Wira Delta Indonesia
WDI MethodEngineering

WDI Method 0.6.32: One Method on 39 Agent Hosts

Until 0.6.32, WDI Method only worked properly on Claude Code. This release records what each of 39 hosts can do, and the method follows that record instead of assuming one host.

Wira Sihombing 5 min read

WDI Method is an open-source review layer for AI coding agents, installed on top of BMad Method. Its skills are folders with a SKILL.md file, a format many coding agents read. That made it easy to believe a repo with the skills in it would work on any host. It did not, and 0.6.32 is the release that fixes it.

What Worked Only on Claude Code

The installer and the validator each kept their own list of five skill folders, and every other host was treated as if it read one of them. update went through as long as the engines that wdi-build drives were somewhere in the repo, even when a selected host could not see them: Kiro, for one, reads only .kiro/skills. The validator also read engines placed under .kiro/skills as "no engines in this repo".

The rule that keeps the daily skills manual-only, the instructions for invoking an engine, and the autopilot loop were written the same way: for Claude Code, and assumed for the rest.

Three Differences Between Hosts

Reading the documentation of each host the installer offers turned up three differences that a shared folder format hides:

  • Where it reads skills and rules. Most hosts read .agents/skills, but Claude Code reads .claude/skills, Kiro .kiro/skills, CodeBuddy .codebuddy/skills, and QwenCoder .qwen/skills. Next to AGENTS.md, many read a rule file of their own: CLAUDE.md, GEMINI.md, QWEN.md, CODEBUDDY.md, .cursorrules, and others.
  • Whether it can hold a skill to manual-only. Ten hosts honor disable-model-invocation: true in a skill's frontmatter. On the rest, the model may start any skill it finds.
  • Whether it has a scheduler. Claude Code, Command Code, and Qoder have a /loop command. CodeBuddy, Cline, Goose, GitHub Copilot, Warp, and AdaL have a scheduler of their own. The other hosts have neither.

How a person types a skill differs as well: /name on most hosts, $name on Codex and Cortex, /skill:name on Kimi Code and Pi, and @name on Windsurf.

0.6.32 writes these facts into one record per host. install and update copy the selected hosts into .control/wdi-method.yaml, and the skills and validate.py read that file instead of assuming Claude Code. The two separate folder lists are gone.

Six hosts were removed because they could not be made to work: iFlow (shut down), Firebender (service ends on 2026-10-31), Roo Code (archived), Hermes Agent (reads only a skill folder in the user's home), Neovate (no documented project skill folder), and Mux (runs other agents and reads no skills itself). npx wdi-method --list-agents prints the 39 that remain.

Manual-Only in Three Layers

Five skills must run only when the owner types them: the four daily tier skills and wdi-explain-to-me. On a host that ignores disable-model-invocation, a model could start one of them on its own. 0.6.32 holds them in three layers:

  1. The host's own lock, where it has one: disable-model-invocation: true on the ten hosts that honor it, deny rules in .claude/settings.json on Claude Code, and "ask" under permission.skill in opencode.json on OpenCode.
  2. A guard line at the top of each of the five skills.
  3. A rule in the method block of AGENTS.md and of each host's own rule file.

On a host with no lock, the last two layers are the lock. They are instructions the model reads, not a mechanism that blocks it, so the docs say which hosts get which layer instead of promising the same guarantee everywhere. A repo without Claude Code no longer gets a .claude/ folder.

Engines and Schedulers per Host

wdi-build drives six engines from mattpocock/skills. Three of them ship with disable-model-invocation: true, and the installer removes that line from the copies in the repo so the method can run them. Before 0.6.32 this covered five folders. It now covers every .<host>/skills folder in the repo and every folder a supported host reads.

That made update stricter. It refuses when the engines are not in a folder every selected host reads, and names the fix for each host. When the engines are already in the repo but in the wrong folder, the installer can copy them: --copy-engines on install or update, or npx wdi-method engines --copy later. The copy comes from the repo's own files, so it needs no network and overwrites nothing.

An engine is invoked through the host's own skill mechanism: a skill tool where the host has one, and elsewhere by reading the engine's whole SKILL.md from the repo. A paraphrase from memory is still not allowed.

The autopilot follows the same record. On a host with /loop or a scheduler, wdi-daily-autopilot repeats every 10 minutes by default. On every other host it runs one iteration each time the owner types it, under the same mandate, ledger, and branch. It never falls back to a shell loop or an OS scheduler.

What Comes in 0.7.0 and 0.8.0

Part of the per-host work above exists because the engines and BMad Method are installed separately. The roadmap removes both prerequisites:

  • 0.7.0: the six engines ship inside wdi-build and wdi-blueprint as supporting files, adapted from a pinned upstream version. The disable-model-invocation patching, the per-host engine placement, and the "install the engines first" step go away.
  • 0.8.0: the BMad workflows the gate skills use are adapted into the wdi-* skills. Each one is checked against the BMad original on the same fixture before it replaces it.

Both are plans, not dates. Both upstream projects are MIT licensed. What WDI Method adapts from them says so file by file, and WDI Method is not affiliated with or endorsed by either.

To update an existing repo, run npx wdi-method@latest update. 0.6.32 needs no wdi-upgrade. If update refuses, it prints three fixes: --copy-engines, npx skills add … --agent <id>, or leaving an unused host out with --agents. Every change is listed in the changelog.