What the plugin costs

The plugin buys ambient memory by spending latency and context. Until 2026-08-11 nobody had measured either, the README's "~10 ms, never on the agent's hot path" was carried forward as though it described the whole plugin, when it describes the capture hook alone.

Reproduce with:

uv run python scripts/measure_hook_overhead.py          # realistic, dedup on
uv run python scripts/measure_hook_overhead.py --fresh  # worst case, dedup defeated
uv run python scripts/measure_hook_overhead.py --json

Measured

Against a local server, real fleet memory, real agent identity, 7 samples per hook, per-session dedup active:

Hook	Fires	p50	p95	Tokens (first → then)
SessionStart	per session	696 ms	729 ms	612 → 612
UserPromptSubmit · inbox	per prompt	65 ms	69 ms	0 → 0
UserPromptSubmit · recall	per prompt	195 ms	221 ms	112 → 88
PreToolUse · gotcha	per tool call	189 ms	204 ms	123 → 0
Stop · capture	per turn	16 ms	17 ms	0 → 0

For a session of 20 prompts and 60 tool calls: ≈ 17.6 s of added wall-clock and ≈ 4,200 tokens of context.

Two things worth reading off that table:

• The dominant cost is PreToolUse, because it fires on every tool call. At 189 ms it is most of the added wall-clock, and it is not the hook anyone would have guessed.

• The drainer does more since these numbers were taken. It now also extracts a token rollup from the same transcript slice and posts it to /usage. That work is in the detached drainer, not the hook, so the 16 ms hot-path figure above still stands, but the measurement predates the change and the drainer's own cost has not been re-measured.

• Stop · capture is genuinely cheap, 16 ms, injects nothing. The original claim holds for the hook it was made about; it just never covered recall, inbox or gotcha.

Numbers to treat with care: this is localhost, so a remote instance is strictly worse; tokens are counted as characters/4; the inbox figure depends entirely on whether that agent has unread messages, and was 547 tokens for an identity that did.

An inert plugin is not a free plugin

With bad credentials every hook returns nothing, logs nothing, and injects nothing. The same session then costs 6.6 s and 0 tokens, all of the latency, none of the benefit, and no signal anywhere that it is happening.

measure_hook_overhead.py exits non-zero when total injection is zero, for exactly this reason. It is also a hazard for any A/B measurement: a treatment arm in this state is a placebo, and would produce a clean, entirely meaningless null result. Any benchmark comparing Artel on/off must assert non-zero injection while it runs, not merely check that the plugin is installed.