Clawdius

Delegate tasks from Claude Code to other AI coding CLIs ("backends") to conserve Claude usage. One universal plugin with a capability-based adapter per backend.

Forked from openai/codex-plugin-cc (Apache-2.0); the Codex app-server runtime is replaced by headless CLI adapters.

Backends

Backend	CLI	Enforced read-only	Resume	Live progress	Transfer	Notes
Grok (xAI)	grok	✅ (plan mode)	✅	✅ streaming-json	seeded	Default backend; structured output for review.
Pi Coding Agent	pi	✅ (tool allowlist)	✅	✅ JSONL feed	seeded	Per-turn token/cost usage.
Antigravity (Google)	agy	⚠️ advisory only	✅	❌ plain text	seeded	Not allowed for review.
Codex (OpenAI)	codex	✅ (OS sandbox)	✅	✅ JSONL feed	native import	codex exec headless; structured output; native review mode. On Windows the read-only sandbox needs codex's setup helper working (see notes).

Pick a backend per run with --backend <id>, or set a per-repo default via /clawdius:setup --default-backend <id>.

Prerequisites

• Claude Code with plugin support
• Node.js >= 18.18
• At least one backend CLI installed and authenticated:
  • Grok CLI — docs.x.ai (Grok CLI section), then grok login
  • Pi Coding Agent — npm install -g @earendil-works/pi-coding-agent, then run pi once to configure credentials
  • Antigravity CLI — antigravity.google, then run agy once to sign in
  • Codex CLI — developers.openai.com/codex/cli, then codex login

Install

From GitHub:

/plugin marketplace add vibecodingmd/clawdius
/plugin install clawdius@clawdius

From a local clone:

/plugin marketplace add <path-to>/clawdius
/plugin install clawdius@clawdius

Then run /clawdius:setup — it reports which backend CLIs are installed and authenticated, with install/login hints for any that are missing.

After installing or updating the plugin, start a new Claude Code session before using /clawdius:transfer: the transcript path it needs is exported by a session-start hook, so a session begun before the install won't have it (pass --source <path-to-claude-jsonl> as a manual fallback).

Commands

• /clawdius:rescue [task] — delegate investigation or a fix to a backend (Grok by default)
• /clawdius:review [focus] — adversarial review of the working tree on a read-only backend; --native uses codex's built-in reviewer, and --native --commit <sha> reviews one commit
• /clawdius:transfer — hand the current Claude Code session to a backend: codex imports it as a real thread (visible turn history, no tokens), grok/pi/agy get a seeded handoff session that /clawdius:rescue --resume can continue
• /clawdius:status [job-id] — show active and recent jobs, including recent progress and last activity
• /clawdius:result [job-id] — show the stored output of a finished job
• /clawdius:cancel [job-id] — cancel an active background job
• /clawdius:setup — check backend install/auth status, set the default backend

Add --write to let a task edit files. /clawdius:rescue --background keeps the rescue in Claude Code's native task list and reports its result on completion; --wait keeps it in the foreground. Direct companion CLI users can still use task --background for a detached job and poll it with /clawdius:status. Use --resume-last to continue the previous session.

Design notes

• Every backend is spawned headless with stdin ignored and a hard kill timeout — several backend CLIs hang forever on an open stdin pipe in non-TTY subprocesses, and expired auth can present as a silent hang.
• Adapters implement checkAvailability, checkAuth, buildTaskArgs, parseTaskOutput, and optional capabilities (streaming, resume, interrupt, native review) declared as flags.
• Job tracking, background workers, and session-scoped status/result/cancel are inherited from the codex-plugin-cc job system.
• Live progress: backends that emit a JSONL event feed (grok streaming-json, pi --mode json, codex exec --json) are parsed line by line as they run. Phase transitions (thinking → responding → finalizing), the backend session id, tool/command activity, and per-turn token usage are written to the job log as they happen, so /clawdius:status <job-id> shows a live preview plus the log's last-activity time during a long background run. The full stdout buffer is still the authoritative source for the final result — streaming is best-effort narration layered on top and never changes the parsed output.
• Session transfer (/clawdius:transfer) has two modes, picked by a per-backend capability flag. Codex supports a protocol-level import: a one-shot codex app-server converts the Claude transcript into a real codex thread with visible turn history, spending no tokens (codex skips byte-identical re-imports, so transferring the same unchanged transcript twice reports an error pointing at the earlier thread). The other backends have no import protocol, so clawdius seeds a fresh read-only session whose first turn ingests a transcript digest; that session is registered as the newest resumable task, so /clawdius:rescue --resume continues it directly.
• Adversarial review (/clawdius:review) only runs on backends that can prove read-only execution (grok, pi, codex). Before a codex review on Windows the adapter runs a token-free codex sandbox probe and refuses the review if the OS sandbox can't be verified, so a review verdict is never produced by a backend that could have edited the tree.

Codex on Windows

The codex read-only sandbox depends on codex-windows-sandbox-setup.exe / codex-command-runner.exe being launchable next to the codex launcher. The standalone installer can leave them in the package's codex-resources folder instead — if /clawdius:setup or a review reports the sandbox is broken, copy both exes next to the launcher and, if prompted, approve the one-time UAC elevation by running codex sandbox cmd /c "echo hi" once in your terminal. Codex tasks that only read/reason work regardless; only shell-command and review runs need the sandbox.

Tests

npm test

License

Apache-2.0. See LICENSE and NOTICE — Clawdius is forked from openai/codex-plugin-cc.