opencode-dcp

Published Opencode plugin that implements Dynamic Context Pruning. Reduces token usage in long sessions by rewriting outgoing LLM requests — replaces stale c…

opencode-dcp

Published OpenCode plugin that implements Dynamic Context Pruning (DCP). Reduces token usage in long sessions by rewriting outgoing LLM requests — replaces stale content with placeholders without modifying session history on disk.

Identity

• npm package: @tarquinen/opencode-dcp
• GitHub org: Opencode-DCP
• Repo: Opencode-DCP/opencode-dynamic-context-pruning
• Author handle: tarquinen
• Docs site: opencodedocs.com/Opencode-DCP/opencode-dynamic-context-pruning/

Install

opencode plugin @tarquinen/opencode-dcp@latest --global

Or manually add to opencode.jsonc:

{
  "plugin": ["@tarquinen/opencode-dcp@latest"]
}

Restart OpenCode. Verify with /dcp.

Key Mechanisms (one-line each)

• Compress — model-driven compress tool that summarizes closed conversation spans (range or experimental message mode)
• Deduplication — automatic; on LLM fetch, keeps only the most recent output for identical (tool, args) pairs
• Purge Errors — automatic; prunes the input of errored tool calls after N turns (default 4), keeps the error message

Config File Locations

DCP-specific config (separate from opencode.jsonc), searched in order, later overrides earlier:

1. ~/.config/opencode/dcp.jsonc — global (auto-created on first run, contains only $schema initially)
2. $OPENCODE_CONFIG_DIR/dcp.jsonc — when env var is set
3. .opencode/dcp.jsonc — project-local

.json variants are also accepted at every level.

Default Protected Tools

task, skill, todowrite, todoread, compress, batch, plan_enter, plan_exit, write, edit

Slash Commands

/dcp, /dcp context, /dcp stats, /dcp sweep [n], /dcp manual [on|off], /dcp compress [focus], /dcp decompress <n>, /dcp recompress <n>

Logs

~/.config/opencode/logs/dcp/daily/ — when debug: true is set in config.

Key Config Knobs

DCP config at ~/.config/opencode/dcp.jsonc (auto-created on first run). All defaults live in code; only override what you need.

"compress": {
  "mode": "range",          // "range" (default) or "message" (experimental)
  "permission": "allow",    // "allow" | "ask" | "deny"
  "minContextLimit": 50000, // below this: reminders off
  "maxContextLimit": 100000,// above this: strong nudges every nudgeFrequency fetches
  "nudgeFrequency": 5,
  "protectUserMessages": false,
  "summaryBuffer": true
}
"strategies": {
  "deduplication": { "enabled": true },
  "purgeErrors": { "enabled": true, "turns": 4 }
}

Both limits accept absolute numbers or "X%" of model context. Per-model overrides via modelMinLimits/modelMaxLimits.

Other knobs: enabled, debug, pruneNotification ("off"|"minimal"|"detailed"), manualMode.enabled, turnProtection, protectedFilePatterns, experimental.allowSubAgents, experimental.customPrompts.

Prompt Cache Trade-Off

DCP rewrites outgoing requests, invalidating prompt-cache prefixes from the prune point onward. Plugin author’s measurements: ~85% cache hit rate with DCP vs ~90% without. Lost cache reads trade against token savings and fewer hallucinations from stale content. No cache impact for request-based billing (GitHub Copilot) or uniform-rate pricing (Cerebras).

Layer Relationship

Layer	When it fires	What it does
OpenCode built-in compaction	Token threshold	End-of-context summarization
lean-session plugin	Same as built-in	Shapes compaction prompt via experimental.session.compacting
DCP	On LLM fetch (dedup/purge) + model decision (compress)	Keeps active payload lean

These compose: DCP keeps context lean → compaction fires later on cleaner input → lean-session shapes what survives compaction.

Related Pages

• Dynamic Context Pruning (DCP) — what it does and why
• OpenCode — host harness