feat(gemini): gemini cleanup and refactor (#300)

Closes #260

## Description

Cleans up and fixes issues with Gemini module.

## Type of Change

- [ ] New module
- [ ] Bug fix
- [X] Feature/enhancement
- [X] Documentation
- [ ] Other

## Module Information

<!-- Delete this section if not applicable -->

**Path:** `registry/coder-labs/modules/gemini`  
**New version:** `v1.1.0`  
**Breaking change:** [X] Yes [ ] No

## Testing & Validation

- [X] Tests pass (`bun test`)
- [X] Code formatted (`bun run fmt`)
- [X] Changes tested locally

## Related Issues
 
#260 
<!-- Link related issues or write "None" if not applicable -->

---------

Co-authored-by: Atif Ali <atif@coder.com>

registry/coder-labs/modules/gemini/README.md

@@ -1,36 +1,41 @@
 ---
 display_name: Gemini CLI
+description: Run Gemini CLI in your workspace for AI pair programming
 icon: ../../../../.icons/gemini.svg
-description: Run Gemini CLI in your workspace with AgentAPI integration
 verified: true
 tags: [agent, gemini, ai, google, tasks]
 ---
 
 # Gemini CLI
 
-Run [Gemini CLI](https://ai.google.dev/gemini-api/docs/cli) in your workspace to access Google's Gemini AI models, and custom pre/post install scripts. This module integrates with [AgentAPI](https://github.com/coder/agentapi) for Coder Tasks compatibility.
+Run [Gemini CLI](https://github.com/google-gemini/gemini-cli) in your workspace to access Google's Gemini AI models for interactive coding assistance and automated task execution.
 
 ```tf
 module "gemini" {
   source   = "registry.coder.com/coder-labs/gemini/coder"
-  version          = "1.0.1"
+  version  = "1.1.0"
   agent_id = coder_agent.example.id
-  gemini_api_key   = var.gemini_api_key
+  folder   = "/home/coder/project"
-  gemini_model     = "gemini-2.5-pro"
-  install_gemini   = true
-  gemini_version   = "latest"
-  agentapi_version = "latest"
 }
 ```
 
+## Features
+
+- **Interactive AI Assistance**: Run Gemini CLI directly in your terminal for coding help
+- **Automated Task Execution**: Execute coding tasks automatically via AgentAPI integration
+- **Multiple AI Models**: Support for Gemini 2.5 Pro, Flash, and other Google AI models
+- **API Key Integration**: Seamless authentication with Gemini API
+- **MCP Server Integration**: Built-in Coder MCP server for task reporting
+- **Persistent Sessions**: Maintain context across workspace sessions
+
 ## Prerequisites
 
-- You must add the [Coder Login](https://registry.coder.com/modules/coder-login/coder) module to your template
 - Node.js and npm will be installed automatically if not present
+- The [Coder Login](https://registry.coder.com/modules/coder/coder-login) module is required
 
-## Usage Example
+## Examples
 
-- Example 1:
+### Basic setup
 
 ```tf
 variable "gemini_api_key" {
@@ -40,39 +45,97 @@ variable "gemini_api_key" {
 }
 
 module "gemini" {
-  count                     = data.coder_workspace.me.start_count
   source         = "registry.coder.com/coder-labs/gemini/coder"
-  version                   = "1.0.1"
+  version        = "1.1.0"
   agent_id       = coder_agent.example.id
-  gemini_api_key            = var.gemini_api_key # we recommend providing this parameter inorder to have a smoother experience (i.e. no google sign-in)
+  gemini_api_key = var.gemini_api_key
-  gemini_model              = "gemini-2.5-flash"
+  folder         = "/home/coder/project"
-  install_gemini            = true
-  gemini_version            = "latest"
-  gemini_instruction_prompt = "Start every response with `Gemini says:`"
 }
 ```
 
-## How it Works
+This basic setup will:
 
-- **Install**: The module installs Gemini CLI using npm (installs Node.js via NVM if needed)
+- Install Gemini CLI in the workspace
-- **Instruction Prompt**: If `GEMINI_INSTRUCTION_PROMPT` and `GEMINI_START_DIRECTORY` are set, creates the directory (if needed) and writes the prompt to `GEMINI.md`
+- Configure authentication with your API key
-- **Start**: Launches Gemini CLI in the specified directory, wrapped by AgentAPI
+- Set Gemini to run in `/home/coder/project` directory
-- **Environment**: Sets `GEMINI_API_KEY`, `GOOGLE_GENAI_USE_VERTEXAI`, `GEMINI_MODEL` for the CLI (if variables provided)
+- Enable interactive use from the terminal
+- Set up MCP server integration for task reporting
+
+### Automated task execution (Experimental)
+
+> This functionality is in early access and is still evolving.
+> For now, we recommend testing it in a demo or staging environment,
+> rather than deploying to production
+>
+> Learn more in [the Coder documentation](https://coder.com/docs/ai-coder)
+
+```tf
+variable "gemini_api_key" {
+  type        = string
+  description = "Gemini API key"
+  sensitive   = true
+}
+
+module "coder-login" {
+  count    = data.coder_workspace.me.start_count
+  source   = "registry.coder.com/coder/coder-login/coder"
+  version  = "~> 1.0"
+  agent_id = coder_agent.example.id
+}
+
+data "coder_parameter" "ai_prompt" {
+  type        = "string"
+  name        = "AI Prompt"
+  default     = ""
+  description = "Task prompt for automated Gemini execution"
+  mutable     = true
+}
+
+module "gemini" {
+  count                = data.coder_workspace.me.start_count
+  source               = "registry.coder.com/coder-labs/gemini/coder"
+  version              = "1.1.0"
+  agent_id             = coder_agent.example.id
+  gemini_api_key       = var.gemini_api_key
+  gemini_model         = "gemini-2.5-flash"
+  folder               = "/home/coder/project"
+  task_prompt          = data.coder_parameter.ai_prompt.value
+  enable_yolo_mode     = true # Auto-approve all tool calls for automation
+  gemini_system_prompt = <<-EOT
+    You are a helpful coding assistant. Always explain your code changes clearly.
+    YOU MUST REPORT ALL TASKS TO CODER.
+  EOT
+}
+```
+
+> [!WARNING]
+> YOLO mode automatically approves all tool calls without user confirmation. The agent has access to your machine's file system and terminal. Only enable in trusted, isolated environments.
+
+### Using Vertex AI (Enterprise)
+
+For enterprise users who prefer Google's Vertex AI platform:
+
+```tf
+module "gemini" {
+  source         = "registry.coder.com/coder-labs/gemini/coder"
+  version        = "1.1.0"
+  agent_id       = coder_agent.example.id
+  gemini_api_key = var.gemini_api_key
+  folder         = "/home/coder/project"
+  use_vertexai   = true
+}
+```
 
 ## Troubleshooting
 
-- If Gemini CLI is not found, ensure `install_gemini = true` and your API key is valid
+- If Gemini CLI is not found, ensure your API key is valid (`install_gemini` defaults to `true`)
-- Node.js and npm are installed automatically if missing (using NVM)
+- Check logs in `~/.gemini-module/` for install/start output
-- Check logs in `/home/coder/.gemini-module/` for install/start output
+- Use the `gemini_api_key` variable to avoid requiring Google sign-in
-- We highly recommend using the `gemini_api_key` variable, this also ensures smooth tasks running without needing to sign in to Google.
 
-> [!IMPORTANT]
+The module creates log files in the workspace's `~/.gemini-module` directory for debugging purposes.
-> To use tasks with Gemini CLI, ensure you have the `gemini_api_key` variable set, and **you pass the `AI Prompt` Parameter**.
-> By default we inject the "theme": "Default" and "selectedAuthType": "gemini-api-key" to your ~/.gemini/settings.json along with the coder mcp server.
-> In `gemini_instruction_prompt` and `AI Prompt` text we recommend using (\`\`) backticks instead of quotes to avoid escaping issues. Eg: gemini_instruction_prompt = "Start every response with \`Gemini says:\` "
 
 ## References
 
-- [Gemini CLI Documentation](https://ai.google.dev/gemini-api/docs/cli)
+- [Gemini CLI Documentation](https://github.com/google-gemini/gemini-cli/blob/main/docs/index.md)
 - [AgentAPI Documentation](https://github.com/coder/agentapi)
-- [Coder AI Agents Guide](https://coder.com/docs/tutorials/ai-agents)
+- [Coder AI Agents Guide](https://coder.com/docs/ai-coder)

registry/coder-labs/modules/gemini/main.test.ts

@@ -8,7 +8,6 @@ import {
 } from "bun:test";
 import { execContainer, readFileContainer, runTerraformInit } from "~test";
 import {
-  loadTestFile,
   writeExecutable,
   setup as setupUtil,
   execModuleScript,
@@ -54,10 +53,24 @@ const setup = async (props?: SetupProps): Promise<{ id: string }> => {
     agentapiMockScript: props?.agentapiMockScript,
   });
   if (!props?.skipGeminiMock) {
+    const geminiMockContent = `#!/bin/bash
+
+if [[ "$1" == "--version" ]]; then
+  echo "HELLO: $(bash -c env)"
+  echo "gemini version v2.5.0"
+  exit 0
+fi
+
+set -e
+
+while true; do
+    echo "$(date) - gemini-mock"
+    sleep 15
+done`;
     await writeExecutable({
       containerId: id,
       filePath: "/usr/bin/gemini",
-      content: await loadTestFile(import.meta.dir, "gemini-mock.sh"),
+      content: geminiMockContent,
     });
   }
   return { id };
@@ -70,7 +83,7 @@ describe("gemini", async () => {
     await runTerraformInit(import.meta.dir);
   });
 
-  test("happy-path", async () => {
+  test("agent-api", async () => {
     const { id } = await setup();
     await execModuleScript(id);
     await expectAgentAPIStarted(id);
@@ -117,7 +130,7 @@ describe("gemini", async () => {
     await execModuleScript(id);
 
     const resp = await readFileContainer(id, "/home/coder/.gemini-module/agentapi-start.log");
-    expect(resp).toContain("gemini_api_key provided !");
+    expect(resp).toContain("Using direct Gemini API with API key");
   });
 
   test("use-vertexai", async () => {
@@ -197,6 +210,20 @@ describe("gemini", async () => {
     expect(resp).toContain(prompt);
   });
 
+  test("task-prompt", async () => {
+    const taskPrompt = "Create a simple Hello World function";
+    const { id } = await setup({
+      moduleVariables: {
+        task_prompt: taskPrompt,
+      },
+    });
+    await execModuleScript(id, {
+      GEMINI_TASK_PROMPT: taskPrompt,
+    });
+    const resp = await readFileContainer(id, "/home/coder/.gemini-module/agentapi-start.log");
+    expect(resp).toContain("Running automated task:");
+  });
+
   test("start-without-prompt", async () => {
     const { id } = await setup();
     await execModuleScript(id);

registry/coder-labs/modules/gemini/main.tf

@@ -74,14 +74,14 @@ variable "use_vertexai" {
 
 variable "install_agentapi" {
   type        = bool
-  description = "Whether to install AgentAPI."
+  description = "Whether to install AgentAPI for web UI and task automation."
   default     = true
 }
 
 variable "agentapi_version" {
   type        = string
   description = "The version of AgentAPI to install."
-  default     = "v0.3.0"
+  default     = "v0.2.3"
 }
 
 variable "gemini_model" {
@@ -102,12 +102,10 @@ variable "post_install_script" {
   default     = null
 }
 
-data "coder_parameter" "ai_prompt" {
+variable "task_prompt" {
-  type        = "string"
+  type        = string
-  name        = "AI Prompt"
+  description = "Task prompt for automated Gemini execution"
   default     = ""
-  description = "Initial prompt for the Gemini CLI"
-  mutable     = true
 }
 
 variable "additional_extensions" {
@@ -122,12 +120,24 @@ variable "gemini_system_prompt" {
   default     = ""
 }
 
+variable "enable_yolo_mode" {
+  type        = bool
+  description = "Enable YOLO mode to automatically approve all tool calls without user confirmation. Use with caution."
+  default     = false
+}
+
 resource "coder_env" "gemini_api_key" {
   agent_id = var.agent_id
   name     = "GEMINI_API_KEY"
   value    = var.gemini_api_key
 }
 
+resource "coder_env" "google_api_key" {
+  agent_id = var.agent_id
+  name     = "GOOGLE_API_KEY"
+  value    = var.gemini_api_key
+}
+
 resource "coder_env" "gemini_use_vertex_ai" {
   agent_id = var.agent_id
   name     = "GOOGLE_GENAI_USE_VERTEXAI"
@@ -181,21 +191,6 @@ module "agentapi" {
   agentapi_version     = var.agentapi_version
   pre_install_script   = var.pre_install_script
   post_install_script  = var.post_install_script
-  start_script         = <<-EOT
-     #!/bin/bash
-     set -o errexit
-     set -o pipefail
-
-     echo -n '${base64encode(local.start_script)}' | base64 -d > /tmp/start.sh
-     chmod +x /tmp/start.sh
-     GEMINI_API_KEY='${var.gemini_api_key}' \
-     GOOGLE_GENAI_USE_VERTEXAI='${var.use_vertexai}' \
-     GEMINI_MODEL='${var.gemini_model}' \
-     GEMINI_START_DIRECTORY='${var.folder}' \
-     GEMINI_TASK_PROMPT='${base64encode(data.coder_parameter.ai_prompt.value)}' \
-     /tmp/start.sh
-   EOT
-
   install_script       = <<-EOT
     #!/bin/bash
     set -o errexit
@@ -209,7 +204,23 @@ module "agentapi" {
     BASE_EXTENSIONS='${base64encode(replace(local.base_extensions, "'", "'\\''"))}' \
     ADDITIONAL_EXTENSIONS='${base64encode(replace(var.additional_extensions != null ? var.additional_extensions : "", "'", "'\\''"))}' \
     GEMINI_START_DIRECTORY='${var.folder}' \
-    GEMINI_INSTRUCTION_PROMPT='${base64encode(var.gemini_system_prompt)}' \
+    GEMINI_SYSTEM_PROMPT='${base64encode(var.gemini_system_prompt)}' \
     /tmp/install.sh
   EOT
+  start_script         = <<-EOT
+     #!/bin/bash
+     set -o errexit
+     set -o pipefail
+
+     echo -n '${base64encode(local.start_script)}' | base64 -d > /tmp/start.sh
+     chmod +x /tmp/start.sh
+     GEMINI_API_KEY='${var.gemini_api_key}' \
+     GOOGLE_API_KEY='${var.gemini_api_key}' \
+     GOOGLE_GENAI_USE_VERTEXAI='${var.use_vertexai}' \
+     GEMINI_YOLO_MODE='${var.enable_yolo_mode}' \
+     GEMINI_MODEL='${var.gemini_model}' \
+     GEMINI_START_DIRECTORY='${var.folder}' \
+     GEMINI_TASK_PROMPT='${var.task_prompt}' \
+     /tmp/start.sh
+   EOT
 }

registry/coder-labs/modules/gemini/scripts/install.sh

@@ -2,7 +2,6 @@
 
 BOLD='\033[0;1m'
 
-# Function to check if a command exists
 command_exists() {
     command -v "$1" >/dev/null 2>&1
 }
@@ -12,7 +11,7 @@ set -o nounset
 ARG_GEMINI_CONFIG=$(echo -n "$ARG_GEMINI_CONFIG" | base64 -d)
 BASE_EXTENSIONS=$(echo -n "$BASE_EXTENSIONS" | base64 -d)
 ADDITIONAL_EXTENSIONS=$(echo -n "$ADDITIONAL_EXTENSIONS" | base64 -d)
-GEMINI_INSTRUCTION_PROMPT=$(echo -n "$GEMINI_INSTRUCTION_PROMPT" | base64 -d)
+GEMINI_SYSTEM_PROMPT=$(echo -n "$GEMINI_SYSTEM_PROMPT" | base64 -d)
 
 echo "--------------------------------"
 printf "gemini_config: %s\n" "$ARG_GEMINI_CONFIG"
@@ -23,7 +22,6 @@ echo "--------------------------------"
 set +o nounset
 
 function install_node() {
-  # borrowed from claude-code module
     if ! command_exists npm; then
       printf "npm not found, checking for Node.js installation...\n"
       if ! command_exists node; then
@@ -52,24 +50,15 @@ function install_node() {
 
 function install_gemini() {
   if [ "${ARG_INSTALL}" = "true" ]; then
-    # we need node to install and run gemini-cli
     install_node
 
-  # If nvm does not exist, we will create a global npm directory (this os to prevent the possibility of EACCESS issues on npm -g)
   if ! command_exists nvm; then
       printf "which node: %s\n" "$(which node)"
       printf "which npm: %s\n" "$(which npm)"
 
-      # Create a directory for global packages
       mkdir -p "$HOME"/.npm-global
-
-      # Configure npm to use it
       npm config set prefix "$HOME/.npm-global"
-
-      # Add to PATH for current session
       export PATH="$HOME/.npm-global/bin:$PATH"
-
-      # Add to shell profile for future sessions
       if ! grep -q "export PATH=$HOME/.npm-global/bin:\$PATH" ~/.bashrc; then
           echo "export PATH=$HOME/.npm-global/bin:\$PATH" >> ~/.bashrc
       fi
@@ -108,7 +97,6 @@ function append_extensions_to_settings_json() {
     fi
     if [ ! -f "$SETTINGS_PATH" ]; then
       printf "%s does not exist. Creating with merged mcpServers structure.\n" "$SETTINGS_PATH"
-      # If ADDITIONAL_EXTENSIONS is not set or empty, use '{}'
       ADD_EXT_JSON='{}'
       if [ -n "${ADDITIONAL_EXTENSIONS:-}" ]; then
         ADD_EXT_JSON="$ADDITIONAL_EXTENSIONS"
@@ -116,10 +104,7 @@ function append_extensions_to_settings_json() {
       printf '{"mcpServers":%s}\n' "$(jq -s 'add' <(echo "$BASE_EXTENSIONS") <(echo "$ADD_EXT_JSON"))" > "$SETTINGS_PATH"
     fi
 
-    # Prepare temp files
     TMP_SETTINGS=$(mktemp)
-
-    # If ADDITIONAL_EXTENSIONS is not set or empty, use '{}'
     ADD_EXT_JSON='{}'
     if [ -n "${ADDITIONAL_EXTENSIONS:-}" ]; then
       printf "[append_extensions_to_settings_json] ADDITIONAL_EXTENSIONS is set.\n"
@@ -133,14 +118,13 @@ function append_extensions_to_settings_json() {
       '.mcpServers = (.mcpServers // {} + $base + $add)' \
       "$SETTINGS_PATH" > "$TMP_SETTINGS" && mv "$TMP_SETTINGS" "$SETTINGS_PATH"
 
-    # Add theme and selectedAuthType fields
     jq '.theme = "Default" | .selectedAuthType = "gemini-api-key"' "$SETTINGS_PATH" > "$TMP_SETTINGS" && mv "$TMP_SETTINGS" "$SETTINGS_PATH"
 
     printf "[append_extensions_to_settings_json] Merge complete.\n"
 }
 
-function add_instruction_prompt_if_exists() {
+function add_system_prompt_if_exists() {
-    if [ -n "${GEMINI_INSTRUCTION_PROMPT:-}" ]; then
+    if [ -n "${GEMINI_SYSTEM_PROMPT:-}" ]; then
         if [ -d "${GEMINI_START_DIRECTORY}" ]; then
             printf "Directory '%s' exists. Changing to it.\\n" "${GEMINI_START_DIRECTORY}"
             cd "${GEMINI_START_DIRECTORY}" || {
@@ -160,16 +144,21 @@ function add_instruction_prompt_if_exists() {
         fi
         touch GEMINI.md
         printf "Setting GEMINI.md\n"
-        echo "${GEMINI_INSTRUCTION_PROMPT}" > GEMINI.md
+        echo "${GEMINI_SYSTEM_PROMPT}" > GEMINI.md
     else
         printf "GEMINI.md is not set.\n"
     fi
 }
 
+function configure_mcp() {
+    export CODER_MCP_APP_STATUS_SLUG="gemini"
+    export CODER_MCP_AI_AGENTAPI_URL="http://localhost:3284"
+    coder exp mcp configure gemini "${GEMINI_START_DIRECTORY}"
+}
 
-# Install Gemini
 install_gemini
 gemini --version
 populate_settings_json
-add_instruction_prompt_if_exists
+add_system_prompt_if_exists
+configure_mcp
 

registry/coder-labs/modules/gemini/scripts/start.sh

@@ -1,6 +1,7 @@
 #!/bin/bash
+set -o errexit
+set -o pipefail
 
-# Load shell environment
 source "$HOME"/.bashrc
 
 command_exists() {
@@ -15,7 +16,8 @@ fi
 
 printf "Version: %s\n" "$(gemini --version)"
 
-GEMINI_TASK_PROMPT=$(echo -n "$GEMINI_TASK_PROMPT" | base64 -d)
+MODULE_DIR="$HOME/.gemini-module"
+mkdir -p "$MODULE_DIR"
 
 if command_exists gemini; then
     printf "Gemini is installed\n"
@@ -43,20 +45,30 @@ else
 fi
 
 if [ -n "$GEMINI_TASK_PROMPT" ]; then
-    printf "Running the task prompt %s\n" "$GEMINI_TASK_PROMPT"
+    printf "Running automated task: %s\n" "$GEMINI_TASK_PROMPT"
     PROMPT="Every step of the way, report tasks to Coder with proper descriptions and statuses. Your task at hand: $GEMINI_TASK_PROMPT"
+    PROMPT_FILE="$MODULE_DIR/prompt.txt"
+    echo -n "$PROMPT" >"$PROMPT_FILE"
     GEMINI_ARGS=(--prompt-interactive "$PROMPT")
 else
-    printf "No task prompt given.\n"
+    printf "Starting Gemini CLI in interactive mode.\n"
     GEMINI_ARGS=()
 fi
 
-if [ -n "$GEMINI_API_KEY" ]; then
+if [ -n "$GEMINI_YOLO_MODE" ] && [ "$GEMINI_YOLO_MODE" = "true" ]; then
-    printf "gemini_api_key provided !\n"
+    printf "YOLO mode enabled - will auto-approve all tool calls\n"
-else
+    GEMINI_ARGS+=(--yolo)
-    printf "gemini_api_key not provided\n"
 fi
 
-# use low width to fit in the tasks UI sidebar. height is adjusted so that width x height ~= 80x1000 characters
+if [ -n "$GEMINI_API_KEY" ] || [ -n "$GOOGLE_API_KEY" ]; then
-# are visible in the terminal screen by default.
+    if [ -n "$GOOGLE_GENAI_USE_VERTEXAI" ] && [ "$GOOGLE_GENAI_USE_VERTEXAI" = "true" ]; then
-agentapi server --term-width 67 --term-height 1190 -- gemini "${GEMINI_ARGS[@]}"
+        printf "Using Vertex AI with API key\n"
+    else
+        printf "Using direct Gemini API with API key\n"
+    fi
+else
+    printf "No API key provided (neither GEMINI_API_KEY nor GOOGLE_API_KEY)\n"
+fi
+
+agentapi server --term-width 67 --term-height 1190 -- \
+    bash -c "$(printf '%q ' gemini "${GEMINI_ARGS[@]}")"

registry/coder-labs/modules/gemini/testdata/gemini-mock.sh

@@ -1,14 +0,0 @@
-#!/bin/bash
-
-if [[ "$1" == "--version" ]]; then
-  echo "HELLO: $(bash -c env)"
-  echo "gemini version v2.5.0"
-  exit 0
-fi
-
-set -e
-
-while true; do
-    echo "$(date) - gemini-mock"
-    sleep 15
-done