Gemini Bridge

Overview¶

Deprecated: the Gemini CLI bridge is kept for compatibility only. The upstream Google CLI flow is obsolete for this package, so new integrations should use another AgentCtrl backend.

The Gemini bridge wraps the gemini CLI (from @google/gemini-cli), Google's terminal-based coding agent. Gemini CLI supports model aliases, approval modes (default, auto_edit, yolo, plan), sandbox isolation, extensions, MCP servers, policy files, session management, and stream-json event streaming. It provides token usage data including cached token counts.

The bridge is implemented by GeminiBridge and configured through GeminiBridgeBuilder. Access the builder through the AgentCtrl facade:

use Cognesy\AgentCtrl\AgentCtrl;
use Cognesy\AgentCtrl\Enum\AgentType;

// Dedicated factory method
$builder = AgentCtrl::gemini();

// Or via the generic factory
$builder = AgentCtrl::make(AgentType::Gemini);

Prerequisites¶

Install Gemini CLI globally:

# npm
npm install -g @google/gemini-cli

# Homebrew
brew install gemini-cli

# npx (no install)
npx @google/gemini-cli

Configure authentication (one of):

# Gemini API key
export GEMINI_API_KEY=...

# Google Cloud API key
export GOOGLE_API_KEY=...

# Or authenticate via Google account (free tier)
gemini

Basic Usage¶

use Cognesy\AgentCtrl\AgentCtrl;

$response = AgentCtrl::gemini()
    ->execute('Explain the architecture of this project.');

echo $response->text();

With model selection:

$response = AgentCtrl::gemini()
    ->withModel('flash')
    ->execute('Review the test suite.');

echo $response->text();

Model Selection¶

Gemini CLI supports model aliases and full model names:

// Model aliases
AgentCtrl::gemini()->withModel('auto');        // Default (gemini-2.5-pro)
AgentCtrl::gemini()->withModel('pro');         // gemini-2.5-pro
AgentCtrl::gemini()->withModel('flash');       // gemini-2.5-flash
AgentCtrl::gemini()->withModel('flash-lite');  // gemini-2.5-flash-lite

// Full model name
AgentCtrl::gemini()->withModel('gemini-2.5-pro');

Approval Modes¶

Gemini CLI supports four approval modes that control how tool execution is approved:

use Cognesy\AgentCtrl\Gemini\Domain\Enum\ApprovalMode;

// Default — prompt for approval on each tool use
AgentCtrl::gemini()->withApprovalMode(ApprovalMode::Default);

// Auto-edit — auto-approve edit tools, prompt for others
AgentCtrl::gemini()->withApprovalMode(ApprovalMode::AutoEdit);

// YOLO — auto-approve all tool executions
AgentCtrl::gemini()->yolo();

// Plan — read-only analysis mode
AgentCtrl::gemini()->planMode();

Sandbox Mode¶

Enable sandboxed execution for process isolation:

AgentCtrl::gemini()
    ->withSandbox()
    ->execute('Analyze the codebase.');

On macOS, this uses Seatbelt (sandbox-exec). Docker, Podman, and gVisor are also supported.

System Prompt¶

Gemini CLI reads instructions from a GEMINI.md file in the project root (similar to CLAUDE.md). You can also set the GEMINI_SYSTEM_MD environment variable to point to a custom system prompt file.

Include Directories¶

Add additional workspace directories for the agent to access:

AgentCtrl::gemini()
    ->withIncludeDirectories(['/projects/shared-lib', '/projects/config'])
    ->execute('Check for shared dependencies.');

Extensions¶

Use specific extensions:

AgentCtrl::gemini()
    ->withExtensions(['my-extension'])
    ->execute('...');

MCP Servers¶

Restrict which MCP servers are available:

AgentCtrl::gemini()
    ->withAllowedMcpServers(['filesystem', 'github'])
    ->execute('...');

Policy Files¶

Load additional policy files for fine-grained tool approval rules:

AgentCtrl::gemini()
    ->withPolicy(['/path/to/policy.yaml'])
    ->execute('...');

Allowed Tools¶

Restrict which tools the agent can use:

AgentCtrl::gemini()
    ->withAllowedTools(['read_file', 'search_files', 'list_directory'])
    ->execute('Analyze the codebase structure.');

Debug Mode¶

Enable debug output for troubleshooting CLI behavior:

AgentCtrl::gemini()
    ->debug()
    ->execute('Analyze the codebase.');

Streaming with Gemini¶

Gemini streams output as JSONL with the stream-json format. The bridge normalizes these into the standard callback API:

use Cognesy\AgentCtrl\AgentCtrl;
use Cognesy\AgentCtrl\Dto\AgentResponse;

$response = AgentCtrl::gemini()
    ->onText(fn(string $text) => print($text))
    ->onToolUse(fn(string $tool, array $input, ?string $output) => print("\n> [{$tool}]\n"))
    ->onError(fn(string $message, ?string $code) => print("\nError: {$message}\n"))
    ->onComplete(fn(AgentResponse $r) => print("\n--- Done ---\n"))
    ->executeStreaming('Analyze the error handling in this codebase.');

Event Normalization¶

Gemini emits stream-json events that are normalized:

message (role=assistant, delta=true) -- Text deltas delivered through onText().
tool_result -- Tool results delivered through onToolUse() with tool name, input (from paired tool_use event), result, and error status.
error -- Errors delivered through onError() with severity and message.
init, tool_use, result -- Lifecycle events available through the wiretap() event system.

Session Management¶

Gemini CLI maintains session history. Agent-Ctrl extracts session IDs from the init event:

// First execution
$first = AgentCtrl::gemini()->execute('Create an implementation plan.');
$sessionId = $first->sessionId();

// Continue the most recent session
$next = AgentCtrl::gemini()
    ->continueSession()
    ->execute('Begin implementing the plan.');

// Resume a specific session by ID
if ($sessionId !== null) {
    $next = AgentCtrl::gemini()
        ->resumeSession((string) $sessionId)
        ->execute('Continue with the next step.');
}

Usage Data¶

Gemini provides token usage data from the result event stats:

$response = AgentCtrl::gemini()
    ->withModel('flash')
    ->execute('Analyze the project dependencies.');

$usage = $response->usage();
if ($usage !== null) {
    echo "Input tokens:    {$usage->input}\n";
    echo "Output tokens:   {$usage->output}\n";
    echo "Total tokens:    {$usage->total()}\n";

    if ($usage->cacheRead !== null) {
        echo "Cached tokens:   {$usage->cacheRead}\n";
    }
}

Data Availability¶

Data Point	Available	Notes
Text output	Yes	Extracted from `message` events (role=assistant, delta=true)
Tool calls	Yes	Normalized from `tool_use` + `tool_result` event pairs
Session ID	Yes	Extracted from `init` event
Token usage	Yes	Input, output, cached tokens from `result` stats
Cost	No	Gemini CLI does not expose cost data
Parse diagnostics	Yes	Malformed JSON line counts and samples

Complete Example¶

use Cognesy\AgentCtrl\AgentCtrl;
use Cognesy\AgentCtrl\Dto\AgentResponse;
use Cognesy\AgentCtrl\Gemini\Domain\Enum\ApprovalMode;

$response = AgentCtrl::gemini()
    ->withModel('pro')
    ->withApprovalMode(ApprovalMode::AutoEdit)
    ->withIncludeDirectories(['/projects/shared'])
    ->withTimeout(300)
    ->inDirectory('/projects/app')
    ->onText(fn(string $text) => print($text))
    ->onToolUse(fn(string $tool, array $input, ?string $output) => print("\n> [{$tool}]\n"))
    ->onComplete(fn(AgentResponse $r) => print("\n--- Complete ---\n"))
    ->executeStreaming('Review the application architecture and suggest improvements.');

if ($response->isSuccess()) {
    echo "\nReview completed successfully.\n";
    echo "Tools used: " . count($response->toolCalls) . "\n";

    $usage = $response->usage();
    if ($usage !== null) {
        echo "Tokens: {$usage->total()} (in: {$usage->input}, out: {$usage->output})\n";
    }
} else {
    echo "\nFailed with exit code: {$response->exitCode}\n";
}

Comparison with Other Bridges¶

Feature	Claude Code	Codex	OpenCode	Pi	Gemini
System prompts	Yes (replace + append)	No	No	Yes (replace + append)	Yes (GEMINI.md file)
Permission modes	Yes (4 levels)	No	No	No	Yes (4 modes)
Turn limits	Yes	No	No	No	Yes (via settings)
Sandbox modes	No	Yes (3 levels)	No	No	Yes (Seatbelt/Docker/Podman/gVisor)
Image input	No	Yes	No	No	No
Thinking levels	No	No	No	Yes (6 levels)	No
Named agents	No	No	Yes	No	No
File attachments	No	No	Yes	Yes (@-prefix)	No
Extensions	No	No	No	Yes (TypeScript)	Yes
Skills	No	No	No	Yes	No
Tool control	No	No	No	Yes (select/disable)	Yes (allowlist)
MCP servers	No	No	No	No	Yes
Policy engine	No	No	No	No	Yes
Session sharing	No	No	Yes	No	No
Session titles	No	No	Yes	No	No
Ephemeral mode	No	No	No	Yes	No
API key override	No	No	No	Yes	No
Token usage	No	Yes (partial)	Yes (full)	Yes	Yes
Cost tracking	No	No	Yes	Yes	No
Multi-provider models	No	No	Yes	Yes	No
Include directories	No	No	No	No	Yes
Debug mode	No	No	No	No	Yes
Free tier	No	No	No	No	Yes

Environment Variables¶

Variable	Description
`GEMINI_API_KEY`	Gemini API key
`GOOGLE_API_KEY`	Google Cloud API key
`GOOGLE_APPLICATION_CREDENTIALS`	Service account JSON path
`GOOGLE_CLOUD_PROJECT`	Project ID for Code Assist
`GOOGLE_GENAI_USE_VERTEXAI`	Enable Vertex AI
`GEMINI_SANDBOX`	Enable sandbox without CLI flag