dsh-compaction-policy
按提供商/模型将 DSH 的自动上下文压缩触发点设置在你需要的位置。防止 DeepSeek Harness 在上下文窗口较小的模型上，远早于配置的阈值压缩对话。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:zyd232/dsh-compaction-policy#6ec165b1a1e4f0470acb149a1764c52043467aa2Compaction Policy 以 dsh-compaction-policy 发布，当前版本为 0.1.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
English | 中文
Stop DeepSeek Harness from condensing your conversation too early — and put the trigger where you want it.
DSH automatically condenses older history when a conversation grows. On a lot of models it starts much earlier than "80% full", and then it does it again a few steps later. This plugin gives you that decision back: it does not edit any config file, it adjusts the running compaction engine.
Probably yes if any of these sound familiar:
If you only ever use very large windows (e.g. 1M), the built-in behaviour is already near 80% and you most likely don't need this.
Before each step DSH checks how large the next request would be and compares it to this trigger:
trigger = min(window × 0.8, window − output reserve − 65536)
The second half of that formula is the catch: it subtracts a fixed 65,536 tokens regardless of how large the window is. Once the window is below about 330,000, that fixed subtraction is the tighter of the two limits, so it becomes the real trigger:
| Context window | Where DSH actually triggers | Where 80% would be |
|---|---|---|
| 128k | ~62k (49%) | 102k |
| 150k | ~84k (56%) | 120k |
| 256k | ~190k (74%) | 205k |
| 1M | ~678k (68%) | 800k |
On very small windows (roughly up to 65k) the fixed margin swallows the whole window: DSH then refuses to configure proactive compaction at all and only reacts after a request is rejected for being too long — which throws away far more history than a planned condensation would.
In DSH 0.2 the compaction policy lives inside each agent preset, and a preset is addressed as one whole definition: neither a profile config file nor a plugin's own config layer can change a single field of it. The only way to write it to disk is to replace the preset body — which pins your preset and stops it from picking up DSH updates.
So this plugin doesn't touch any file. It changes the behaviour of the compaction engine while DSH runs.
/compact command
are never touched.DSH ships quickly, so the support line is stated rather than implied.
| DSH version | Status |
|---|---|
0.2.0-rc.2 | Verified — this release was developed and tested against it |
later 0.2.x / 0.3.x | Expected to work. The @deepseek-ai/* peers are open-ended, so a newer harness still installs and loads the plugin; if one of the internal seams it uses changes, it stops intervening instead of breaking your session |
earlier than 0.2.0-rc.2 | Not supported — the plugin wraps the per-agent-preset compaction engine, which only exists in the 0.2 line |
The machine-readable form of this table is the open-ended @deepseek-ai/* range in package.json, and the verified build is recorded there as dsh.verifiedWith.
Two checks tell you it is running on your build:
If it does nothing on your version, that is the designed failure mode rather than a crash: please open an issue with the output of dsh --version.
Last verified: DSH 0.2.0-rc.2, 2026-10-01.
dsh plugin --profile <your profile> add github:zyd232/dsh-compaction-policy
Use the profile you actually run — desktop for the desktop app, web for
dsh web, or your own profile name. Then restart DSH: plugin rows are only
read at startup, so without a restart nothing changes.
You do not need to start a new conversation. Existing conversations are picked up on their next step.
To remove it again:
dsh plugin --profile <your profile> remove dsh-compaction-policy
Open Settings and look at the left sidebar: you will find a Compaction policy entry alongside the built-in sections. The panel follows DSH's language (Settings → General → Language) and defaults to English; Chinese is included.
Everything there is saved to DSH's normal settings document and applies immediately — no restart needed. Each field also has a restore default button that removes your override.
| Field | Default | What it means |
|---|---|---|
| Enable | on | Turn the plugin off to get DSH's stock behaviour back. |
| Routes | * | Which routes it governs: provider or provider/model, comma or space separated. * = every route. |
| Trigger ratio | 0.8 | Condense once the context reaches this share of the window. 0.8 = 80%, DSH's own default. |
| Output reserve (tokens) | 24576 | The margin kept free for the model's answer. DSH's own value is 65536; this is the knob that fixes the early trigger. |
General rule first, examples second.
0.8 unless you want extra safety. On a very
small window some people prefer 0.7.16384–32768 covers most chat and coding work; a model that
writes very long files may want more. Too small a reserve risks a request being
rejected because the answer had nowhere to go.llama-cpp (only local models served by llama.cpp), ollama, openai, or
ollama/* to spell out every model of that provider.With the default reserve of 24,576 the trigger becomes a clean 80% of the window for every window above roughly 123k, and stays proportional below that.
Compaction is recorded in the session log. If you want numbers instead of a feeling, take the request size immediately before a compaction and divide it by the model's window: it should now land near your trigger ratio instead of the old ~50% mark.
Visually, the context ring should keep growing past the point where it used to drop.
agentPresets.serviceFor(…, 'compaction')) and wraps its
compactIfNeeded method. If a future DSH renames that service or changes that
signature, the plugin needs an update — until then it simply does nothing rather
than misbehave.@deepseek-ai/* peer ranges are intentionally open-ended
(>=0.2.0-rc.2, no upper bound), so a future harness release can never prevent
the plugin from installing or loading.