Summary

Adds support for multiple Antigravity accounts with automatic round-robin load balancing and failover.

Problem

• Single Antigravity account can get rate limited, blocking all requests
• Antigravity rate limits cause requests to hang indefinitely rather than returning errors
• Upstream @mariozechner/pi-ai expects credentials under a single google-antigravity key

Solution

Multi-account storage

Accounts stored with namespaced keys in oauth.json:

{
  "google-antigravity": { ... },
  "google-antigravity:[email protected]": { ... }
}

Round-robin selection

Each request picks the account with the oldest lastUsed timestamp (least-recently-used).

Automatic failover

On retryable errors (401, 403, 429, quota, rate limit, timeout), marks the account as failed and automatically retries with the next available account.

Timeout-based failover (new)

Since Antigravity rate limits cause hangs rather than errors, we added timeout detection:

• When a request times out (based on agent.timeoutSeconds), treat it as a rate limit
• Mark the account as failed and try the next available account
• Logs: Antigravity account X timed out (possible rate limit). Trying next account...

Exponential backoff

Failed accounts get cooldown: 1min → 5min → 25min → 1hr max. State tracked in antigravity-state.json.

The storage swap trick

We create a temp storage object with the selected account mapped to google-antigravity, so the upstream library's token refresh works transparently without modification.

CLI Commands

clawdbot antigravity add     # Add new account (opens browser auth)
clawdbot antigravity list    # List all accounts  
clawdbot antigravity remove  # Remove an account

Files changed

• src/agents/model-auth.ts - Core multi-account logic
• src/agents/pi-embedded-runner.ts - Failover retry loop with timeout detection
• src/cli/antigravity-cli.ts - CLI command registration
• src/cli/program.ts - Register antigravity subcommand
• src/commands/antigravity.ts - Add/list/remove implementations

Configuration

The failover timeout is controlled by agent.timeoutSeconds (default: 600s / 10 min). For faster failover on rate limits, lower this value:

{
  "agent": {
    "timeoutSeconds": 180
  }
}

Trade-off: Lower timeout = faster failover but may cut off legitimately long-running tasks.

Testing

• Tested with 4 accounts - round-robin alternation confirmed via antigravity-state.json timestamps
• Timeout failover confirmed in production: account marked failed after timeout, switched to next account successfully
• Verified failover logs appear correctly