Back to directory

dsh-fast-mode

In a curated listMaintenance: Active

regrevia/dsh-fast-mode

Adds a per-conversation 'fast mode' toggle (lightning chip + settings panel) that trims output narration, widens tool fan-out and bounds/cancels stalled tool calls to cut turn latency.

View on GitHubHomepage
$ dsh plugin install dsh-fast-mode

16

stars

3

forks

JavaScript

Language

MIT

License

2026-10-24

Created

2026-11-18

Last push

MIT-licensed native DSH plugin whose per-session lightning chip toggles prompt-side brevity and reasoning-hygiene switches plus a stalled-tool timeout, backed by a published paired benchmark and a self-measurement harness.

DSH integration

Native runtime

Author-claimed

Safety audit

Unaudited

Last verified

2026-11-18

License

MIT

01What can it help you accomplish?

  • Cut the latency of a DeepSeek Harness session without changing the model or any request field

    Fewer round trips and less narration: the author's 30-pair bench reports wall-clock time halved, output tokens −42% and grading 438/438 green

    DSH users running long agentic coding tasks who want faster turns without a quality trade

  • Check whether the speed-up holds on your own repository and task list

    A dry run of the expanded commands, or paired real runs writing `pairs.jsonl` and a `report.md` with average speed-up and variance under `audit-out/<timestamp>/`

    Teams that want local evidence before enabling a latency feature for everyone

02How to install into DeepSeek Harness

Prerequisites

  • DSH with the `dsh plugin` CLI available

Installation steps

  1. 01

    $ dsh plugin install dsh-fast-mode

  2. 02

    Click the lightning chip inside a session to switch fast mode on; right-click it to open the settings panel — the README stresses it only affects the current session

Verify the integration

Not specified by the author

03DSH integration and capability boundaries

DSH integrationNative runtime

Native DSH plugin installed with `dsh plugin install dsh-fast-mode`; it adds a lightning chip to the session UI whose right-click panel toggles prompt-side switches that apply to that session only.

  • Per-session fast mode

    a click (or right-click) on the lightning chip in a session→the switches applied to that session only

    changes the prompt guidance of the current session only
  • Concise output (精简输出)

    any turn in a fast-mode session→no tool-call previews, no restating of results, no progress notes — the final answer is not shortened

  • Three reasoning-hygiene switches

    turns where the agent is unsure, plans too long or repeats a call→reminders to read once before asserting, to act after at most one line of plan, and to change method when a call repeats or three results match

    injects reminder text into the prompt; no request field is written
  • Stalled-tool auto-cancel

    a single tool call that runs too long→the call is cancelled and control returns; the threshold is at least 150 s and sits after the host has moved long commands to the background

    can cancel a long-running tool call
  • Self-measurement harness

    `node audit/run.mjs --dry-run --cwd <repo> --tasks <list>`, then the same command without `--dry-run`→paired real runs (3 rounds per task per arm) writing `pairs.jsonl` raw readings and a `report.md` under `audit-out/<timestamp>/`

    runs your tasks repeatedly and writes an audit-out/ directory

04Who is it for? When not to use it?

Good for

  • DSH users running long agentic coding tasks who want faster turns without a quality trade
  • Teams that want local evidence before enabling a latency feature for everyone

Not for

  • Not a universal speed-up: 1–5-step small fixes and data reshaping cannot pay the overhead back (perfrefactor 18% slower, csvpipe 12% slower), and implementation tasks occasionally take one extra step.

05Compatibility, maintenance and safety notes

  • Not a universal speed-up: 1–5-step small fixes and data reshaping cannot pay the overhead back (perfrefactor 18% slower, csvpipe 12% slower), and implementation tasks occasionally take one extra step.
  • Only two presets are offered in the panel; four other knobs (step cap, opening file list, batched file reads, truncated tool output) were withdrawn because no stable speed-up could be measured — their config keys remain but they are not exposed as switches.
  • The published numbers come from the repository's own bench: the README says your machine and tasks may differ, and reports with fewer than 6 pairs only state that the sample is insufficient.
  • The tool-timeout auto-cancel is a failure guard, not a speed-up: the threshold is at least 150 seconds and only fires after the host has already moved long commands to the background.
2026-10-242026-11-15v1.7.45

MIT · README 分支 main · 最新 release v1.7.45 (2026-11-18)

06Frequently asked questions

How do I turn fast mode on in DeepSeek Harness?

Run `dsh plugin install dsh-fast-mode`, then click the lightning chip inside a session. Right-clicking the chip opens the settings panel, and the README stresses that the switches only affect the current session.

Does it change the model, the reasoning effort or the request?

No. Every switch is prompt-side: it adds no request field, does not change the model, does not lower reasoning effort and does not change the completion criteria. Only the wording of the guidance the agent follows changes.

Which switches does the panel expose?

Concise output (on by default), plus three reasoning switches — verify before asserting, act before explaining, change method on repetition — that load with the standard preset, and a fifth line that auto-cancels a stalled tool call (on by default). The conservative preset turns everything off.

Will it always be faster?

No. The README states that 1–5-step small fixes and data reshaping lose more than they gain (perfrefactor 18% slower, csvpipe 12% slower) and that implementation tasks sometimes take one extra step. Its measured gains come from 30 paired tasks on the repo's own bench.

How do I measure the effect on my own tasks?

Run `node audit/run.mjs --dry-run --cwd <repo> --tasks <list>` to see the expanded commands, then drop `--dry-run` for paired real runs (3 rounds per task per arm). Outputs land in `audit-out/<timestamp>/` as `pairs.jsonl` plus a `report.md`; with fewer than 6 pairs the report only says the sample is insufficient.

08Data and sources

  • Author-claimedgithub.com77ba97235dd9…

    dsh plugin install dsh-fast-mode

  • Author-claimedgithub.com77ba97235dd9…

    装完在会话里点闪电。右键闪电打开面板,**只影响当前会话**。

This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-18. Found an error? Submit a correction.

🏆

Best DeepSeek Harness Plugins

Twelve plugins worth installing first — picked from the whole catalog, across every category.

DSH Plugins is an independent community directory of DeepSeek Harness plugins. Not affiliated with or endorsed by DeepSeek. Third-party plugins are not security-audited — review the source before installing.

New DeepSeek Harness plugins, weekly. No spam.