Configuration

Most projects need no configuration beyond rdf init and /r-init. Your rules live in plain files you can edit.

  1. Project files
  2. Convention inheritance
  3. Use other AI tools
  4. Environment variables
  5. Keep deployed files in sync
  6. Check your environment

Project files

Path Written by Purpose
CLAUDE.md rdf init Project working instructions, merged from your stack’s profiles. Excluded from commits by default
MEMORY.md rdf init Placeholder for session-persistent project facts
.rdf/governance/reference/ rdf init Reference docs from matching profiles
.rdf/governance/index.md /r-init Project identity, loaded first by agents
.rdf/governance/architecture.md /r-init How the project is laid out
.rdf/governance/conventions.md /r-init How code is written here
.rdf/governance/verification.md /r-init Lint and test commands that QA runs
.rdf/governance/constraints.md /r-init Hard limits, which also raise a phase’s gate scope
.rdf/governance/anti-patterns.md /r-init Patterns to reject, which can mark paths security-sensitive
.rdf/governance/ignore.md /r-init Paths agents should skip
.rdf/work-output/ Agents Session state and agent results
.git/info/exclude rdf init Keeps working files out of commits

After the code changes, run /r-refresh to update governance. It preserves your edits. /r-init --force deletes .rdf/governance/, including the reference docs rdf init copied there, and starts over.

Convention inheritance

The most specific rule wins:

project CLAUDE.md  >  workspace CLAUDE.md  >  profile defaults  >  core defaults

Agents read the chain at session start and apply it in that order.

Use other AI tools

rdf init targets Claude Code by default. Pass --tools LIST (comma-separated) to wire up other surfaces in the target project. An existing AGENTS.md is left untouched, and AGENTS.md composition requires a git repository.

Value Adds
claude-code The default; a no-op when combined with others
agent-skills Symlinks .agents/skills
agents-md Composes AGENTS.md from the project’s own CLAUDE.md
codex, antigravity Both of the above

See Multi-tool parity.

Environment variables

Most setups need none of these.

Variable Default Purpose
RDF_TARGET ~/.claude Deploy target for the claude-code adapter
RDF_ACCOUNT Unset Account for rdf run and rdf review when no @name is given. Set by rdf use, and overrides rdf account default. RDF_CLAUDE_ACCOUNT is still read as a fallback
RDF_WORKSPACE Parent of the RDF checkout Workspace root for rdf doctor --all. Set it when the checkout sits directly in $HOME, because --all refuses to scan a home directory
RDF_CLAUDE_PROJECTS $CLAUDE_CONFIG_DIR/projects, else ~/.claude/projects Transcript root for rdf tokens
CODEX_HOME ~/.codex Codex home read by rdf tokens
RDF_TOKENS_CACHE ~/.rdf/cache/tokens rdf tokens row cache
RDF_TOKENS_JOBS CPU count Parallel extraction workers

Keep deployed files in sync

Generated output is symlinked, so regenerating updates deployed files in place. A new command also needs rdf deploy, because each skill has its own symlink.

canonical edits  -->  rdf generate  -->  deployed (~/.claude/, ~/.gemini/, ...)
deployed edits   <--  rdf sync      <--  emergency path

Emergency edits to ~/.claude/ are allowed, but pull them back with rdf sync. rdf doctor --scope content-drift detects divergence using per-file .rdf-hash sidecars.

Check your environment

/r-env compares this host with what the project needs and prints exact fixes for your package manager. It changes nothing on its own, and offers to run the fixes you pick. It checks:

  • A code intelligence plugin and language server for each primary language (detected from manifests, ignoring samples, docs and vendored code)
  • The tools RDF relies on, the hard fd limit and sandbox dependencies
  • Claude Code currency on your update channel
  • .claude settings and memory files
  • The age of the rdf tokens price table

/r-start shows a one-line brief, and rdf doctor runs it offline as its env scope. To silence a finding, add its id to ~/.rdf/env-ignore.