Trellis v0.5.0: AI Code Orchestration Gets a Complete Architectural Overhaul

Trellis v0.5.0 Beta is live. This is the largest architecture overhaul since 0.4.0 GA. In short: AI orchestration has shifted from “you tell it what to do” to “it sees the context and knows what to do.”

For developers who have been using Trellis to manage AI coding workflows, this upgrade is significant enough to make you reassess your workflow configurations.

Starting with the most important change: Skill-first architecture

In the 0.4.x era, Trellis operated in a command-driven way—you had to manually input commands like /trellis:before-dev or /trellis:brainstorm to trigger specific orchestration behaviors. That worked, but essentially remained “human-driven AI.”

v0.5.0 flips that logic around.

Apart from the three core flow commands /start, /continue, and /finish-work, all other Trellis commands have been refactored into auto-triggered skills. What does that mean? While working, the AI will automatically activate the corresponding skill whenever it sees a suitable context—no need to manually type commands.

The five commands converted are:

• before-dev — Pre-development checks
• brainstorm — Solution brainstorming
• break-loop — Escape infinite loops
• check — Code review
• update-spec — Update specification documents

For example: previously you had to manually say “check the requirements for me” before coding; now when the AI sees you open a new task, it knows to run before-dev. If it notices you revising the same issue multiple times, it auto-triggers break-loop to help you change strategies.

Of course, if you prefer to control every step yourself, skills can still be invoked manually. This isn’t an either/or design—it defaults to automation, with manual control as an option.

This shift in design philosophy deserves emphasis. The mainstream thinking for AI coding tools is still “human issues commands, AI executes.” Trellis aims to take a step forward—allowing the AI, once equipped with contextual awareness, to autonomously decide what to do and when. This aligns with Cursor’s Agent mode and Copilot’s multi-file edit approach, but Trellis operates at a higher orchestration layer, not at the single code-generation level.

File structure migration

With the architecture changed, the file organization has naturally evolved too. Files previously stored under commands/<name>.md are now moved to skills/trellis-<name>/SKILL.md.

There are 65 rename migrations that will run automatically during the upgrade. If you’ve modified files locally (many advanced users customize prompt templates), the migration script will prompt you for each change—whether to preserve your custom version or overwrite it with the new template.

Additionally, command and skill templates across all platforms are now merged into a single source:

packages/cli/src/templates/common/
├── commands/   (3 core commands)
└── skills/     (5 auto-triggered skills)

This resolves a long-standing issue—Trellis previously supported multiple platforms, each maintaining its own templates, often causing drift (“Platform A updated while Platform B didn’t”). Now, with a single source plus a cross-platform output adapter, one update syncs across all platforms.

For maintainers, this reduces workload substantially; for users, it ensures consistent experience across platforms.

7 platforms upgraded to Agent-capable status

This is another major highlight in v0.5.0. The following seven platforms have been upgraded from “command-only” to fully agent-capable:

• Qoder
• CodeBuddy
• Factory Droid
• Cursor
• Gemini CLI
• Kiro
• GitHub Copilot

What exactly does “Agent-capable” mean? Two things:

First, sub-agent definition. Each platform can now generate three types of sub-agents:

• implement — Responsible for actual code implementation
• check — Responsible for code review and verification
• research — Responsible for information gathering and research

These sub-agents are generated in each platform’s native format. For instance, Cursor uses its recognizable agent configuration; Copilot uses its own format. You don’t need to worry about the differences—Trellis smooths them out.

Second, the hook system. Implemented with Python scripts under the shared-hooks/ directory, it includes three core hooks:

• session-start — Initialization on session start
• inject-subagent-context — Inject context into sub-agents
• statusline — Update status bar information

The pattern—single implementation + cross-platform output adapter—applies the same idea as template merging.

Notably, community feedback confirms that Codex sub-agents are also supported, meaning Trellis’s coverage at the agent orchestration layer is expanding rapidly.

What problem does this actually solve?

Stepping back, what Trellis is building is the “orchestration layer” for AI coding tools.

Current AI coding tools—Cursor, Copilot, Kiro, Gemini CLI—each have their own agent capabilities, but they are isolated. A workflow you’ve set up in Cursor has to be recreated in Copilot. Best practices you’ve refined in one project don’t carry over to another.

Trellis aims to provide a unified orchestration layer above these tools—define your workflow once, run it anywhere. The Skill-first + Sub-agent combo in v0.5.0 effectively evolves the orchestration layer from “passive response” to “active collaboration.”

It’s somewhat like Terraform for cloud platforms—you don’t manage infrastructure directly, but through an abstraction layer. The difference is that Trellis manages AI coding workflows, not infrastructure.

Upgrade notes

To install or upgrade:

# Fresh install
npm install -g @mindfoldhq/trellis@beta

# Upgrade from 0.4.x (must include --migrate)
trellis update --migrate

--migrate is mandatory this time; skipping it will throw an error. The reason is simple—the file structure has changed, and old configurations won’t run without migration.

Possible upgrade scenarios:

1. If you’ve customized templates under the commands directory, the migration script will ask how to handle them.
2. The 65 rename migrations will run automatically; manual intervention is usually unnecessary.
3. Once migration completes, it’s recommended to run trellis check to verify configuration integrity.

Some in the community noted that documentation hasn’t been fully updated for v0.5.0 yet. Understandable for a beta, but it’s something the team should catch up on soon. Developers who rely on docs may need to combine changelog and community discussions to explore new features for now.

Some thoughts

My take:

The Skill-first direction is right. The trend in AI coding tools is shifting from “tool” to “collaborator”—and a collaborator shouldn’t wait for explicit orders every time. Having AI autonomously decide what to do based on context is product intuition done right.

But beta-stage risks are obvious: how accurate are auto-triggers? Could there be false activations—like “AI thinks it should brainstorm” when you’re just fixing a typo? Only extensive real-world use will surface these edge cases.

The addition of sub-agent support is another highlight. A pain point in current AI coding setups is the single-agent ceiling—one agent handling coding, reviewing, and research quickly hits context window limits. Splitting these into specialized sub-agents collaborating together could theoretically overcome this. Doing it at the orchestration layer, as Trellis does, is far more efficient than each platform implementing separately.

The coverage across seven platforms shows Trellis’s strategy clearly—it doesn’t bind to any one AI coding tool but operates as an abstraction above them. At a time when AI coding tools are flourishing, that’s valuable—no one knows which tool will ultimately prevail, but the orchestration layer hedges against that uncertainty.

There’s already talk in the community about wrapping Trellis as a Cursor Composer plugin. That’s a reasonable idea—embedding Trellis as a plugin rather than a standalone CLI would lower barriers to use. But it would also require deep integration with each platform’s plugin ecosystem—a nontrivial engineering effort.

Comparison with similar tools

There aren’t many tools focusing on AI coding orchestration yet. Current comparisons include:

• Aider — An AI coding assistant focused on Git integration, but not cross-platform orchestration
• Claude Code /project system — Project-level context management, but locked to the Claude ecosystem
• Cursor Rules — An internal rule system that works only within Cursor

Trellis’s differentiation lies in its “cross-platform + orchestration layer” combination. If you only use one AI coding tool, Trellis’s value is limited; but if you switch between tools—or have a team using different tools—Trellis’s unified orchestration becomes very meaningful.

Summary

Trellis v0.5.0 Beta is an ambitious architectural rewrite. Skill-first makes orchestration proactive, Sub-agent breaks the single-agent capability ceiling, and seven-platform coverage enables cross-tool collaboration.

As a beta, rough edges are inevitable—outdated documentation, auto-trigger boundary cases, migration compatibility issues—but the direction is clear and the architecture clean.

If you’re using multiple AI coding tools or care about workflow automation for AI coding, it’s worth trying now. Feedback during the beta stage is the most influential in shaping the product’s future.

---

References

• Trellis v0.5.0 Beta Release Discussion — Original thread on the Linux.do community, including full update notes and feedback