stream-speed

Records how fast each model starts answering and how fast it writes, and shows it with /speed.

Hermes fires three observer hooks around every streamed model response: on_stream_start, on_stream_delta and on_stream_end. stream-speed pairs them into one row per response: the time until the first visible text, the total time, and how many characters arrived. /speed turns those rows into medians per model, so you can see which model or provider keeps you waiting before the first word and which one writes slowly once it starts.

Install

hermes plugins install stream-speed --enable

Requires Hermes 0.21.5 or newer. No API key, no configuration, no dependencies.

Use

/speed          last 24 hours, by model
/speed all      everything recorded
/speed clear    delete the record

Example (/speed all, from the end-to-end test: two one-shot turns against a loopback provider that sends 16 characters every 0.15 s):

Streaming speed, all recorded:

model       streams  failed  first text    p90  chars/s
fake-model        2       0       0.92s  1.00s      111

• first text is the median time from the start of the streaming request to the first visible text; p90 is the slow end of the same number.
• chars/s is the median writing speed between the first text and the end of the stream. It counts characters, not tokens: the streaming hooks carry no token counts.
• failed counts streams that errored or did not finish.
• Streams with no visible text (a response that is only tool calls) are counted, shown with no first-text time, and mentioned in a footer line.

How accurate it is

Hermes delivers stream hooks off the token path, on a queue with its own worker thread per hook, so each timestamp is taken when the event is delivered, a few milliseconds after the wire. first text starts when Hermes begins the streaming request, so it includes connection setup as well as the provider's own delay; that is the wait you actually sit through. In the test above the provider's 107 chars/s is measured as 110-111. Reasoning deltas are ignored: first text means text you can read.

If the first-delta event is still queued when the end event arrives, the plugin waits up to 0.25 s for it; a stream whose order still cannot be established is recorded with an unknown first-text time and is never reported as zero.

What is recorded

One JSON line per streamed response in <HERMES_HOME>/plugin-data/stream-speed/streams.jsonl (per profile):

ts, session_id, iteration, provider, model, surface, ttft_s, duration_s, chars, finished, failed.

Not recorded: the streamed text, reasoning text, prompts, error messages (only a failed flag). The file is bounded: once it passes 2 MB it is rewritten to its newest 5,000 rows.

Security and footprint

• register() only registers three hooks and one command (/speed), plus an atexit handler that waits at most one second for a stream still open at exit (so a one-shot run records its last answer).
• No network access, no subprocesses, no downloads, no credentials read.
• Writes only streams.jsonl under its own plugin data directory. /speed clear deletes it.
• The streamed text passes through the hook payloads in memory; only its length is kept.
• Hooks are observer-only and never raise: an unwritable record is logged at debug level.
• In memory it tracks at most 256 open streams and drops any older than an hour.
• Standard library only.

Development

python -m pytest tests
PYTHONPATH=<hermes checkout> python tests/e2e/stream_speed_e2e.py
hermes plugins validate stream-speed

MIT licensed.