# dsh-dcp
**Dynamic context pruning for the DeepSeek Harness — let the model tidy up after itself.**
[](https://www.npmjs.com/package/@zzxxxxxx/dsh-dcp)
[](LICENSE)
[](https://github.com/deepseek-ai/deepseek-harness)
**English** | [中文](README.md)
add ...` only adds a dependency to the profile — the plugin row itself is mounted by the bundle patch this package ships. - Update: from source, `git pull && npm install && npm run build`; from git, `add` again. Restart either way. - Remove: `dsh plugin --profile
remove @zzxxxxxx/dsh-dcp`, then restart.
- The browser half (the readout pill and the settings page) appears only in profiles with a Web UI; the model tools, compaction and deduplication work everywhere else.
## Usage
- **`/dcp-compact [focus]`** asks the model to run one compaction pass now, with optional focus text to narrow it. The host's own `/compact` is a separate path; the two do not interfere.
- **Model-facing tools**: `compact` (fold a span into a summary the model writes), `compact_targets` (list handles and legal cut marks), `recall` (read an original back; `query` filters lines, `offset` pages through it). Availability follows `compaction.permission`.
- **Automatic**: one deduplication pass per turn boundary; while context pressure is above a threshold the model is reminded at most once every `nudgeFrequency` nodes.
- **Manual mode** (`manualMode.enabled`): nudges stop, and the model may compact once only after you run `/dcp-compact`.
- **Readout**: the pill under the input shows this session's compaction blocks, prune count and tokens saved; click it for the breakdown.
## Configuration
**Settings → DCP context** (the host's `settings.section`, id `dsh-dcp`) lists the 12 fields that can be changed at runtime and applies them immediately; the other 8 live in the plugin row's `config:` and need a restart.
```jsonc
{
"pruneNotification": "minimal", // off | minimal
"protectedFilePatterns": [],
"turnProtection": { "enabled": false, "turns": 4 },
"experimental": { "allowSubAgents": false, "customPrompts": false },
"commands": { "protectedTools": ["subagent*", "skill", "compact", "write", "edit", "plan_enter", "plan_exit"] },
"manualMode": { "enabled": false, "automaticStrategies": true },
"compaction": {
"permission": "allow", // allow | ask | deny
"maxContextLimit": "67%", // share of the window, or absolute tokens
"minContextLimit": "33%",
"modelMaxLimits": {}, "modelMinLimits": {},
"nudgeFrequency": 5,
"protectedTools": ["subagent*", "skill"],
"protectUserMessages": false,
"recall": true
},
"strategies": { "deduplication": { "enabled": true, "protectedTools": [] } }
}
```
| Field | Default | Meaning | Hot |
| --- | --- | --- | :---: |
| `pruneNotification` | `"minimal"` | What an automatic prune pass reports: `minimal` is one line, `off` leaves it to the panel | ✓ |
| `protectedFilePatterns` | `[]` | Path globs: matching calls are never rewritten, and their bodies are preserved in a summary's appendix | |
| `turnProtection.enabled` | `false` | Keep the newest turns out of the strategy candidates | ✓ |
| `turnProtection.turns` | `4` | How many turns stay protected | ✓ |
| `experimental.allowSubAgents` | `false` | Let DCP prune subagent sessions too (off leaves them alone) | ✓ |
| `experimental.customPrompts` | `false` | Enable prompt overrides under `$DSH_HOME/dcp-prompts/` | ✓ |
| `commands.protectedTools` | `subagent*` `skill` `compact` `write` `edit` `plan_enter` `plan_exit` | One of the three protection lists; also governs the `/dcp-compact` sweep and the panel | |
| `manualMode.enabled` | `false` | Stop nudging; the model may compact once per explicit request | ✓ |
| `manualMode.automaticStrategies` | `true` | Keep deduplicating while manual mode is on | ✓ |
| `compaction.permission` | `"allow"` | `allow` runs directly; `ask` goes through the Harness approval seam; `deny` leaves the model tools unregistered | |
| `compaction.maxContextLimit` | `"67%"` | Strong threshold: above it the model is told to compact now | ✓ |
| `compaction.minContextLimit` | `"33%"` | Weak threshold: below it nothing nudges | ✓ |
| `compaction.modelMaxLimits` | `{}` | Per-`"provider/model"` override for the strong threshold | |
| `compaction.modelMinLimits` | `{}` | Per-`"provider/model"` override for the weak threshold | |
| `compaction.nudgeFrequency` | `5` | Inject a nudge at most once every N conversation nodes | ✓ |
| `compaction.protectedTools` | `subagent*` `skill` | Bodies are appended verbatim to a summary, and exempt from deduplication | |
| `compaction.protectUserMessages` | `false` | Preserve user messages verbatim (a large pasted prompt never compresses away) | ✓ |
| `compaction.recall` | `true` | Register the `recall` tool | |
| `strategies.deduplication.enabled` | `true` | Rewrite the outputs of older duplicate calls | ✓ |
| `strategies.deduplication.protectedTools` | `[]` | Extra exemptions from deduplication | |
A few semantics:
- A percentage threshold needs the route to report a context window; when it reports none, nothing nudges — the plugin does not fall back to a fixed token count.
- All three protection lists reach deduplication: naming a tool in `compaction.protectedTools`, `commands.protectedTools` or `strategies.deduplication.protectedTools` keeps its older outputs from being rewritten (`commands`' default list is what exempts `write`/`edit`).
- `protectedFilePatterns` governs both deduplication and compaction: matching calls are never rewritten, and their bodies are kept in a summary's appendix. Globs are fail-open: a value over 8192 characters, a pattern over 4096 characters, or a match running past 1,048,576 steps counts as "no match" — that loses protection, never adds it; collecting the paths one call declared scans at most 10,000 nodes.
- With `experimental.customPrompts` on, `$DSH_HOME/dcp-prompts/overrides/