All plugins

stream-speed

How fast each model starts answering and how fast it writes: /speed shows the median and p90 time to first text and the median characters per second per model, provider or day, or lists the newest streams (ten, or up to 100 with /speed recent 50), for the last 24 hours, any number of hours or days, the latest session or everything recorded, from the streaming hooks. Disclosure — writes per-response stream timings (ids, model, timings, character count, no text) to plugin-data/stream-speed/streams.jsonl, trimmed to 5,000 rows past 2 MB; atexit waits up to 1 s; no network

olric3015Toolsv1.4.0Updated 1 month ago

About

How fast each model starts answering and how fast it writes: /speed shows the median and p90 time to first text and the median characters per second per model, provider or day, or lists the newest streams (ten, or up to 100 with /speed recent 50), for the last 24 hours, any number of hours or days, the latest session or everything recorded, from the streaming hooks. Disclosure — writes per-response stream timings (ids, model, timings, character count, no text) to plugin-data/stream-speed/streams.jsonl, trimmed to 5,000 rows past 2 MB; atexit waits up to 1 s; no network

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 session     the most recent session that streamed a response
/speed all         everything recorded
/speed 7d          the last 7 days (any number of hours or days: 6h, 30d)
/speed providers   add to any of the above to group by provider instead of model
/speed days        add to any of the above to group by day, newest first
/speed recent      add to any of the above to list the newest 10 streams instead
/speed recent 25   the newest 25 instead (1 to 100)
/speed clear       delete the record

/speed session is the quick check after a conversation that felt slow. /speed providers
compares providers that serve you the same model, and a time window (/speed 6h, /speed 7d)
shows whether a slowdown is new. The record keeps its newest 5,000 streams, so a long window on
a busy install reaches back only as far as those rows do.

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.

/speed days puts one day on each line, newest day first, so you can see the day a provider
became slow. Days are your computer's local calendar days, and a day with no streams has no
line. Example (the plugin's own output for a sample week):

Streaming speed, last 7 days:

day         streams  failed  first text    p90  chars/s
2026-10-06        4       1       4.20s  5.10s       79
2026-10-05        2       0       1.15s  1.30s      127
2026-10-04        3       0       0.90s  1.10s      139

1 stream(s) carried no text (tool calls only, or failed before any).

/speed recent lists the newest 10 streams of the window one by one (/speed recent 25 lists
25, up to 100), newest first, so the one
slow answer behind a median can be found. total is the whole stream, and result is ok,
failed (the stream errored) or unfinished (it ended without finishing). Example (the
plugin's own output for sample rows):

Streaming speed, last 7 days, newest 5 of 5:

when       model    provider  first text   total  chars/s      result
1m ago   m-small      custom       1.12s   4.95s      452          ok
40m ago  m-large  openrouter       0.74s   7.88s      370          ok
42m ago  m-large  openrouter           -  30.00s        -  unfinished
3h ago   m-small      custom       1.37s   6.02s      406          ok
2d ago   m-large  openrouter       0.82s   9.41s      363          ok

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

Changes

  • 1.4.0: /speed recent takes a row count, /speed recent 25 (1 to 100; ten without
    one). What is recorded did not change.
  • 1.3.0: /speed recent, the newest 10 streams of any window, newest first, with their
    model, provider, time to first text, total time, writing speed and result. What is recorded
    did not change.
  • 1.2.0: /speed days, one line per day with streams, newest day first; works with every
    window (/speed 30d days). The record also reads correctly after a Windows tool has saved it
    with a byte-order mark. What is recorded did not change.
  • 1.1.0: /speed session, time windows (/speed 6h, /speed 7d, any number of hours or
    days) and grouping by provider (/speed providers). What is recorded did not change.
  • 1.0.0: first release.

MIT licensed.

README from the reviewed commit d7b48bb; it updates when the author re-pins.

More by olric3015

18 plugins by olric3015 in the catalog.

olric3015approval-ledgerLocal ledger of Hermes's approval prompts (dangerous commands, protected writes, MCP consent, smart-mode verdicts): /approval-log shows how often each rule asked, how many prompts were approved, denied or left unanswered, and how long a person took to answer, per rule, per surface or per day, for the last 24 hours, any number of hours or days, the latest session or everything recorded; /approval-log recent lists the newest decisions (ten, or up to 100 with /approval-log recent 50). Disclosure — writes decision metadata (session id, surface, rule label, answer, timings; no command, description or session key) to plugin-data/approval-ledger/approvals.jsonl, trimmed to 5,000 rows past 2 MB; observer only, cannot answer or change an approval; no networkCommunitytoolsolric3015aux-ledgerLocal ledger of Hermes's auxiliary LLM calls (titling, compression, vision, approval): /aux shows tokens, time and failures per task, model, provider or day, or lists the newest calls (ten, or up to 100 with /aux recent 50), for the last 24 hours, any number of hours or days, the latest session or everything recorded. Disclosure — writes auxiliary-call metadata (ids, models, timings, token counts, no prompt or response text) to plugin-data/aux-ledger/calls.jsonl, trimmed to 5,000 rows past 2 MB; no networkCommunitytoolsolric3015b64-ref@b64: context references that decode encoded text for the model: @b64:<value> takes base64 (standard or URL-safe, padded or not), hex (0x48656c6c6f, 48:65:6c, \x48\x65), percent-encoding or a data: URI, told apart by the value's shape or picked with base64:, base64url:, hex: or url:, and attaches what it holds: text as text with control and invisible characters escaped, JSON pretty-printed, gzip and zlib inflated in memory (at most 1 MiB, one layer), and binary named by its format (PNG and GIF with their size, JPEG, PDF, ZIP, ELF, SQLite and more) with a hex dump of its first 128 bytes; notes point out a value that was cut short, padding that is off, a second layer of encoding and a JWT inside; the TUI and Desktop composer preview a value as you type it. Disclosure — reads only the text of the reference itself, using Python's base64, json, urllib.parse and zlib; nothing decoded is run, opened or saved; no network, no files read or written, no subprocess, no credentials read; the plugin keeps no copy of what it decodes, but the @b64: reference and the decoded contents become part of the message sent to your model provider and saved in the session history, so decode only values you are willing to share with it; register() only adds the @b64: reference provider.Communitytoolsolric3015chmod-ref@chmod: context references that explain a file permission mode for the model: @chmod:<mode> takes an octal mode (755, 0755, 4755), ls -l text (-rwsr-xr-x, with the file type letter and a trailing + . or @) or an st_mode as stat and git write it (100644), and attaches both forms, each class's rights in words (directory rights named as such), what setuid, setgid and the sticky bit do and where they are ignored, and notes such as a mode everyone can change, one private to the owner, or a class holding rights the class before it lacks; @chmod:"644 u+x,go-w" applies a symbolic change step by step the way GNU chmod does (u g o a, + - =, r w x X s t, copies like u=g), @chmod:"100644 100755" shows what changes between two modes, and @chmod:"umask 027" shows what new files and directories get; the TUI and Desktop composer preview a mode as you type it. Disclosure — reads only the text of the reference itself, using Python's re; it never looks at a file on disk; no network, no files read or written, no permission changed, no subprocess, no credentials read; the plugin keeps no copy of what you write, but the @chmod: reference and the explanation become part of the message sent to your model provider and saved in the session history; register() only adds the @chmod: reference provider.Communitytools

Developer Mode

Explore built-in Hermes toolsets and actions.

HermesAgent

Developer?

Create a Hermes extension for Desktop or agent workflows.

Developer Section