Running PAI on anything other than Claude Code

[Rewdog]

I’ve been running one of the first PAI releases with opencode.ai, gradually merging in new features. At first, it felt like the project was heading toward portability — PAI_DIRECTORY and similar moves suggested we could use it with different clients (Claude, Codex, Opencode, etc.).

Recent updates suggest otherwise:

• v0.5.0 moved heavily into Claude skills.
• v0.6.0 added the .claude/ structure and now assumes a Claude Code environment.

That’s fine if Claude is the official direction, but I’d like clarity before I stop adapting it for opencode.ai and potentially add yet another paid subscription.

Questions:

1. Is PAI officially intended to stay Claude Code–specific going forward?
2. Is anyone maintaining a client-agnostic fork that preserves the early modular design?

Totally fine if the project is prioritizing efficiency using Claude vs flexibility. I really appreciate your effort @danielmiessler in finding a way to open source your system— just want to know whether I’m fighting the intended direction or if portability is still part of the plan.

[smolcompute]

It's still early days. Can't speak for Daniel obviously, but I think the introduction of skills made compatibility difficult until the other CLI tools catch up. Anthropic also seems to be well alined with some of PAI's principles, like progressive context discovery via the file system.

Given the speed at which all these tools are evolving, it doesn't make much sense to expect a definitive direction. Everything is exploration right now.

[danielmiessler]

Great question here.

PAI is going to continue to push toward being platform agnostic.

The reason you're seeing more of a push towards Claude Code is for a couple of reasons:

1. It's the most advanced agentic AI platform out there
2. It's the platform most people are using
3. MOST IMPORTANTLY: Even the Claude Code-leaning components are still surface-level, e.g., the name of the container directory, the use of "Skills" which is an Anthropic-branded thing

What makes me ok with moving to Skills is that it's all just Markdown + Code.

If you point any agent in the Gemini or Codex ecosystem at a Skill directory, it has most everything it needs to navigate the file structure, process the instructions, load the context, and do 95% of what Claude Code would have done with it. The main difference is just the default loading of the front matter as system prompt, but the file structure of frontmatter with description and system prompt and then the body with instructions, context, and code is universal.

So it's a tradeoff between maximally useful to one particular platform while still providing structure and content that can be used by any of them, and I feel like we're hitting that balance pretty well right now.

That being said, it's something I'm paying close attention to.

If Anthropic moves away from the Markdown paradigm, or if Codex / Gemini suddenly becomes overwhelmingly better, we will adjust.

But the one thing we'll always do is try to make the core components of a personal AI system as transparent and usable as possible, regardless of the platform being used.

[Steffen025]

Update: PAI-OpenCode Migration Progress & Feature Gap Discovery

I've been actively working on migrating PAI 2.0 to OpenCode and wanted to share a concrete finding from our v0.4 milestone (Agent Delegation).

What's Working

The core migration is progressing well:

• Skills translation ✅ (v0.3 complete)
• Agent delegation via Task wrapper (v0.4 in progress)
• All 7 core agents (intern, engineer, architect, researcher, designer, pentester, writer) being migrated

Feature Gap: Agent-level Resume

We discovered that OpenCode doesn't support resuming a specific subagent with its previous context:

PAI has:

// Resume agent with full context preserved
Task({ resume: "agent-id", prompt: "Continue..." })

OpenCode has:

• Session resume (/sessions) for entire conversations
• Session hierarchy for parent/child subagent work
• But NO agent-level resume by ID

Impact

Multi-phase workflows (like SpecFirst: Specify → Clarify → Analyze → Apply) lose context between phases. Each Task call starts fresh.

Next Steps

We've:

1. Deferred this to v0.5 of our migration
2. Filed a feature request at sst/opencode: [FEATURE]: Agent-level resume for subagents (resumeSessionId parameter) anomalyco/opencode#6584
3. Documented the technical gap in detail

Would love to hear if anyone else has tackled this or has workaround ideas!

---

From: PAI-OpenCode v0.4 Agent Delegation specification