Multi-Tool Parity

RDF emits governance, commands, and skills for several AI coding tools from one canonical source. Reach (Wave 2) recasts the tool set around three first-class citizens — Claude Code, Codex CLI, and Antigravity CLI — plus one frozen legacy tier, Gemini CLI (enterprise). Design rationale lives in docs/specs/2026-07-15-scale-reach-design.md §13.

1. Matrix

Tool	Command surface	Skill surface (.agents/skills/)	Context file	Hooks	Generate target
Claude Code	.claude/skills/<cmd>/SKILL.md + intent description: frontmatter (same adp_emit_skills library emitter as .agents/skills/, own output tree)	own skills/ tree, not the shared one	CLAUDE.md (NOT AGENTS.md)	hooks.json (manual merge)	claude-code
Codex CLI	.agents/skills/<cmd>/SKILL.md (native scan)	shared .agents/skills/	AGENTS.md (Agent Skills + AGENTS.md composite)	deferred (openai.yaml, §13.7)	codex
Antigravity CLI	skills (fuzzy-matched slash)	shared .agents/skills/	AGENTS.md + GEMINI.md	deferred (.agents/hooks.json, §13.7)	antigravity (composite)
Gemini CLI (enterprise)	.gemini/commands/*.toml	via agy plugin import gemini	GEMINI.md	—	gemini-cli

Tiers: the first three rows are first-class; Gemini CLI is legacy / frozen (kept for enterprise paid-API users, changed only for the TOML-escaping fix).

Claude Code — plugin install vs. symlink deploy. The Claude Code row above describes the symlink deploy (rdf deploy claude-code), which also populates the ~/.rdf/state/ session-state helpers. A plugin install (/plugin install rdf@rdf) ships commands and agents plus a SessionStart hook that bootstraps the same ~/.rdf/state/ helpers on the plugin’s first session start after install — so the 10 state-backed commands (r-spec, r-plan, r-build, r-ship, r-status, r-save, r-refresh, r-context-audit, r-vpe, r-util-mem-compact) run in a degraded mode only before that first restart. r-start degrades gracefully (probe + fallback); the others require the helpers for full session-scoped state.

2. The .agents/skills/ shared convention

.agents/skills/ is a workspace-level shared directory at the repo root — one tree that Codex, Antigravity, and every AAIF-compliant client read. Because it is shared, it is emitted exactly once by a dedicated adapter:

rdf generate agent-skills     # emits .agents/skills/<cmd>/SKILL.md

It is NOT duplicated into per-tool output trees. Each emitted SKILL.md carries name + description frontmatter only (name equals the command basename and the parent directory name, per the AAIF constraint); optional AAIF fields (license, metadata, allowed-tools) are deliberately not emitted (§13.4).

Deploy targets a workspace root:

rdf deploy --project-root /path/to/workspace agent-skills

This symlinks .agents/skills/ into the target workspace so rdf generate agent-skills updates deployed skills in place. agent-skills is the one project-scoped deploy target whose --project-root defaults to the current directory; agents-md, codex, and antigravity require it.

2b. Composites and AGENTS.md

codex and antigravity are the same composite on both verbs — the .agents/skills/ symlink plus AGENTS.md (copy-skip: an existing file is left alone unless --force):

rdf generate --project-root /path/to/proj antigravity   # optional here
rdf deploy   --project-root /path/to/proj antigravity   # required here

agents-md is also a standalone generate/deploy target. Generating it with a --project-root composes <root>/AGENTS.md from that repo’s own CLAUDE.md — falling back to its .rdf/governance/index.md, then a stub — and refuses a root that is not a git repository. Without --project-root, generate refreshes this checkout’s tracked adapters/agents-md/output/AGENTS.md: RDF’s self case, and the file that rdf deploy agents-md copies into a project root. rdf generate accepts its flags and target in any order.

rdf init --tools codex|antigravity|agent-skills|agents-md does both steps at init time — the composites expand to agent-skills agents-md, so skills are symlinked into the new project and its AGENTS.md is composed from the CLAUDE.md that init just wrote (spec docs/specs/2026-09-02-skills-native-adapter-consolidation-design.md §5).

Bounded surface. The shared skills tree covers the 10 lifecycle commands — /r-spec, /r-plan, /r-build, /r-ship, /r-start, /r-save, /r-status, /r-audit, /r-refresh, /r-init — not all 37 RDF commands (source of truth: adapters/agent-skills/skill-meta.json). Utility commands stay Claude-Code-only for now.

3. Gemini `` lossy edge

Gemini CLI command TOML uses token substitution. RDF command bodies are **not translated**: the canonical body is emitted verbatim into the prompt, so a body that reads `$ARGUMENTS` keeps that literal token — Gemini performs no substitution on it. What RDF adds is a **NOTE** comment (emitted only when the body references `$ARGUMENTS`) documenting that Gemini exposes just (the whole invocation string) and does not tokenize positional forms: a command invoked as /r-build 3 receives the raw 3, unbound to any placeholder. The body’s $ARGUMENTS therefore stays unsubstituted on the Gemini surface — an accepted limitation of the frozen legacy tier. The NOTE flags it so a Gemini CLI user knows to pass the argument through the `` form instead.

4. Legacy gemini-cli tier

Gemini CLI stopped serving free/Pro/Ultra tiers on 2026-06-18; enterprise and paid-API users are unaffected and the OSS repo stays active. RDF keeps the gemini-cli adapter frozen for those users: its GEMINI.md context and command TOML are unchanged except the one TOML-escaping fix.

• TOML '''-literal fix. The canonical command body contains characters that basic TOML strings ("...", """...""") interpret as escapes (regex \b, sed \|). Emitting those bodies into '''-literal multi-line strings — which do not process backslash escapes — makes the output parse strictly.
• Migration source. agy plugin import gemini converts .gemini/commands/*.toml INTO Antigravity skills, so strictly-valid TOML is the migration source. Fixing the escaping directly serves Antigravity adoption — a frozen legacy artifact that doubles as the transition input.

5. Deferred surfaces (probe-gated)

The following ship only after a live-docs probe confirms a stable schema, per the audit lesson that runtime-observed facts get an engineer-validation pass before code. Cross-reference spec §13.7:

• Antigravity hooks (<workspace>/.agents/hooks.json) — schema undocumented.
• Antigravity subagents (markdown) — schema actively churning.
• Antigravity plugins (~/.gemini/antigravity-cli/plugins/<name>/plugin.json).
• Codex agents/openai.yaml per-skill metadata — Codex-specific, optional.
• Global/user-level ~/.agents/skills/ scanning — unverified.
• SKILL.md optional AAIF fields (license, metadata, allowed-tools).
• MCP server work — a standing §3 non-goal.