DeepSeek Harness's own fast mode: the same model, half the output, up to twice as fast.
Several switches you combine freely, all prompt-side: no model swap, no lower reasoning effort, no request field written.
Best measured: 0.505x wall clock, output tokens −42%, 438/438 checks green — benchmark and repro.
---
## At a glance
30 complete task pairs, same model, same settings, only three switches apart:
| Metric | Result |
|---|---|
| **Wall clock** | **0.505x** (about twice as fast; 26 of 30 pairs faster) |
| **Output tokens** | **−42%** (16037 → 9380) |
| **Reasoning characters** | **−51%** (42380 → 20872) |
| **Task quality** | **438/438** checks green, 30/30 runs passing |
All three switches plus the panel defaults is the fastest setting; the bolt alone, with the three newer switches off, is **0.823x** over the full 138-pair pool (18 % saved).
## Install
```bash
dsh plugin install dsh-fast-mode
```
Click the bolt in a conversation. Right-click it for the eight-row panel — every row affects that conversation only.
## What the eight buttons do
All eight are prompt-side: no request parameter, no model change, no change to completion checks. The first two shape output, the last one shapes how much context arrives, the five in between shape the route.
| Button | How it works | Default |
|---|---|---|
| **Concise output** | No call previews, no restating results, no progress reports; the final answer is not shortened | On |
| **Step cap** | A hard cap; at the limit it reports status instead of truncating the task | No cap |
| **File inventory** | Hands the model a directory listing up front, skipping the opening `ls`/`find` | Off |
| **Batched reads** | Independent reads go out as one batch, fewer round trips | Off |
| **Trim tool output** | Keeps head, tail and matching error lines instead of the whole output | Off |
| **Verify before guessing** | Read the uncertain fact once; not pure speed — it saves the retry a wrong guess costs | Off |
| **Act before explaining** | At most one line of plan; past 4000 characters of reasoning, a nudge to act on evidence | Off |
| **Nudge on repeats** | The same call again, or three identical results: one line saying to change the approach | Off |
The middle five arrived in 1.7.0 and ship **off**; the reasons are under Limits below. Trim tool output reached the panel in 1.7.12 — it is the one switch that changes the evidence the model sees, so it ships off, and off means the original text.
## Benchmark
All three on (`trio17`) vs all off, paired per task, 4000 bootstrap resamples:
| Block | Pairs | Wall clock (geometric) | 95 % CI | Faster |
|---|---|---|---|---|
| DP17 | 12 | 0.647x | [0.501, 0.893] | 10/12 |
| DP18 | 9 | 0.491x | [0.314, 0.788] | 8/9 |
| DP20b (shipped build) | 9 | 0.372x | [0.240, 0.563] | 8/9 |
| **Pooled** | **30** | **0.505x** | **[0.401, 0.635]** | **26/30** |
On the shipped build: steps 13.2 → 6.7, wall clock 464.9 s → 173.1 s. In every block, all-three-on is the only arm whose interval stays below 1. Raw data and recompute commands are in [BENCHMARK.md](BENCHMARK.md).
## Why it is faster, and where it is not
A step costs roughly round-trip overhead × steps + decode time × output tokens + tool wait. The model is not in that equation, so this only moves programmatic quantities: fewer round trips, fewer re-reads, fewer repeated edits.
- **Some tasks get slower — do not expect a win everywhere.** 1–5 step repairs and data reshaping do not repay the hint cost (`perfrefactor` 1.18x, `csvpipe` 1.12x); implementation-shaped tasks occasionally take one extra step (one 1.04x pair in DP20).
- **Quality first.** In every accepted batch both arms produced identical hidden assertions; safety checks, destructive operations and accessibility paths always run the full validation.
- **The switch is yours.** Only concise output is on by default; the three newer switches stay off because their payoff varies by task shape.
Recommendation: **most of the time, turn all three on.** If you want the safest single switch, take *Verify before guessing*; for long reasoning chains take *Act before explaining*; *Nudge on repeats* alone measured 1.007x and fired 3 times across twelve sessions — harmless in the combination, not worth switching alone.
## License
MIT. Reference: [docs/REFERENCE.en.md](docs/REFERENCE.en.md). Changes: [CHANGELOG.md](CHANGELOG.md).
-blue){kind=icon}
