Antigravity Global Agent Config (dotgemini)

Personal global configuration and custom skills for Google Antigravity CLI.

Note

Repository Naming vs. Installation Paths: The repository is named dotagent to reflect dotfile configuration for AI agents. Google Antigravity CLI expects its global settings to reside at ~/.gemini/ on disk. Setup cloning instructions target ~/.gemini for Antigravity-compatibility.

Platform Support

This configuration is developed and tested on macOS and Linux (validated in CI). It is not supported on Windows — the branch resolver script used by the adversarial-review and explain-diff skills will exit with an error on platforms without fcntl, and several paths assume POSIX semantics.

Setup on a New Machine

1. Install Antigravity CLI: Download and install the Antigravity CLI. This is required for executing Antigravity skills (adversarial-review, explain-diff, make-feature, session-sync), environment runner wrappers, and subagent delegation.

   Verify your installation:

   agy --version

   [!NOTE] Installation Order: If you clone the repository before installing the CLI, the CLI will safely initialize its state inside the existing ~/.gemini folder without conflicts when you install it.

2. Backup existing config (if any): Ensure $HOME/.gemini.bak does not already exist before running the backup command. If both directories exist, the backup step will fail with an error to prevent overwriting or nesting:

   if [ -e "$HOME/.gemini" ] || [ -L "$HOME/.gemini" ]; then
     if [ -e "$HOME/.gemini.bak" ] || [ -L "$HOME/.gemini.bak" ]; then
       echo "Error: Both $HOME/.gemini and $HOME/.gemini.bak exist." >&2
       echo "Please manually rename or remove $HOME/.gemini.bak before proceeding." >&2
       exit 1
     else
       mv "$HOME/.gemini" "$HOME/.gemini.bak"
     fi
   fi

3. Clone this repository: Clone the repository to ~/.gemini so that the Antigravity CLI can locate it:

   git clone https://github.com/jerrylin96/dotagent.git ~/.gemini
   # Or via SSH: git clone git@github.com:jerrylin96/dotagent.git ~/.gemini

   Note: The repository was previously named dotgemini; GitHub automatically redirects old URLs and clones to dotagent.

4. Restore local settings and credentials (if applicable):

   # Restore settings, credentials, and machine state safely if backup exists
   if [ -d "$HOME/.gemini.bak" ]; then
     mkdir -p "$HOME/.gemini/antigravity-cli"
     [ -f "$HOME/.gemini.bak/antigravity-cli/settings.json" ] && cp "$HOME/.gemini.bak/antigravity-cli/settings.json" "$HOME/.gemini/antigravity-cli/settings.json"
     
     # Restore credentials and machine IDs
     [ -f "$HOME/.gemini.bak/google_accounts.json" ] && cp "$HOME/.gemini.bak/google_accounts.json" "$HOME/.gemini/google_accounts.json"
     [ -f "$HOME/.gemini.bak/oauth_creds.json" ] && cp "$HOME/.gemini.bak/oauth_creds.json" "$HOME/.gemini/oauth_creds.json"
     [ -f "$HOME/.gemini.bak/antigravity-cli/installation_id" ] && cp "$HOME/.gemini.bak/antigravity-cli/installation_id" "$HOME/.gemini/antigravity-cli/installation_id"
   else
     echo "No backup directory found to restore settings."
   fi

Note: Credentials, local settings (antigravity-cli/settings.json), terminal logs (*.log), subagent run data (brain/), and conversation logs (conversations/) are automatically excluded via .gitignore to prevent leaking secrets or tracking local session state.

---

What's Included

1. Global Rules (AGENTS.md, aliased as GEMINI.md)

• Defines the Ponytail (Lazy Senior Dev Mode) YAGNI ladder.
• Enforces the Caveman (Terse Style) communication formatting to save ~65% output tokens.
• Enforces structured software development lifecycle gates.
• Also exposed as GEMINI.md (a symlink to AGENTS.md for backward compatibility). Antigravity CLI seamlessly reads the rules by default.

2. Isolated Execution Environments

• Integrates with setup_review_env.py to automatically bootstrap CPU-compatible testing and linting virtual environments under ~/.gemini/tmp/<workspace-hash> using uv.

  [!WARNING] This provides dependency/test isolation, NOT a security sandbox. Running setup scripts or installing dependencies (via uv/pip) on untrusted repositories can execute arbitrary build hooks or code under your user credentials.

• Isolates dynamic branch workspaces to prevent test collisions.

3. Custom Registered Skills (skills/)

• adversarial-review — Git worktree-based adversarial code review helper.

  • User Workflow:
    1. Prepare Baseline: Prior to starting the review, checkout the intended reference branch (the baseline, e.g., main or a specific release branch) in your active workspace:

       git checkout <your-reference-branch>

    2. Activate Review: Trigger the review by typing /adversarial-review in the chat.
    3. Select Feature Branch or PR: Antigravity will automatically fetch the latest updates, detect your current checked-out branch as the reference branch, and present a list of all other local and remote branches. Select the feature branch (the target containing the new changes to review) when prompted. You can also name a pull request directly (e.g. "review PR #42" or paste the PR/MR URL) — the resolver fetches the PR head ref from the remote, so fork PRs work without a local branch (GitHub/GitLab/Gitea; not Bitbucket).
    4. Under the Hood: Antigravity checks out the selected feature branch to a clean, isolated worktree under ~/.gemini/tmp/worktrees/, leaving your active workspace untouched on the reference branch. It then generates the diff between the two branches, runs tests/linters in the feature branch worktree, and performs the adversarial review.

• explain-diff — Interactive, read-only walkthrough explaining what changed between a feature branch, pull request, or specific commit/range and a reference branch.

  • User Workflow: Trigger with /explain-diff. Branch selection works exactly like adversarial-review (it reuses the same branch resolver and worktree cache). You then get an overall summary of the changeset and a numbered menu of changed files; pick one for a hunk-by-hunk explanation, ask follow-up questions, and return to the menu until done. Nothing is executed or modified — no tests, no linters.

• make-feature — Git worktree-based isolated feature branch development engine and mandatory software lifecycle pipeline.

  • User Workflow: Trigger with /make-feature (or automatically whenever preparing to write or edit codebase files).
    1. Phase 1 (Spec & Plan): Draft /spec and /plan artifacts. Antigravity pauses for explicit human approval before creating worktrees or writing code.
    2. Phase 2 (Build & Worktree): Creates a feature branch prefixed with gemini/ checked out to an isolated worktree under ~/.gemini/tmp/worktrees/. Performs all file modifications and runs tests via isolated environment wrappers (run_in_env.py).
    3. Phase 3 (Push & Subagent Review): Pushes the feature branch to origin and delegates an isolated background subagent to execute an /adversarial-review loop until approved.
    4. Phase 4 (Human Signoff & Merge): Antigravity pauses, presents the review report, and awaits explicit user confirmation (/signoff) before merging into the target branch and pruning the worktree.

• signoff — Socratic human-comprehension and risk-ownership audit before merging.

  • User Workflow: Trigger with /signoff before merging a feature branch. Antigravity reverse-interviews you across 4 core axes (mechanics, trade-offs, failure boundaries, ownership). Upon verified comprehension and explicit user approval, it appends a flat block of Signoff-* git trailers (including a traceable SHA-256 transcript reference digest) to the commit.

• /ponytail — Lazy senior developer instructions.

• /caveman — Token-efficient caveman communication mode.

• incremental-implementation — Vertical slicing development guidelines.

• test-driven-development — Red-green-refactor testing patterns.

• debugging-and-error-recovery — Systematic triage and root-cause fix strategies.

Running Tests

To run all unit tests for the configured skills and environment setup:

python3 scripts/run_tests.py

This script automatically ensures that the correct isolated virtual environment is initialized/up-to-date and runs all discoverable tests using pytest with importlib import mode.

If uv is available but the environment setup fails, run_tests.py will exit with a non-zero status by default to prevent masking issues. To bypass setup errors and fall back to testing on the host environment anyway, run with the ALLOW_HOST_TEST_FALLBACK=1 environment variable set. If uv is entirely missing, host testing fallback occurs automatically.

---

Sharing and Collaborator Onboarding

Warning

While this repository commits no credentials, local settings, or session history (they are automatically gitignored), replacing a collaborator's local ~/.gemini directory with this repository will replace their own global rules, custom skills, and preferences with yours.

Furthermore, collaborators will lose their local configurations and credentials unless they back up and restore:

• google_accounts.json
• oauth_creds.json
• antigravity-cli/settings.json
• antigravity-cli/installation_id (optional machine state ID)

Collaborators should only clone this as a reference or keep it isolated. For sharing rules/skills across team projects, prefer Project-Level Integration below to build a shared, repository-specific review workflow, or copy only the specific skill subset needed.

Project-Level Integration (Recommended for Sharing)

If you want to bake this isolated environment setup directly into a specific project repository so that any agent working on it automatically uses it:

1. Copy the Env Manager: Place the setup_review_env.py script in the project repository (e.g., scripts/setup_review_env.py).
2. Add a Project-Level Guide: Add an AGENTS.md to the project's root containing:

   ## Isolated Testing & Execution Environment
   All testing and validation commands must run inside the CPU-isolated virtual environment:
   1. Initialize: `python3 scripts/setup_review_env.py <workspace_path>`
   2. Run Tests: `python3 ~/.gemini/scripts/run_in_env.py <workspace_path> pytest`
   *(Note: The setup_review_env.py script automatically manages and prints details of the isolated environment.)*