Antigravity Bridge

Antigravity Bridge

Bidirectional knowledge bridge between OpenClaw and Google Antigravity IDE. Both agents share the same knowledge, tasks, and learnings — no manual sync needed.

Prerequisites

• Google Antigravity IDE installed (with ~/.gemini/antigravity/ directory)
• A project with .agent/ directory (Antigravity's standard agent config)
• Python 3.10+ (for sync scripts)

Configuration

Before first use, create a config file:

python3 "$(dirname "$0")/scripts/configure.py"

Or create ~/.openclaw/antigravity-bridge.json manually:

{
  "knowledge_dir": "~/.gemini/antigravity/knowledge",
  "project_dir": "~/path/to/your/project",
  "agent_dir": "~/path/to/your/project/.agent",
  "gemini_md": "~/path/to/your/project/.gemini/GEMINI.md",
  "auto_commit": false
}

Sync Command — The Core Workflow

Running the sync

python3 scripts/sync_knowledge.py

This outputs structured JSON to stdout with two sections:

• diff — what changed since last sync (or "first sync" if no previous state)
• current — full snapshot of current Antigravity state

State is tracked in ~/.openclaw/workspace/antigravity-sync-state.json.

Agent responsibilities after sync

The agent (not the script) is responsible for updating OpenClaw memory.

After running the sync script, the agent MUST:

1. Read the JSON output, focusing on the diff section
2. If diff.is_first_sync is true or diff.summary shows changes:
   • Update MEMORY.md with significant changes:
     • Task count changes (done/todo deltas)
     • New or removed KI topics
     • New active tasks or phase changes
     • New lessons learned (from memory files)
   • Append a sync log entry to memory/YYYY-MM-DD.md:

     HH:MM — Antigravity Bridge sync: <diff.summary>
     

3. If diff.summary is "No changes since last sync":
   • No updates needed. Optionally log the sync attempt.

What goes where

What NOT to do

• Do NOT create standalone reference docs (no ANTIGRAVITY.md, no antigravity-sync.md)
• Do NOT dump raw sync data into files — distill into MEMORY.md
• Do NOT ask the user whether to update MEMORY.md — just do it

Other Commands

Diff Tasks (diff)

Show what changed in tasks since last sync:

python3 scripts/diff_tasks.py

Next Task (pick-task) — mirrors Antigravity's /next-task

Gather project context for intelligent task selection:

python3 scripts/pick_task.py

This outputs structured JSON context (does NOT modify tasks.md). The agent uses this to reason.

Agent workflow after running pick-task:

1. Read the JSON output (tasks, git log, sessions, memory)
2. Analyze:
   • Active [>] tasks that need finishing
   • Session handoffs (continuation prompts from Antigravity)
   • Recent commits (what was just worked on)
   • Phase dependencies (what's unblocked)
3. Recommend 2-3 tasks to the user with:
   • Context: why this task now
   • Scope: what's involved
   • Effort: Small / Medium / Large
4. Present as numbered options:

   🎯 Recommended Next Tasks — Reply 1, 2, or 3
   
   ### Option 1: [Task Name] ⭐
   - Context: ...
   - Scope: ...
   - Effort: Medium
   
   ### Option 2: [Task Name]
   ...
   

5. User picks → agent marks task [>] in tasks.md
6. Agent spawns coding sub-agent with task brief (rules + relevant memory + KIs)
7. After completion: mark [x], run self-improve

Priority criteria (in order):

1. Active but incomplete ([>] tasks)
2. Unblocking (enables other work)
3. Quick wins (low effort, high value)
4. Technical debt (flagged in audits/lessons)
5. Natural progression (next step in current phase)

The script gathers data. The agent thinks. The user decides.

Self-Improve (self-improve)

Update BOTH knowledge systems with session learnings:

python3 scripts/self_improve.py --topic "<topic>" --lesson "<what was learned>"

What it does:

1. OpenClaw side: Updates MEMORY.md and memory/YYYY-MM-DD.md
2. Antigravity side:
   • Creates/updates .agent/memory/lessons-learned-<topic>.md
   • Creates/updates KI artifacts in knowledge/<topic>/artifacts/
   • Updates metadata.json and timestamps.json
3. Optionally commits changes

Write Knowledge Item (write-ki)

Write directly to Antigravity's native Knowledge Item system:

python3 scripts/write_ki.py \
  --topic "my_topic" \
  --title "My Topic" \
  --summary "What this knowledge covers" \
  --artifact "pattern_name" \
  --content "# Pattern\n\nDetailed content here..."

Create Antigravity Skill (create-agy-skill)

Generate the reverse-direction Antigravity skill:

python3 scripts/create_agy_skill.py --project-dir ~/path/to/project

Architecture

┌─────────────────────┐       ┌─────────────────────┐
│    Antigravity       │       │    OpenClaw          │
│    (Gemini)          │       │    (Any LLM)         │
│                      │       │                      │
│  knowledge/       ◄──┼───────┼──► MEMORY.md         │
│  .agent/tasks.md  ◄──┼───────┼──► MEMORY.md (tasks) │
│  .agent/memory/   ◄──┼───────┼──► MEMORY.md (lessons)│
│  .agent/sessions/ ◄──┼───────┼──► MEMORY.md (handoff)│
│  .agent/rules/    ───┼───────┼──► MEMORY.md (counts) │
│  .agent/skills/   ───┼───────┼──► MEMORY.md (counts) │
│  .agent/workflows/───┼───────┼──► MEMORY.md (counts) │
└─────────────────────┘       └─────────────────────┘
         │                              │
         └──── state.json ◄─────────────┘
              (diff tracking)

State File

~/.openclaw/workspace/antigravity-sync-state.json tracks:

• Last sync timestamp
• KI topic hashes and artifact counts
• Task file hash and counts
• Memory/rules/skills/workflows file hashes

This enables precise change detection without markdown diffing.

Security & Privacy

• All data stays local. No external API calls. No cloud sync.
• Scripts only read/write files within configured directories.
• No credentials or tokens are required.
• The skill never modifies code — only knowledge/task/memory files.

External Endpoints

None. This skill operates entirely on local files.

Trust Statement

This skill reads and writes files within your Antigravity knowledge directory (~/.gemini/antigravity/) and your project's .agent/ directory. It does not send any data to external services. Only install if you trust the skill to modify these directories.