replace('\', '/')
if ($hook -match '\s' -or -not (Test-Path -LiteralPath $hook)) {
Write-Error 'hook file missing, or its path has a space'
} else {
$old = [Environment]::GetEnvironmentVariable('BUN_OPTIONS', 'User')
if (-not " $old ".Contains(" --preload=$hook ")) {
[Environment]::SetEnvironmentVariable('BUN_OPTIONS', "$old --preload=$hook".Trim(), 'User')
}
}
A User variable, not $PROFILE, reaches the desktop app and every process not started from
PowerShell. Keep the forward slashes. The guard stops on a missing file or a space in the path (bun
splits BUN_OPTIONS on whitespace): point $hook at an 8.3 short name or a folder without spaces
that only your account can write to. Then fully quit the desktop app. Run claude daemon stop --any. Open new terminals.
Desktop app
The Code tab runs its own bundled Claude Code, not the one on your PATH, so the Code tab needs a global
BUN_OPTIONS. For the bundled binary on Windows, see Tested versions. Cowork is probably not
covered: by omgapnt's reasoning, Cowork runs in its own VM.
Env
| variable |
values |
effect |
CC_AGENT_SUMMARY_MODE |
off (default), pass |
off suppresses the summary timer. pass is the A/B control: loads and logs, arms every timer untouched. Any other value warns on stderr and acts as off. |
CC_AGENT_SUMMARY_LOG |
file path, optional |
Appends preload active mode=… bun=… hooked=global.setTimeout on load, one line per matched timer (first 20), and an exit count. |
CC_AGENT_SUMMARY_MARKER |
file path, optional |
Overwritten on load with the preload active … line. |
Uninstall
Order matters: a BUN_OPTIONS that names a missing file, or a preload that throws on load, stops Claude Code from starting
(error: preload not found). So:
- Remove the
--preload= entry from wherever you set it (Windows: the User variable), or the wrapper.
- Quit every running session and the desktop app. Run
claude daemon stop --any.
- In a new terminal, confirm
BUN_OPTIONS no longer names the file. Delete the file last.
Nuances
- Background processes: the daemon, pty host and pre-warmed
bg-spare sessions inherit the
environment of the client that started the daemon, so a daemon started before BUN_OPTIONS
was set keeps spawning unhooked sessions. On Linux, /proc/<pid>/environ of each must carry
--preload; if it does not, restart the daemon (Install step 3).
- Headless runs need it too:
claude -p, cron, CI, editor integrations.
- npm
--ignore-scripts: bin/claude stays a shell stub and Bun never starts, so the hook
never runs.
- Upgrading from v2: overwrite the file in place. Switch
BUN_OPTIONS to the = form.
Restart the daemon. If both forms name the same path, bun loads it once. Two different paths
(say an old copy kept beside the new one) both load: a second v3 or v4 copy is a no-op, but v2
has no such guard, so do not leave a v2 path in BUN_OPTIONS.
Verify (after every update)
-
Preload loads:
BUN_OPTIONS="--preload=$HOME/.local/share/no-agent-summary.js" CC_AGENT_SUMMARY_LOG=/tmp/nas.log claude --version && cat /tmp/nas.log
Expect preload active mode=off …. No line: BUN_OPTIONS did not reach the process, the path
is not absolute, or the gate does not match your install.
On Windows, in a new terminal: $env:CC_AGENT_SUMMARY_LOG = "$env:TEMP\nas.log"; claude --version; Get-Content "$env:TEMP\nas.log".
Fail-open
Matching is by source text: an anchor must sit inside the callback passed to the timer. If an
update breaks that, the hook matches nothing and summaries silently come back. Ways it can happen:
- the log line or query-source id is renamed;
- the callback is passed via
fn.bind(...) (no source in toString) or another wrapper;
- the strings move to a shared constant outside the callback;
- the timer moves off the global
setTimeout (setInterval, require("timers"),
timers/promises, a scheduler loop).
Effect
- We have run the hook under every Claude Code process on our machine since 2026-10-26 (then
2.1.301). Our telemetry since then shows zero
agent_summary requests outside our no-hook and
pass control runs, and no other side query grew to take its place.
- Subagents, background agents, forked skills and background sessions run normally, without the
summary fork.
- Visible change: the per-agent status line stops updating; SDK
task_progress has no summary.
- Model-side change: after a compaction the model no longer gets the
Progress: line for a
still-running agent. We have not measured whether that changes its behaviour.
- Overhead: one
Function.prototype.toString per distinct timer callback, cached in a WeakMap.
claude --version: ~10 ms → ~17 ms (one measurement).
Tested versions
Linux x64 binary from npm @anthropic-ai/claude-code-linux-x64 (its sha256 equals the native
installer's manifest.json checksum). Static: in the extracted bundle, the callback of the
summary timer carries a hook anchor. Live: claude -p --model haiku with
CLAUDE_CODE_AGENT_PROGRESS_SUMMARIES=1 (headless runs don't arm the summary timer without it),
one subagent doing 8 steps of ~20 s, counting [AgentSummary] Forking in --debug-file.
Hook version per row: v4 (current, run 2026-11-03), v3 (re-run 2026-10-29), unmarked v2.
v4 changes only the gate.
| version |
published |
sha256 |
static |
summary forks in a ~2.5-min agent run: no hook / hook pass / hook off |
notes |
| 2.1.317 v4 |
2026-11-02 |
eb224419e344 |
OK |
5 / 5 / 0 |
latest on 2026-11-03 |
| 2.1.312 v3 |
2026-10-27 |
4f4178ce41d8 |
OK |
5 / 6 / 0 |
the pass run took 232 s on a loaded machine, so one more 30 s tick fit; off with the gist file itself: 0 |
| 2.1.301 v3 |
2026-10-15 |
bc1aad9982c7 |
OK |
5 / 5 / 0 |
run through the Nix-packaged binary rather than the npm one |
| 2.1.312 |
2026-10-27 |
4f4178ce41d8 |
OK |
5 / 5 / 0 |
latest on 2026-10-28; interactive: 6 no hook, 0 off |
| 2.1.310 |
2026-10-23 |
b73676430af1 |
OK |
5 / 5 / 0 |
|
| 2.1.308 |
2026-10-22 |
356733f185fa |
OK |
5 / 6 / 0 |
stable on 2026-10-28; the pass run's agent took 9 steps instead of 8 |
| 2.1.301 |
2026-10-15 |
bc1aad9982c7 |
OK |
5 / 5 / 0 |
same with earlier hook builds; interactive off also 0 |
| 2.1.280 |
2026-09-22 |
1e08503dbdf3 |
OK |
1 / 1 / 0 (timer fired 11 / 11 / 0) |
interactive: -p cannot enable summaries on this version. The agent sat at a permission prompt, so later timers skipped the unchanged transcript |
| 2.1.257 |
2026-09-01 |
9a64bda9d872 |
OK |
1 / 1 / 0 (timer fired 11 / 11 / 0) |
same as 2.1.280 |
v4 on Windows: Windows 11 Enterprise 24H2 (build 26100) in a KVM virtual machine, native
installer, bun 1.4.4, PowerShell 5.1, same live test as above, 2026-11-03: 2.1.316 5 / 5 / 0;
2.1.317 5 / 5 / 0 (a 6th no-hook fork was aborted empty because the agent had finished).
5 / 5 / 0 = the fork happens on this version, merely loading the hook changes nothing, and off
removes it.
Reported by others
| platform |
version |
checked |
by |
| macOS 26.7.1, native darwin-arm64, bun 1.4.4 |
2.1.312 |
static: the 3–5 words prompt, the timer (30 s, clamped 30–240 s), no off switch; live: the preload loads, and an interactive session with a background subagent logged summary timer #1 suppressed (setTimeout, delay=30000). Token numbers not checked (no OTEL export) |
lostsock, HN, 2026-10-29 |
| macOS, native darwin-arm64, bun 1.4.4 |
2.1.312 |
v3. Headless -p --model haiku with CLAUDE_CODE_AGENT_PROGRESS_SUMMARIES=1: 5 / 5 / 0. Interactive, background agent, nothing opted in: 5 pass / 0 off |
declanbx, 2026-10-30 |
| Windows |
2.1.312 |
5 / 5 / 0. A derivative with its own gate fix, not this hook |
omgapnt, 2026-10-29 |
| Windows, desktop app's bundled binary |
2.1.309 |
5 / 5 / 0. Same test, same derivative; headless, not through the app's UI |
omgapnt, 2026-10-29 |
| Windows 11, bun 1.4.4 |
2.1.314 |
Headless, one subagent, three sleep 20 calls, CLAUDE_CODE_AGENT_PROGRESS_SUMMARIES=1. pass: matched=3 suppressed=0. off: matched=1 suppressed=1. Used anhdoox's own gate, not v4's |
anhdoox, 2026-10-30 |
| not stated |
2.1.308 |
5 / 5 / 0 |
Bedz01, 2026-10-29 |
| Windows 11, embedded bun 1.4.3 |
2.1.293 (npm), 2.1.301 (VS Code extension bundle) |
Load only: with their own gate fix the hook reports preload active mode=off. Suppression not run (not signed in) |
apoapostolov, 2026-10-30 |
Why there is no setting for this
- The documented switch only turns it on. The SDK option
agentProgressSummaries defaults to
false, but in the bundle it is one term of an OR, and every code path that writes it sets
true. So false changes nothing, and the undocumented CLAUDE_CODE_AGENT_PROGRESS_SUMMARIES
can only enable summaries too.
- The interval belongs to the server. It comes from a per-model-family server flag and is
clamped to 30–240 s. The default is already the minimum, and no value disables the timer.
- The obvious env vars miss.
DISABLE_NON_ESSENTIAL_MODEL_CALLS, which some packagers set,
does not occur in the binary at all. CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1
stops the summary, but only by taking background agents with it. CLAUDE_CODE_FORK_SUBAGENT=0 costs you
fork subagents and can help only in interactive sessions: forked skills are always summarized,
and reports on 2.1.312 disagree on whether it works at all
(#86879,
#105981).
- The prompt cache makes it cheaper, not cheap. About 98 % of these calls re-read the
transcript from cache; the ~2 % that hit a cold cache carry over a third of the summary's cost.
- Cost tools that read transcripts don't see it. These requests are not written to the
session logs on disk (#104358); only
the OpenTelemetry export shows them. The docs section on background token usage doesn't list it
either, while Agent View's row summaries are documented as a Haiku-class model refreshing every
few minutes.
- A cheaper
short mode sits behind a server flag. In the 2.1.317 binary, the flag
tengu_groovy_ember ({"mode":"short"}) switches the summary to short notes of about 1.5k
tokens written by Haiku, but only when a first-party Haiku model is available; otherwise the
summary stays full. claude -p runs that opt in via CLAUDE_CODE_AGENT_PROGRESS_SUMMARIES do
not consult the flag. On our account the flag is unset, so our live runs used full mode. The
hook covers both modes: both run from the same timer callback, which carries the anchor.
Timeline (2026)
| version (published) |
what changed |
| 2.1.32 (Feb 5) |
the code first appears; dead in public builds |
| 2.1.72 (Mar 9) |
reachable as an SDK opt-in |
| 2.1.117 (Apr 21) |
reachable in the interactive CLI behind the fork gate or a server rollout |
| ≤ 2.1.220 (Jul 24) |
forked skills are always summarized |
| 2.1.232 (Aug 13) |
on by default for interactive background agents |
| 2.1.300 (Oct 14) |
interval becomes server-tunable, 30–240 s |
| 2.1.310 (Oct 23) |
the cheaper short mode lands, switched off |
In our telemetry its weekly share of spend went from 3.5–7 % in July (the fork gate was already
open for us) to 7.3 % and then 11.8 % in the week we picked up 2.1.233, and to about 16 % by
mid-October.
FAQ
Does this apply to the Agent SDK or ACP? Mostly no. In the 2.1.301 bundle (2.1.310 has the same
call sites) a background Agent-tool agent is summarized only in interactive mode, unless the host
opts in with agentProgressSummaries: true. Fork subagents and forked skills always are, SDK
included. The preload works there too, provided BUN_OPTIONS reaches the claude process.
Not checked: whether a given ACP adapter sets it, and whether versions after 2.1.301 changed the
non-interactive default.
Doesn't "awaySummaryEnabled": false turn this off? No. That setting controls the welcome-back
recap (gist comment).
See also (not audited by us): claude-summary-tuner, v1.0.2 or later only (earlier versions wiped other apps' startup entries when the logger was on); declanbx's fork.
Changelog
- v4 — Windows: the gate also accepts
Bun.main = B:/~BUN/root/cli, normalises backslashes
in Bun.main and process.execPath, and lowercases the basename (v3 never matched there).
README: the compaction route in The problem; new Windows, Desktop app,
Uninstall, FAQ and Contributions sections; our Windows runs and community results in
Tested versions.
- docs — first macOS report (darwin-arm64, 2.1.312) added to
Tested versions.
- v3 —
--preload=<path> everywhere (the space form broke bun subcommands); the hook installs
only inside Claude Code (Bun.main plus process.execPath, see
Which processes it touches); a second load in the same process is
a no-op.
- v3 — wraps only the global
setTimeout, the one the summary timer is armed through (v2
also wrapped setInterval and require("timers")).
Contributions
- Bedz01: the
--preload= form, the Claude-Code-only gate, an independent 5 / 5 / 0 on 2.1.308.
- omgapnt: first Windows and desktop-app results, Windows path tips.
- anhdoox: the Windows
Bun.main and execPath probe, a working gate, the user-variable install.
- declanbx: macOS results, the ~8 % replay estimate, the build-check fork.
- Genbox: first look at the desktop app's source.
- lostsock: the first macOS report, on HN.
- wtfsayo: the correction that a summary can return to the model after a compaction.
- apoapostolov: an independent Windows gate probe on 2.1.293 and 2.1.301, and the failed-preload hazard.
- tonyvince: the answer to the
awaySummaryEnabled question.
License and disclaimer
MIT © 2026 Flamerix AI. Provided as is, without warranty of any kind;
you use it at your own risk. Flamerix AI is not liable for any damage, data loss or account
action arising from its use. Changing a client's runtime behaviour may conflict with the terms of
the service you use it with; checking that is on you.
"Claude" and "Claude Code" are trademarks of Anthropic, PBC. This project is independent and is
not affiliated with, sponsored or endorsed by Anthropic.