JSPM

  • Created
  • Published
  • Downloads 2124
  • Score
    100M100P100Q141719F
  • License MIT

Opinionated OpenCode agent harness — PRIME, plan, build, QA, skills, MCP wiring, hashline editing.

Package Exports

  • @glrs-dev/harness-plugin-opencode
  • @glrs-dev/harness-plugin-opencode/cli

Readme

@glrs-dev/harness-plugin-opencode

OpenCode agent harness. 27 agents, 7 commands, 5 tools, 18 skills.

Docs: glrs.dev/harness

Install

curl -fsSL https://glrs.dev/install.sh | bash

Or: npm i -g @glrs-dev/cli && glrs harness install && opencode

Commands

/fresh ENG-1234              # branch + start [SPEAR](https://www.edge.ceo/p/introducing-spear-the-management) workflow
/fresh add rate limiting     # same, from description
/ship                        # squash, push, open PR
/review 87                   # adversarial code review
/research how does auth work # parallel codebase search
/costs                       # LLM spend
/dispatches                  # subagent history

Autopilot

glrs loop "implement the auth middleware"

Agents

Agent Tier What it does
prime mid SPEAR end-to-end — Sonnet orchestrator, delegates hard work to Opus (default)
prime-heavy deep PRIME on Opus — use for heavyweight orchestration
plan deep Planner with gap analysis
build mid Plan executor
research deep Parallel codebase research

Plus 22 subagents, autopilot variants, and cost-optimized tiers. Full list at glrs.dev/harness/agents.

Configuration

Model overrides, MCP servers, env vars: glrs.dev/harness/config

Telemetry

The harness sends anonymous usage events via Counted to help prioritize work: per-model token speed and cost (model_turn), tool and skill usage with best-effort success (tool_used), and post-edit type-check results (post_edit_verify). No cookies, no fingerprinting, no PII — never repo names, branch names, paths, prompts, or arguments; properties are public model/provider ids, enums, booleans, and counts only. Tracking never blocks or breaks a session and a dead network can never delay it.

Opt out with either:

export DO_NOT_TRACK=1        # the cross-tool Do Not Track standard
export GLRS_NO_ANALYTICS=1   # glrs-specific

Security

Report vulnerabilities per SECURITY.md. Not a sandbox — treat the agent like a dev with shell access.

License

MIT