OpenCode Config Bundle

A collection of OpenCode AI agent configuration presets with planning-first multi-tier workflow support.

Overview

This bundle provides drop-in OpenCode agent configurations that route work through schema-validated JSON handoff artifacts before planning, execution, and review. The goal is fewer ambiguous changes, less rework, and tighter safety boundaries with a small, explicit contract between agents.

Contract Ownership

This bundle implements the V2 bundle contract defined by the opencode-agents repository. The contract schema and specification are maintained in the CLI repository.

For the full contract specification, see Bundle Manifest Reference in the opencode-agents documentation.

Versioning

This bundle follows its own versioning scheme (bundle_version in the manifest), independent of the opencode-helper CLI version:

• Semantic Versioning: Follows semver (e.g., v1.0.0, v1.1.0, v2.0.0)
• Contract Compliance: Each bundle version declares which manifest_version it complies with

Required Files

A valid V2 bundle must include:

<bundle-root>/
  opencode-bundle.manifest.json  <- contract compliance marker
  <preset-entrypoint>.json       <- preset configurations
  .opencode/schemas/
    handoff.schema.json          <- from CLI contract
    result.schema.json           <- from CLI contract

Bundle Contents

Preset	Description
openai	OpenAI-based multi-tier agent configuration (GPT-5 series)
mixed	Mixed model stack (Claude for routing/planning/review, Codex for execution)
kimi	Kimi-based multi-tier agent configuration
big-pickle	Big Pickle model-based configuration
minimax	MiniMax-based configuration

Usage with opencode-helper

Register this bundle as a config source:

opencode-helper source add qbicsoftware/opencode-config-bundle --name qbic

Apply a preset to your project:

opencode-helper bundle apply qbic --preset openai --project-root ./myproject

Design Philosophy

Planning-First Execution

The core insight behind these configurations is that unplanned implementation is expensive to undo. Before any file is touched, the system asks: is this task concrete and scoped enough to implement directly? If not, a dedicated planning agent runs first.

Model Tier Strategy

Four model tiers are used, selected on the principle: use the cheapest model that can do the job correctly.

• Standard (claude-sonnet-4-6 / gpt-5.4): Planning, routing decisions, review
• Fast (claude-haiku-4-5 / gpt-5.2): Cheap routing, narrow doc edits
• Mini (gpt-5.1-codex-mini): Trivial and localized code edits
• Codex (gpt-5.3-codex / big-pickle): Primary implementation execution

Agent Architecture

All configurations define four functional tiers:

1. Routing Agents - Entry points that classify and delegate work
2. Planning Agents - Produce structured execution plans
3. Execution Agents - Implement changes based on plans
4. Validation Agents - Review quality and safety

License

AGPL-3.0 - see the LICENSE file for details.