@harnesstrim/adapter-opencode

HarnessTrim adapter for OpenCode. A thin plugin that:

• tool.execute.after — slims noisy tool output (test runs, git diffs, ...) in place using the shared @harnesstrim/core reducers, before it enters the model's context.
• experimental.session.compacting — injects compact-handoff guidance so compaction keeps decision-relevant state instead of narrated history.

The plugin is deliberately thin: all detection and reduction logic lives in @harnesstrim/core (and is unit-tested there without a live harness). See the repo PLAN.md.

Install (local / development)

From an OpenCode project, reference the plugin in opencode.json. During development you can point at the workspace package directly:

{
  "plugin": ["@harnesstrim/adapter-opencode"]
}

With options:

{
  "plugin": [
    ["@harnesstrim/adapter-opencode", { "mode": "active", "minLength": 400, "debug": false }]
  ]
}

Configuration

Options (in opencode.json) take precedence over environment variables, then defaults.

| Option | Env var | Default | Meaning | | ------------------- | ------------------------- | ---------- | -------------------------------------------------------------- | | mode | HARNESSTRIM_MODE | active | active reduces in place; dryrun logs only; off disables. | | minLength | HARNESSTRIM_MIN_LENGTH | 400 | Tool outputs shorter than this (chars) are left untouched. | | debug | HARNESSTRIM_DEBUG | false | Emit one-line [harnesstrim] diagnostics to stderr. | | compactionHandoff | — | true | Inject handoff guidance on compaction. | | telemetry | HARNESSTRIM_TELEMETRY | false | Append a TrimEvent JSONL record per reduction. | | telemetryPath | HARNESSTRIM_TELEMETRY_PATH | .harnesstrim/metrics.jsonl | Where telemetry JSONL is appended. |

Telemetry is off by default. When on, each reduction appends one JSON line; read the aggregate with harnesstrim metrics <path>. In dryrun mode telemetry still records what would be reduced, so you can size the win before switching to active.

Start with mode: "dryrun" and debug: true to see what would be reduced without changing model input, then switch to active once you're comfortable.

Runtime notes

• OpenCode plugins run on Bun; this package targets Bun's runtime (it imports TypeScript directly). No build step.
• Reduction is measured at runtime by character count (a cheap proxy) — no tokenizer is bundled into the harness process. Token-accurate reduction numbers live in the repo's benchmarks/.