oc-dw — OpenCode Dynamic Workflows

Claude Dynamic Workflows, but for OpenCode. Install once, use inside OpenCode forever.

oc-dw brings multi-agent workflow orchestration directly into OpenCode. No separate CLI to run, no servers to manage, no dashboard to open. Just tell OpenCode what you want done and it plans, fans out, verifies, and synthesizes — exactly like Claude Code's built-in dynamic workflows.

Install

npm i -g oc-dw

Setup

One command wires everything into OpenCode:

oc-dw setup

This does three things automatically:

1. Adds "oc-dw" to your opencode.json plugin list
2. Installs the /workflow custom command
3. Installs the /ultrawork custom command

Restart OpenCode. Done.

Usage

Inside OpenCode, just ask:

/workflow Refactor the auth layer to use JWT tokens instead of sessions

Or for fire-and-forget mode that runs until completion:

/ultrawork Add comprehensive test coverage for the payment module

That's it. The agent calls dynamic_workflow_run behind the scenes, which:

1. Plans — Generates phases, tasks, dependencies, and model routing
2. Fans out — Spawns isolated worker sessions with scoped prompts
3. Verifies — Independent verifiers check results against acceptance criteria
4. Synthesizes — Merges all outputs into a final report
5. Checkpoints — File-backed state so workflows are resumable

Artifacts are stored under .opencode/dynamic-workflows/runs/<workflow-id>/.

What it does

Feature	What it means
Planning	LLM generates a structured plan with phases, tasks, dependencies, model roles, and quality gates
DAG execution	Tasks run in dependency order with bounded concurrency
Multi-round verification	Independent verifier agents check every result
Adversarial review	Critic and adversary agents debate results with convergence detection
Skill constraints	Apply reusable rules like test-driven, security-first, strict-types to all workers
Templates	Built-in templates for deep-research, codebase-audit, large-migration, test-generation, documentation-update
Resumable	File-backed checkpoints keep state across restarts
Model-agnostic	Route any role to any OpenCode provider/model

Customizing workflows

The agent picks defaults automatically, but you can be explicit:

Run a dynamic workflow with:
- template: deep-research
- skill: security-first
- skill: test-driven
- adversarial_review: true
- effort: ultra
- stopping_condition: "All findings are cross-referenced with source code and documented in FINDINGS.md"

Available templates:

• deep-research — Multi-angle investigation with cross-checking
• codebase-audit — Systematic security/quality audit
• large-migration — Framework/language migration with parity verification
• test-generation — Comprehensive test generation with edge cases
• documentation-update — Doc update/generation for a codebase

Available skills:

• no-casts — Prohibit type assertions or any
• test-driven — Require tests for every change
• minimal-diff — Smallest possible change
• strict-types — Enforce TypeScript strict mode
• security-first — Evaluate all changes for security
• performance-aware — Consider complexity and I/O efficiency
• docs-required — Document all public APIs
• backward-compat — Maintain backward compatibility

Architecture

• Planner — Uses structured output to create the workflow plan
• Workers — Each task runs in an isolated OpenCode session with scoped context
• Verifiers — Independent agents check against acceptance criteria
• Adversaries — Cross-check findings with convergence detection
• Synthesizer — Chunks and merges results into a final report
• State store — File-backed for resumability and inspection

Development

npm install
npm run typecheck
npm run build
npm test

License

MIT