@skybluejacket/opencode-context-compress
v0.4.3
Published
Manual context compression plugin for OpenCode sessions
Downloads
454
Maintainers
Readme
opencode-context-compress
Manual-first context compression. You own the when.
OpenCode plugin for explicit, user-triggered context compression. It helps the model fold completed conversation phases into durable technical summaries only when you ask for it.
Core Behavior
- No autonomous context management loops.
- No automatic nudges or per-turn injections.
- Compression runs only when you trigger
/compress manage. - During
/compress manage, the agent can usecompress_mapandcompress(subject to permissions). - After that management turn completes, its trigger, tool calls, tool outputs, map snapshots, and assistant chatter are hidden from future model prompts.
Commands
/compressor/compress help: show command help./compress manage: send a lean context-management reminder to the active agent./compress context: show token usage breakdown for the current session./compress stats: show session and all-time compression totals.
/compress manage is the only command that intentionally creates a model-visible turn.
Agentic Workflow
When /compress manage runs, the plugin opens a single model-visible management turn with a short reminder. Inside that turn the agent can:
- Call
compress_mapto fetch the current<compress-context-map>snapshot. - Call
compresswith one range at a time to replace completed phases with topical blocks. - Read the refreshed map returned by
compressand continue iterating in the same turn if needed.
The manual boundary stays absolute: outside a user-triggered /compress manage turn, the plugin does not prompt for compression or open any background workflow.
While the management turn is running, the agent can see its own maps and tool results. On later turns, the model-visible context contains only compressed [bN] blocks, normal conversation between compression runs, and the active tail. The cleanup leaves no marker or placeholder behind.
Context Map
compress_map and compress both use the same structured map format:
<compress-context-map>
[1] user: "Let's implement JWT auth"
[2-4] assistant: 5 tool calls - auth exploration (~1,240 tokens)
[b0] [compressed] "Prior database migration debugging" (~420 tokens)
[5] user: "Looks good, now add tests"
[6-8] assistant: 4 tool calls - test implementation (~2,180 tokens)
---
Total: 8 messages + 1 block | ~6,500 tokens
</compress-context-map>The agent decides what counts as the active tail. Older completed work should be compressed more tersely than the most recent completed phase. Block labels follow where their anchors appear in the conversation stream, so re-compressing one block does not renumber unrelated blocks.
Installation
npm (Recommended)
npm install @skybluejacket/opencode-context-compressThen add it to your config:
| Platform | Global | Project-level |
| --- | --- | --- |
| OpenCode | ~/.config/opencode/opencode.jsonc | .opencode/opencode.jsonc |
| OpenCodeOrchestra | ~/.config/oco/oco.jsonc | oco.jsonc or .oco/oco.jsonc |
{
"plugin": ["@skybluejacket/opencode-context-compress"]
}From Source
Clone the repo, build, and reference the compiled entry file directly:
git clone https://github.com/AidenGeunGeun/opencode-context-compress.git
cd opencode-context-compress
npm install
npm run build{
"plugin": ["file:///absolute/path/to/opencode-context-compress/dist/index.js"]
}Configuration
Config files are loaded and merged in this order:
~/.config/opencode/compress.jsonc(orcompress.json)$OPENCODE_CONFIG_DIR/compress.jsonc(orcompress.json)<project>/.opencode/compress.jsonc(orcompress.json)
If no global config exists, the plugin creates ~/.config/opencode/compress.jsonc with:
{
"$schema": "compress.schema.json"
}Default runtime config:
{
"enabled": true,
"debug": false,
"notification": "detailed",
"notificationType": "chat",
"commands": {
"enabled": true,
"protectedTools": ["task", "todowrite", "todoread", "compress", "compress_map", "batch", "plan_enter", "plan_exit"]
},
"turnProtection": {
"enabled": false,
"turns": 4
},
"protectedFilePatterns": [],
"tools": {
"settings": {
"protectedTools": ["task", "todowrite", "todoread", "compress", "compress_map", "batch", "plan_enter", "plan_exit"]
},
"compress": {
"permission": "allow",
"showCompression": false
},
"compress_map": {
"permission": "allow"
}
}
}Persistence
Session state is stored at:
~/.local/share/opencode/storage/plugin/compress/<sessionId>.json
Stored fields include:
- compressed tool IDs
- compressed message IDs
- compression summaries
- completed management-turn cleanup markers
- per-session compression stats
The raw conversation history still exists in OpenCode storage, but completed /compress manage machinery is suppressed from future model prompts. Restarting the session reloads the saved cleanup markers, so old management turns do not reappear in the model-visible stream.
Development
npm install
npm run generate:prompts
npx tsc --noEmit
npm testPrompt utility docs are in scripts/README.md.
License
MIT
