opencode-gliner-prune

Local, non-destructive context pruning for OpenCode using fastino/gliner2.5-small-v1.

The plugin analyzes completed tool interactions locally, keeps recent and mutating operations, preserves exact evidence spans, and changes only the cloned context sent to the model. OpenCode's stored session remains unchanged.

Requirements

• Apple Silicon macOS
• OpenCode 1.18.31
• uv

Install

Add the npm package to opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-gliner-prune"]
}

Restart OpenCode. The plugin immediately prepares and loads the local model in the background. On first launch it downloads the checkpoint; subsequent loading and inference are local.

When context has accumulated, run:

/gliner-prune

The completion toast reports the reduction, changed-action counts, preserved evidence, and confirms that stored history was not modified. A new substantive prompt automatically restores full context; exact continuations keep pruning active.

For local development, build from the repository root and use:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": [
    "file:///absolute/path/to/gliner25-compaction/integrations/opencode"
  ]
}

Transcript content stays on the machine and inference is forced offline after the checkpoint is available.

Safety model

• Mutating tools, question answers, recent interactions, unresolved errors, requirements, URLs, and generated identifiers are protected.
• Unknown tools and shell commands are treated as mutating unless they match a narrow observational allowlist.
• Low-confidence or malformed model output resolves to keep_full.
• A worker crash, timeout, or invalid span sends the original full context.
• A substantive new user prompt clears active pruning automatically. Only an explicit continuation such as Continue. carries decisions forward.

Token counts shown by the plugin are estimates. Character counts cover the canonical text and tool payloads analyzed by GLiNER, not OpenCode's provider framing or metadata.

See the shared architecture, safety policy, and evaluation guide.