Files
zerogravity/docs/extension-server-analysis.md
Nikketryhard 6a07786c4e feat: implement headless LS authentication via state sync
Reverse-engineered the UnifiedStateSyncUpdate protocol:
- initial_state field is bytes (not string), contains serialized Topic proto
- Map key for OAuth is 'oauthTokenInfoSentinelKey'
- Row.value is base64-encoded OAuthTokenInfo protobuf
- OAuthTokenInfo includes access_token, token_type, expiry (Timestamp)
- Set far-future expiry (2099) to prevent token expiry errors

Also fixed:
- PushUnifiedStateSyncUpdate returns proper empty proto response
- Stream keep-alive avoids sending empty envelopes (LS rejects nil updates)
- uss-enterprisePreferences topic handled (empty initial state)
2026-02-15 21:40:35 -06:00

351 lines
13 KiB
Markdown

# Extension Server Protocol — Deep Analysis
Source: `strings` analysis of `/usr/share/antigravity/resources/app/extensions/antigravity/bin/language_server_linux_x64`
---
## Architecture Overview
```mermaid
graph TD
subgraph "Antigravity App (Electron)"
EXT["Extension (TypeScript)"]
EXTSRV["Extension Server<br/>(ConnectRPC, HTTP/1.1)"]
end
subgraph "Language Server (Go binary)"
LS["LanguageServerService<br/>(API Server, ConnectRPC)"]
ESC["ExtensionServerClient<br/>(ConnectRPC client)"]
STATESYNC["UnifiedStateSyncClient"]
CAC["CodeAssistClient"]
end
subgraph "Google Cloud"
GOOGLE["daily-cloudcode-pa.googleapis.com"]
end
EXT -->|"starts"| EXTSRV
ESC -->|"HTTP/1.1<br/>application/connect+proto<br/>(ServerStream RPC)"| EXTSRV
LS -->|"HTTPS<br/>application/connect+proto"| GOOGLE
STATESYNC -->|"via ESC"| EXTSRV
CAC -->|"GetOAuthToken()"| STATESYNC
style EXTSRV fill:#f96,stroke:#333
style STATESYNC fill:#69f,stroke:#333
```
### Communication Paths
| From → To | Protocol | Content-Type | Direction |
| --------------------- | -------- | ------------------------------ | -------------------- |
| Proxy → LS API Server | HTTP/1.1 | `application/json` (JSON-RPC) | Unary |
| Proxy → LS API Server | HTTP/1.1 | `application/proto` (protobuf) | Unary |
| LS → Extension Server | HTTP/1.1 | `application/connect+proto` | **Server-Streaming** |
| LS → Google API | HTTPS/H2 | `application/connect+proto` | Streaming |
> [!IMPORTANT]
> **ALL ExtensionServerService methods use Server-Streaming Connect RPC**, even methods like `GetSecretValue` and `LanguageServerStarted` that are semantically unary.
> This means every response must use Connect streaming envelope framing.
---
## Connect Streaming Protocol
The LS uses `connect-go` v1 (`connectrpc.com/connect/v1`). All extension server calls use **ServerStream**, which requires:
### Request Format (LS → Stub)
```
POST /exa.extension_server_pb.ExtensionServerService/{Method} HTTP/1.1
Content-Type: application/connect+proto
Connect-Protocol-Version: 1
```
Body: raw protobuf (NOT envelope-framed for client→server in ServerStream)
### Response Format (Stub → LS)
```
HTTP/1.1 200 OK
Content-Type: application/connect+proto
```
Body: sequence of **envelope frames**:
```
┌─────────┬──────────────────┬───────────────┐
│ Flags │ Message Length │ Message Data │
│ (1 byte)│ (4 bytes, BE) │ (N bytes) │
├─────────┼──────────────────┼───────────────┤
│ 0x00 │ protobuf length │ protobuf data │ ← data message (optional, 0 or more)
│ 0x02 │ trailer length │ JSON trailer │ ← end-of-stream (required, exactly 1)
└─────────┴──────────────────┴───────────────┘
```
### Examples
**Empty success** (most methods):
```
\x02 \x00\x00\x00\x02 {}
```
= flag(0x02) + length(2) + trailer("{}")
**Success with data** (e.g., GetSecretValue):
```
\x00 \x00\x00\x00\x0e \x0a\x0c ya29.a0A... ← data: protobuf response
\x02 \x00\x00\x00\x02 {} ← end-of-stream trailer
```
**Error response** (e.g., not found):
```
\x02 \x00\x00\x00\x22 {"error":{"code":"not_found"}}
```
> [!WARNING]
> Sending `Content-Length: 0` with `Content-Type: application/connect+proto` causes `unexpected EOF` — the client expects at least the end-of-stream envelope.
---
## Extension Server Service — Method Reference
### All Methods Are ServerStream
From binary analysis, every method on `ExtensionServerService` is wrapped as:
```go
*connect.ServerStreamForClient[...Response]
```
This includes ALL of these methods:
| Method | Request Proto | Response Proto | Purpose |
| ---------------------------------- | ----------------------------------------- | ------------------------------------ | ------------------ |
| `LanguageServerStarted` | `LanguageServerStartedRequest` | `LanguageServerStartedResponse` | Init notification |
| `GetSecretValue` | `GetSecretValueRequest{Key}` | `GetSecretValueResponse{Value}` | Secret store read |
| `StoreSecretValue` | `StoreSecretValueRequest{Key,Value}` | `StoreSecretValueResponse` | Secret store write |
| `SubscribeToUnifiedStateSyncTopic` | `SubscribeToUnifiedStateSyncTopicRequest` | `UnifiedStateSyncUpdate` (stream) | Real-time state |
| `PushUnifiedStateSyncUpdate` | `PushUnifiedStateSyncUpdateRequest` | `PushUnifiedStateSyncUpdateResponse` | Push state updates |
| `GetChromeDevtoolsMcpUrl` | `GetChromeDevtoolsMcpUrlRequest` | `GetChromeDevtoolsMcpUrlResponse` | Chrome DevTools |
| `FetchMCPAuthToken` | `FetchMCPAuthTokenRequest` | `FetchMCPAuthTokenResponse` | MCP auth |
| `AddAnnotation` | `AddAnnotationRequest` | `AddAnnotationResponse` | IDE annotation |
| `OpenFilePointer` | `OpenFilePointerRequest` | `OpenFilePointerResponse` | Open file in IDE |
| `ExecuteCommand` | `ExecuteCommandRequest` | `TerminalShellCommandStreamChunk` | Run terminal cmd |
| ... | ... | ... | 53+ total methods |
---
## OAuth Token Flow
```mermaid
sequenceDiagram
participant LS as Language Server
participant ESC as ExtensionServerClient
participant STUB as Stub Extension Server
participant SYNC as UnifiedStateSyncClient
Note over LS: Startup
LS->>ESC: LanguageServerStarted()
ESC->>STUB: POST .../LanguageServerStarted
STUB-->>ESC: 200 OK (empty envelope)
Note over LS: OAuth Token Request
LS->>SYNC: GetOAuthTokenInfo()
SYNC->>ESC: GetSecretValue({key: "oauth_token"})
ESC->>STUB: POST .../GetSecretValue
STUB-->>ESC: 200 OK (envelope with {value: "ya29.a0..."})
ESC-->>SYNC: {value: "ya29.a0..."}
SYNC-->>LS: OAuthTokenInfo{access_token: "ya29.a0..."}
Note over LS: Using Token
LS->>LS: Set auth_token on API requests
LS->>LS: GetUserStatus, LoadCodeAssist, etc.
```
### Key Insight: How the LS Gets OAuth Tokens
The call chain at `server.go:558` (`Failed to get OAuth token`):
1. **`CodeAssistClient.GetOAuthToken()`** — entry point for getting OAuth
2.**`UnifiedStateSyncClient.GetOAuthTokenInfo()`** — state sync layer
3.**`ExtensionServerClient.GetSecretValue(key)`** — calls extension server
4.**Our stub** must respond with the token value
### Proto Structures
**GetSecretValueRequest** (from binary):
```protobuf
message GetSecretValueRequest {
string key = 1; // protobuf:"bytes,1,opt,name=key"
}
```
**GetSecretValueResponse** (from binary):
```protobuf
message GetSecretValueResponse {
string value = 1; // protobuf:"bytes,1,opt,name=value"
}
```
**OAuthTokenInfo** (from binary):
```protobuf
message OAuthTokenInfo {
string access_token = ?; // .GetAccessToken()
string expiry = ?; // .GetExpiry()
string refresh_token = ?; // .GetRefreshToken()
string token_type = ?; // .GetTokenType()
}
```
**SaveOAuthTokenInfoRequest** contains:
```protobuf
message SaveOAuthTokenInfoRequest {
OAuthTokenInfo token_info = 1; // protobuf:"bytes,1,opt,name=token_info"
}
```
---
## UnifiedStateSyncTopic
The state sync system is central to the LS's operation:
```mermaid
graph LR
LSS["LS State Sync"] -->|"Subscribe"| EXT["Extension Server"]
EXT -->|"Stream updates"| LSS
LSS -->|"Push"| EXT
LSS --> |"Reads"| OAuth["OAuth Token"]
LSS --> |"Reads"| Settings["User Settings"]
LSS --> |"Reads"| Models["Available Models"]
LSS --> |"Reads"| Trajectories["Trajectory Summaries"]
```
### What the LS reads via state sync:
From `statesync.(*UnifiedStateSyncClient)` methods:
- `GetOAuthTokenInfo()` — OAuth access token
- `GetSecureModeEnabled()` — security settings
- `GetBrowserToolsEnabled()` — browser tool access
- `GetBrowserAllowlist()` — allowed browser URLs
- `GetBrowserJsExecutionPolicy()` — JS execution policy
- `GetTerminalAutoExecutionPolicy()` — terminal auto-run policy
- `GetTerminalAllowedCommands()` — allowed terminal commands
- `GetTerminalDeniedCommands()` — denied terminal commands
- `GetGcpProjectID()` — GCP project binding
- `GetAllowAgentAccessNonWorkspaceFiles()` — file access policy
- `GetAllowCascadeAccessGitignoreFiles()` — gitignore policy
- `GetAutoContinueOnMaxGeneratorInvocations()` — auto-continue
- `GetDisableCascadeAutoFixLints()` — lint autofix
- `GetTabGitignoreAccess()` — tab-level gitignore
### SubscribeToUnifiedStateSyncTopic
This is a **long-lived server stream**. The LS subscribes at startup and expects to receive `UnifiedStateSyncUpdate` messages over time. If the stub closes the connection, the LS will reconnect.
For a minimal stub, we can:
1. Accept the subscription request
2. Send an end-of-stream trailer immediately (empty data)
3. Close the connection
The LS will poll/retry, but this won't break functionality — it just means settings won't be dynamically updated (which is fine for headless mode).
---
## Proxy → LS Communication
The proxy communicates with the LS API server (`LanguageServerService`) differently:
| Method | Content-Type | Protocol |
| --------------------------------- | ------------------- | ------------- |
| `call_json()` (most methods) | `application/json` | Unary Connect |
| `call_proto()` (protobuf methods) | `application/proto` | Unary Connect |
Both use `Connect-Protocol-Version: 1` header.
> [!NOTE]
> The LS API server accepts unary Connect calls with `application/json` or `application/proto`.
> This is different from the Extension Server which expects streaming `application/connect+proto`.
---
## What the Headless Stub Must Implement
### Critical (blocks all requests):
1. **GetSecretValue** — Must return the OAuth token when requested
- Request body contains `key` (field 1, string)
- Response must contain `value` (field 1, string) = the OAuth token
- Must use Connect streaming envelope framing
2. **LanguageServerStarted** — Must acknowledge startup
- Empty success response (just end-of-stream trailer)
### Important (blocks some features):
3. **SubscribeToUnifiedStateSyncTopic** — Settings subscription
- Can return empty end-of-stream (no data messages)
- LS will retry but won't crash
4. **PushUnifiedStateSyncUpdate** — State push
- Can return empty success
### Nice-to-have (non-blocking):
5. All other methods — return empty success
- `GetChromeDevtoolsMcpUrl`, `ShowAnnotation`, `OpenFilePointer`, etc.
---
## Current Stub Issues (from latest debug log)
### Issue 1: "key not found"
```
E0215 20:05:56.311541 server.go:558] Failed to get OAuth token: key not found
```
The `GetSecretValue` response doesn't match what the LS expects. The LS calls `GetSecretValue` with a specific key, but our stub ignores the key and always returns the token. The "key not found" error suggests the LS's state sync layer caches by key and doesn't find the expected entry.
**Root cause**: The LS doesn't just call `GetSecretValue` — it goes through the `UnifiedStateSyncClient` which uses `GetRow(key)`. The state sync is a key-value store. The LS looks up a specific key in state sync, and the state sync client calls `GetSecretValue` on the extension server. Since our stub returns an empty protobuf for everything except `GetSecretValue`, the state sync's initial `SubscribeToUnifiedStateSyncTopic` gets no data, and subsequent `GetRow()` calls return "key not found".
### Issue 2: "unknown model key MODEL_PLACEHOLDER_M18"
```
E0215 20:05:56.358443 interceptor.go:74] SendUserCascadeMessage: unknown model key MODEL_PLACEHOLDER_M18
```
The model configuration isn't loaded because `Cache(loadCodeAssistResponse)` failed. This cache depends on `userInfo` which depends on the OAuth token. Fix the token flow and this should resolve.
### Issue 3: "mkdir permission denied"
```
E0215 20:05:56.311614 log.go:380] Failed to create artifacts directory...mkdir /tmp/antigravity-standalone/.gemini/antigravity-standalone/brain/.../: permission denied
```
The LS tries to create directories under the `gemini_dir`. This is non-fatal but noisy.
---
## Recommended Fix Strategy
The current approach of parsing individual methods won't scale — ALL 53+ methods are ServerStream and need envelope framing.
**Better approach**: Instead of understanding every method, ensure:
1. **Every response** uses Connect streaming envelope framing (`0x02 + len + {}` minimum)
2. **GetSecretValue** returns the token in a data envelope before the end-of-stream
3. **Content-Type** is always `application/connect+proto`
4. **Connection: close** to avoid HTTP keep-alive issues
5. Create the `gemini_dir` with proper permissions before spawning the LS