This is the canonical shared policy for all terminal-capable agents in this workspace. Applies to Codex, Gemini Code Assist, Cline, Kilo Code, Roo Code, Claude Code, and similar agents.
All executable terminal commands must run through:
$HOME/.local/bin/modal-agent-runner -c "<raw command string>"or
$HOME/.local/bin/modal-agent-runner -- <binary> <arg1> <arg2>The runner forwards commands to scripts/modal_exec.sh, which executes them in Modal via modal_tasks.py.
AGENTS.mdmirrors this file for Codex/Kilo/Roo/Claude-style agents..rules/antigravity.rulesand.rules/codex.rulespoint to this policy for Codex rules loading..clinerulesmirrors this file for Cline..vscode/settings.jsondefinesgeminicodeassist.rulesfor Gemini Code Assist..gemini/bin/bashand.gemini/.envforce Geminibash -cshell invocations throughmodal-agent-runner.
Do not run local direct commands such as:
ls,cat,find,rgpython,python3,pip,uvnode,npm,pnpm,yarnpytest,ruff,mypy,make,bash,zsh
unless they are inside the Modal execution path above.
Some operations are not terminal compute and remain local by design:
- UI rendering and editor interactions
- shell builtins (
cd,export,alias) - source-control operations (
git ...) - Modal control-plane submission command itself (
python3 -m modal ...) - network/chat inference performed by the provider backend
Run:
make antigravity-policy-checkSuccessful routing should show:
hostnameprintsmodalIN_MODAL_TASK_RUNNER=1