OpenCode is fast. Shipping with discipline is harder.
SkyFlow OpenCode turns OpenCode into a pipeline-first engineering workflow with explicit planning, approval gates, verification checkpoints, and risk-based review depth.
SkyFlow OpenCode is not just a collection of prompts.
It is a structured OpenCode preset for people who want their agent to behave more like a delivery-minded engineering lead:
- clarify requirements before implementation
- stop for approval before code changes
- verify with evidence before handoff
- increase review depth when risk increases
- keep the workflow reproducible across sessions
If you want more control than “just vibe-code until it works”, this preset is for you.
Tip
Good fit for: bug fixes, internal tools, production changes, risky refactors, and teams that want process without building a full custom harness.
| Area | Capability | What it gives you |
|---|---|---|
| 🧭 | Root Orchestrator | A primary agent that enforces mode selection, planning, and approval before action |
| 📋 | Pipeline Modes | MINI, FAST, BALANCED, COMPREHENSIVE, HARDENING depending on task risk |
| 🔍 | Tech Scout | Requirement clarification plus real codebase context before implementation |
| 🧪 | Red-First Testing | BALANCED and above can require executable failing tests before code changes |
| ✅ | Independent QA | Separate verification step for builds, tests, evidence, and coverage gaps |
| 🛡️ | Risk Gates | Optional security, performance, dependency, and release review stages |
| 🗂️ | Blackboard Workflow | A stateful task record that keeps long runs structured and auditable |
| 🌏 | Chinese-First UX | User-facing responses and reports are designed for Simplified Chinese workflows |
Most OpenCode setups optimize for raw speed.
SkyFlow OpenCode optimizes for controlled delivery.
That means:
- fewer ambiguous starts
- fewer silent assumption jumps
- fewer “looks done” handoffs
- more explicit gates
- more reusable execution patterns
It is opinionated on purpose.
The main agent should not immediately start coding.
It should first decide:
- what kind of task this is
- what execution mode it needs
- what evidence is required
- which sub-agent should act next
- whether the pipeline should continue or pause for approval
SkyFlow packages that behavior into a reusable preset.
SkyFlow OpenCode scales its process by task risk.
| Mode | Best for | What happens |
|---|---|---|
MINI |
tiny localized changes | lightweight scout → implementation → build check |
FAST |
ordinary delivery work | requirement doc + approval gate + implementation + QA |
BALANCED |
core bug fixes and feature work | adds red tests and post-green refactor review |
COMPREHENSIVE |
important flows and external interfaces | adds unified static compliance audit |
HARDENING |
auth, payments, sensitive data, high-risk code | adds security, performance, dependency, and release gates |
SkyFlow currently ships with the following agent roles:
main— root orchestrator, state machine, approval gatekeeperbuild— primary coding agent templatetech_scout— requirement clarification and context buildertest_red_author— creates executable failing tests firstimpl_green_coder— implements the minimum change required to go greenquality_assurance— independently verifies builds, tests, and evidencerefactor_reviewer— improves readability after green verificationcompliance_auditor— unified static audit for security, performance, and dependenciessecurity_reviewer— security-focused review gateperf_reviewer— performance-focused review gatedependency_guard— dependency and license risk gaterelease_gate— final go / no-go decision
SkyFlow OpenCode is a configuration pack / preset.
It is not only a runtime plugin, and it is not a theme.
This repository currently includes:
agents/— custom primary and sub-agent promptsopencode.json— provider and MCP defaultspackage.json— runtime dependency baselinecommands/,plugins/,skills/— extension directories for future expansion
This layout keeps the project easy to inspect, version, publish, and evolve.
SkyFlow-OpenCode/
agents/
commands/
plugins/
skills/
.env.example
opencode.json
package.json
README.md
README.zh-CN.md
Point OpenCode to this repository as the config directory:
export OPENCODE_CONFIG_DIR=/path/to/SkyFlow-OpenCode
opencodeOn Windows PowerShell:
$env:OPENCODE_CONFIG_DIR = "C:\path\to\SkyFlow-OpenCode"
opencodeCopy these into your OpenCode config root:
agents/opencode.json- optional future
commands/,plugins/,skills/
This repository intentionally avoids hardcoded secrets.
Required environment variables:
CCODEZH_API_KEYCONTEXT7_API_KEYTAVILY_API_KEY
Example:
export CCODEZH_API_KEY="your-key"
export CONTEXT7_API_KEY="your-key"
export TAVILY_API_KEY="your-key"You can use .env.example as a local reference.
Start tasks by explicitly selecting a mode:
Use BALANCED mode for this bug fix and show me the plan first.
Examples:
Use MINI mode to fix a tiny style issueUse FAST mode to add request validationUse BALANCED mode to fix a core logic bugUse HARDENING mode to review authentication changes
- plan before modification
- human approval before code changes
- minimal implementation over clever implementation
- verification before delivery
- review depth matched to task risk
- structured and consistent user-facing output
- never commit real API keys
- keep
opencode.jsonsafe to publish - prefer
{env:...}or{file:...}for credentials - review remote MCP settings before using this in shared or production-like environments
Note
SkyFlow OpenCode is intentionally strict. That is a feature, not an accident.
Repository name
skyflow-opencode
Short description
Pipeline-first OpenCode preset with orchestrated agents, QA gates, and risk-based review modes.
Suggested topics
opencode ai-agent multi-agent workflow qa prompt-engineering coding-agent
- add a one-command installer
- add built-in plugin behaviors for guardrails and hooks
- add reusable commands for mode shortcuts
- add skill packs for backend, frontend, and security-heavy tasks
- add publish-ready assets for GitHub and npm release flows
SkyFlow OpenCode is an opinionated preset for OpenCode. It does not replace engineering judgment. Review prompts, permissions, providers, and MCP access before using it in production or in shared repositories.