Every agent

Reference. Procoder works with any AI coding agent, not just Claude Code. One canonical AGENTS.md carries the always-on contract; every host gets the thinnest adapter that serves it — a rule-file copy, a manifest pointing at the existing files, or a hook. procoder agents prints anything missing or drifted, and drift blocks the gate like any other mirror.

The adapter rule: adapters stay thin. Logic lives in the binary; content lives in AGENTS.md and commands/; an adapter only points or copies.

Instruction tier — a rules file, nothing else

Host	File	Install
Zed, Amp, Jules, Swival, VS Code agents…	AGENTS.md	nothing — read natively
Cursor	.cursor/rules/procoder.mdc	copy into your repo
Windsurf	.windsurf/rules/procoder.md	copy into your repo
Cline	.clinerules/procoder.md	copy into your repo
Kilo Code	.kilocode/rules/procoder.md	copy into your repo
Roo Code	.roo/rules/procoder.md	copy into your repo
Kiro	.kiro/steering/procoder.md	copy into your repo
Antigravity	.agents/rules/procoder.md	copy into your repo
Qoder	.qoder/rules/procoder.md	copy into your repo
Copilot in editors	.github/copilot-instructions.md	copy into your repo
OpenAI Codex (repo docs)	.codex/AGENTS.md	copy into your repo

Every copy is byte-pinned to AGENTS.md (host frontmatter aside) by the gate; edit the master, then procoder agents prints the refreshed copies.

Plugin tier — manifests and hooks

Host	Entry	Notes
Claude Code	.claude-plugin/plugin.json	the reference host — fully tested
Codex CLI	.codex-plugin/plugin.json	shares Claude's hooks file; the binary answers in Codex's JSON shape by host detection
GitHub Copilot CLI	.github/plugin/plugin.json	own hook schema (hooks/copilot-hooks.json, bash+powershell); session-start injection only
Gemini CLI / Antigravity	gemini-extension.json	contextFileName: AGENTS.md; deliberately no root hooks/hooks.json (Gemini would auto-load it with incompatible events — its absence is enforced by the gate)
OpenCode	opencode.json → .opencode/plugins/procoder.mjs	JS shim injects AGENTS.md per turn; .opencode/command/*.md are generated twins of commands/*.md (parity pinned by test)
Grok Build	root plugin.json	skills only, no hooks
Devin CLI	.devin-plugin/plugin.json	metadata + skills
Qoder (plugin tier)	.qoder-plugin/plugin.json	skills + rules pointer
pi	package.json pi block → pi-extension/index.js	injects AGENTS.md at agent start
Hermes Agent	plugin.yaml + __init__.py	pre_llm_call hook injects AGENTS.md

Host detection lives in the binary (internal/host): COPILOT_PLUGIN_DATA (or a VS Code plugin-root path) → Copilot, PLUGIN_DATA → Codex, QODER_SESSION_ID → Qoder, else Claude. procoder principles --hook answers in each host's session-start shape.

Test coverage per host

Claude Code is the reference host — every command, hook, and skill is exercised there daily. The other adapters follow each host's published plugin shape (largely proven in the wild by the ponytail plugin's portability layer, which this design follows) and are not continuously tested against live installs. If one misbehaves on your host, that is a bug — report it with the host name and the adapter file.

Manifest versions are pinned to .claude-plugin/plugin.json by the gate, so a release can never leave one host stale.

The everywhere-binary

All tiers assume the procoder binary is reachable. Claude Code gets it via the plugin automatically; for every other host, put the right dist/<platform>/procoder on PATH (or go install). The instruction tier degrades gracefully: without the binary the rules still shape behaviour, and the agent is told what to install.