Bug Report: Antigravity 1.23.2 — Agent Chat Silently Drops All Messages

# Bug Report: Antigravity 1.23.2 — Agent Chat Silently Drops All Messages

**Reporter**: Paid subscriber (Google AI Pro/Ultra, $200/mo tier)

**Date filed**: 2026-04-30

**Severity**: P0 — Core agent feature unusable, blocking daily workflow for 2+ days

**Reproducibility**: 100% on affected machine

-–

## Summary

Antigravity 1.23.2 agent chat is completely broken. Messages typed in the Agent panel or Agent Manager are accepted by the UI (rendered as chat session titles) but **never forwarded to the language server or Google CloudCode backend**. Zero `streamGenerateContent` calls leave the client. The conversation appears active but the agent never responds — silent failure with no error toast surfaced to user.

This started **on or around 2026-04-29** after running 1.23.2 successfully for 13 days since the Apr 16 release.

-–

## Environment

| Field | Value |

|—|—|

| App version | Antigravity 1.23.2 (User install) |

| Extension version | google.antigravity 0.2.0 |

| OS | Windows 11 |

| Install path | `C:\Antigravity\` |

| User data | `%APPDATA%\Antigravity\` |

| Install method | Originally direct download; reinstalled via `winget install Google.Antigravity --force` during debugging |

| Account | Signed in with valid Google account (“Tom Winchester (Google Auth)” shown in avatar dropdown) |

| Network | Outbound to `daily-cloudcode-pa.googleapis.com` works, polls succeed |

| Smart App Control | Off (was On until 2026-04-29 — see below) |

-–

## Reproduction Steps

1. Launch Antigravity (welcome screen or with workspace open — both fail identically)

2. Sign in with Google account → confirmed signed-in

3. Open Agent panel (`Ctrl+L`) OR Agent Manager (`Ctrl+E`)

4. Type any message (e.g., “alo”, “test”, “hello”)

5. Press Enter / click send

**Expected**: Message sent to selected model (Gemini 3.1 Pro / Claude Sonnet 4.6 / etc.), agent responds.

**Actual**: Message displays as conversation title with retry icon (↶) but no response ever arrives. No error shown in UI.

Tested with:

- Plan mode ON (worktree-based) → same silent failure (also produces “failed to remove worktree: not a git repository” toast when workspace isn’t git)

- Plan mode OFF (Chat mode) → same silent failure

- Multiple models (Gemini 3.1 Pro High, Claude Sonnet 4.6 Thinking) → same

- Welcome screen with no workspace open → same

- Workspace opened (git repo) → same

- New conversation, fresh sign-in, after full state wipe → same

-–

## Diagnostic Evidence

### 1. Language server log (`%APPDATA%\Antigravity\logs\\window1\exthost\google.antigravity\Antigravity.log`)

After successful auth, only two API endpoints are ever called:

```

URL: https://daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist

URL: https://daily-cloudcode-pa.googleapis.com/v1internal:fetchAvailableModels

```

**Zero** calls to any send-message / `streamGenerateContent` endpoint, even after multiple message submissions. The user input never reaches the Go language server.

Language server experiences SIGTERM/restart loop:

```

Language server started

LS lspClient started successfully

… (1-2 seconds elapse) …

Language server killed with signal SIGTERM

Starting language server process with pid

Language server started

… (1-2 seconds elapse) …

Language server killed with signal SIGTERM

```

### 2. DevTools Console (Help → Toggle Developer Tools)

Floods with this error, repeating every few seconds:

```

ERR [Extension Host] FetchError: Unleash Repository error:

at ClientRequest. (c:\Antigravity\resources\app\extensions\antigravity\dist\extension.js:2:812944)

at ClientRequest.emit (node:events:519:28)

at emitErrorEvent (node:_http_client:107:11)

at Socket.socketErrorListener (node:_http_client:574:5)

```

The port number is dynamic (changes per session: 49720, 56116, etc.) but **the proxy server never binds to it** — verified via `Get-NetTCPConnection -State Listen` showing no listener at the expected port while other Antigravity ports (LSP gRPC at 49718, 49719, etc.) bind successfully.

### 3. Manifest declaration errors (DevTools Console)

Extension declares menu items referencing commands not defined in the `commands` section:

```

ERR [google.antigravity]: Menu item references a command ‘antigravity.importAntigravitySettings’ which is not defined in the ‘commands’ section.

ERR [google.antigravity]: Menu item references a command ‘antigravity.importAntigravityExtensions’ which is not defined in the ‘commands’ section.

ERR [google.antigravity]: Menu item references a command ‘antigravity.prioritized.chat.open’ which is not defined in the ‘commands’ section.

```

`antigravity.prioritized.chat.open` is particularly suspicious given the symptom is chat send doing nothing.

### 4. Localhost binding test

```powershell

Test-NetConnection -ComputerName 127.0.0.1 -Port 49720

# PingSucceeded: True

# TcpTestSucceeded: False ← no listener on the port the proxy URL targets

Test-NetConnection -ComputerName 127.0.0.1 -Port 49718

# TcpTestSucceeded: True ← LSP gRPC binds fine

```

Port 49720 is not reserved (`netsh interface ipv4 show excludedportrange tcp` confirms), no firewall rule blocks it (`Get-NetFirewallApplicationFilter` returns no Antigravity-specific rules), no other process owns it. The Unleash proxy simply never starts.

### 5. Smart App Control note

A separate prior issue: from ~2026-04-24, Microsoft’s SAC reputation database began blocking unsigned modules inside Antigravity:

```

Code Integrity event 3033/3077:

attempted to load \Device\HarddiskVolume3\Antigravity\resources\app\node_modules\windows-foreground-love\build\Release\foreground_love.node

did not meet the Enterprise signing level requirements

Policy ID: {0283ac0f-fff1-49ae-ada1-…} ← Smart App Control policy

```

This was resolved on 2026-04-29 by disabling SAC. The current bug is **not** related to SAC and persists with SAC fully off.

-–

## Things Already Ruled Out (Please Don’t Suggest These)

- :cross_mark: Sign-in / OAuth issue → auth succeeds, `loadCodeAssist` returns 200

- :cross_mark: Workspace state corruption → wiped `workspaceStorage`, `globalStorage`, `state.vscdb` entirely; same bug

- :cross_mark: Plan mode / git worktree requirement → Chat mode fails identically

- :cross_mark: Firewall / antivirus → no block rules, port 49720 unreserved, no Defender threat detection in last 24h

- :cross_mark: Smart App Control → disabled, CodeIntegrity errors stopped, bug persists

- :cross_mark: Network connectivity → Google CloudCode endpoints reachable, polls succeed

- :cross_mark: Stale install → reinstalled clean via `winget --force` (folders + registry + protocol handler removed first)

- :cross_mark: Stale credentials → wiped Windows Credential Manager entries (none present anyway)

- :cross_mark: User settings override → `settings.json` is `{}` (default)

- :cross_mark: Account permission → same Google account works on other AI products (Gemini app, AI Studio)

-–

## Likely Root Cause

The Unleash feature-flag proxy server (started in-process by the `google.antigravity` extension at activation) fails to bind to its localhost port. Because feature flags can’t load, the chat command (`antigravity.prioritized.chat.open` and related) is gated off — UI accepts input but the send pipeline is disabled at the flag level. No error is surfaced.

This may be a **server-side change** Google rolled out around 2026-04-29 that’s incompatible with the Apr 16 1.23.2 client (proxy code can’t handle new Unleash response format and crashes silently before binding), OR a build defect in 1.23.2 that was previously masked by cached feature-flag state and only surfaced after our fresh install / state wipe.

-–

## Impact

- **2+ days of agent downtime** on a paid subscription

- Workflow blocked: hybrid sandbox setup that uses Antigravity to cross-check tasks from another agent — replacement requires switching providers

- No workaround discovered after ~2 hours of systematic debugging

- Bug is silent (no error toast) — users without DevTools console access have no way to know it’s a backend/build issue and not user error

-–

## Requests

1. **Acknowledge** the bug exists in 1.23.2 (or confirm if other users are affected)

2. **ETA** for fix or rollback of the backend change

3. **Service credit** for the days of downtime — at $200/mo this is roughly $13/day prorated

4. **Workaround**, if any: env var, settings.json key, or older installer URL to downgrade to a known-good version

-–

## Attachments to include with the support ticket

If the support form allows file uploads, attach:

1. `%APPDATA%\Antigravity\logs\\window1\exthost\google.antigravity\Antigravity.log` (language server log)

2. `%APPDATA%\Antigravity\logs\\window1\exthost\exthost.log` (extension host log)

3. DevTools Console screenshot showing the FetchError flood

4. Output of these PowerShell commands:

```powershell

winget list --id Google.Antigravity

Get-NetTCPConnection -State Listen -LocalAddress 127.0.0.1 | Where-Object { $_.OwningProcess -in (Get-Process Antigravity).Id } | Sort-Object LocalPort

Get-MpComputerStatus | Select-Object SmartAppControlState

```

-–