Runtime Adapters

Claude Code is the primary runtime. Codex, OpenCode, and Pi receive shared skills and instructions from one devkit release through optional adapters; this page says how to switch them on and what they write.

Activate

bash

devkit config codex auto|on|off       # default auto: active while `codex` is installed
devkit config opencode auto|on|off    # default off
devkit config pi auto|on|off          # default off

A mode change takes effect on the next devkit apply or devkit sync. devkit config show prints the modes; devkit doctor checks the written files.

What each adapter writes

Runtime	Shared source	Form	Destination
Codex	Portable skills	od-<name> skills	~/.agents/skills/
Codex	Agents	TOML definitions	~/.codex/agents/
Codex	Shared and Codex guard hooks	Native hook config	~/.codex/hooks.json
Codex	Guardrails	Rule files	~/.codex/rules/
Codex	Global instructions	Managed block	~/.codex/AGENTS.md
OpenCode	Tool access	tools, permission and subagent_depth keys, added only, never removed; permission."*": "allow" sits first, so opencode starts without permission prompts and a personal ask/deny still wins	~/.config/opencode/opencode.json
OpenCode	Global instructions	Managed block	~/.config/opencode/AGENTS.md
Pi	Global instructions	Managed block marked od managed (onedot-devkit pi)	~/.pi/agent/AGENTS.md
Pi	Shared hooks	Compiled event, matcher, command manifest	~/.pi/agent/extensions/devkit-hooks.manifest.json
Pi	Hook bridge	Managed TypeScript adapter	~/.pi/agent/extensions/devkit-hooks.ts

The adapters leave same-named personal entries and everything outside a managed block untouched; keeping a personal rule is reported as a warning, not a failure.

Pi loads devkit-hooks.ts as a native extension and reads the compiled manifest at session start. Delegates keep --no-extensions; write-capable Pi delegates load only Devkit's hook bridge, and a missing bridge refuses the dispatch. Turning Pi off removes its managed hook paths and, when Codex is disabled, the shared skill links Pi needed. Unlinking devkit removes only devkit-owned links.

Skill names

Devkit materializes shared skills as od-<name> under ~/.agents/skills/ when Codex or Pi is enabled. Pi discovers those directories under the same od- names without a second mirror; disabling Codex does not remove Pi's skills. OpenCode uses its native shared-skill discovery. devkit path skill <name> stays prefix-free in every runtime; devkit run starts shared scripts.

Pi MCP connections are configured separately through pi-mcp-adapter; this instructions adapter does not provision them. An imports: ["codex"] entry in Pi's adapter config reuses Codex MCP server definitions without starting Codex. Set lifecycle: "eager" (or "keep-alive" for a persistent remote server) in a Pi adapter override when those servers must connect at session start.

Models and permissions

Each agent gets the same model class as in Claude Code, mapped to the runtime's model names. Read-only agents cannot write; agents that edit get write and shell access.

/delegate can send review, research or implementation to Codex or another installed runtime; see initiative. devkit path resume [project-dir] resolves Codex's pre-compact snapshot for a project.

OpenCode prerequisites

A shared skill in OpenCode needs the built-ins read, glob, grep, list and bash. Without them the agent still answers, but as a chatbot: no repository search, no devkit run, and /friction-feedback cannot even report it. lsp and websearch are optional.

OpenCode removes a built-in from the toolset when any merged config layer denies it — permission.read: "deny", or tools.read: false without a permission override. Layers merge global → OPENCODE_CONFIG → project opencode.json/opencode.jsonc → .opencode/; a global deny stays in force until that file changes or a project layer sets the same key to "allow".

Two more places hide a deny: a per-agent permission block overrides the top level, so agent.explore.permission.bash: "deny" blinds that subagent even when the top level allows it, and experimental.primary_tools lists tools that only primary agents keep — a built-in named there is denied to every subagent.

devkit apply writes this into ~/.config/opencode/opencode.json and leaves a differing personal value alone:

json

{
  "$schema": "https://opencode.ai/config.json",
  "subagent_depth": 2,
  "permission": { "read": "allow", "glob": "allow", "grep": "allow", "list": "allow", "bash": "allow" }
}

bash can stay granular ("bash": { "git *": "allow", "*": "ask" }); only an all-deny rule counts as missing. subagent_depth is a top-level key (not under experimental): shared skills delegate to subagents that delegate again, and OpenCode's default of 1 stops the second hop with Subagent depth limit reached (1).

bash

devkit doctor --target opencode   # only OpenCode checks, exit 1 on a finding
devkit doctor --target pi         # only Pi managed-block and hook-bridge checks

The report names each missing built-in, the file that denies it, each agent with its own deny, every required tool in experimental.primary_tools, and a subagent_depth below 2. The same checks run inside a plain devkit doctor whenever the adapter is on or a project ships its own opencode.json.

OpenCode reads config, agents, skills and plugins once at start; restart it after every change. If a broken config prevents the start: OPENCODE_DISABLE_PROJECT_CONFIG=1 skips the project layer, OPENCODE_CONFIG=/path/file.json loads one more file, OPENCODE_CONFIG_CONTENT='{"$schema":"https://opencode.ai/config.json"}' injects inline JSON.