jcarr
/
gemini-cli
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
gemini-cli/docs/core/telemetry.md

11 KiB
Raw Blame History

Gemini CLI Observability Guide

Telemetry provides crucial data about the Gemini CLI's performance, health, and usage. By enabling it, you can monitor operations, debug issues, and optimize tool usage through traces, metrics, and structured logs.

This entire system is built on the OpenTelemetry (OTEL) standard, allowing you to send data to any compatible backend, from your local terminal to a cloud service.

Enabling Telemetry

You can enable telemetry in multiple ways. Configuration is primarily managed via the .gemini/settings.json file and environment variables, but CLI flags can override these settings for a specific session.

A Note on Sandbox Mode: Telemetry is not compatible with sandbox mode at this time. Turn off sandbox mode before enabling telemetry. Tracked in #894.

Order of Precedence:

  1. CLI Flag (--telemetry): These override all other settings for the current session.
  2. Workspace Settings File (.gemini/settings.json): If no CLI flag is used, the telemetry value from this project-specific file is used.
  3. User Settings File (~/.gemini/settings.json): If not set by a flag or workspace settings, the value from this global user file is used.
  4. Default: If telemetry is not configured by a flag or in any settings file, it is disabled.

Add these lines to enable telemetry by in workspace (.gemini/settings.json) or user (~/.gemini/settings.json) settings:

{
  "telemetry": true,
  "sandbox": false
}

Running an OTEL Collector

An OTEL Collector is a service that receives, processes, and exports telemetry data. The CLI sends data using the OTLP/gRPC protocol.

Learn more about OTEL exporter standard configuration in documentation.

Configuration

  1. Install otelcol-contrib or use docker
  1. Create a folder for the OTEL configurations:
mkdir .gemini/otel

Local (Automated Script)

For the most straightforward local setup, use the scripts/local_telemetry.js script. This script automates the entire process of setting up a local telemetry pipeline, including configuring the necessary settings in your .gemini/settings.json file. The script installs otelcol-contrib (The OpenTelemetry Collector) and jaeger (The Jaeger UI for viewing traces). To use it:

  1. Run the Script: Execute the script from the root of the repository:

    ./scripts/local_telemetry.js

    The script will:

    • Download Jaeger and OTEL if needed.
    • Start a local Jaeger instance.
    • Start an OTEL collector configured to receive data from the Gemini CLI.
    • Automatically enable telemetry in your workspace settings.
    • On exit, disable telemetry.

  2. View Traces: Open your web browser and navigate to http://localhost:16686 to access the Jaeger UI. Here you can inspect detailed traces of Gemini CLI operations.

  3. Inspect Logs and Metrics: The script redirects the OTEL collector's output (which includes logs and metrics) to .gemini/otel/collector.log. You can monitor this file to see the raw telemetry data:

    tail -f .gemini/otel/collector.log

  4. Stop the Services: Press Ctrl+C in the terminal where the script is running to stop the OTEL Collector and Jaeger services.

Local (Manual Setup)

1. Create a Configuration File

Create the file .gemini/otel/collector-local.yaml with the following:

cat <<EOF > .gemini/otel/collector-local.yaml
receivers:
  otlp:
    protocols:
      grpc:
        endpoint: "0.0.0.0:4317"

processors:
  batch:
    timeout: 1s

exporters:
  debug:
    verbosity: detailed

service:
  telemetry:
    logs:
      level: "debug"
  pipelines:
    traces:
      receivers: [otlp]
      processors: [batch]
      exporters: [debug]
    metrics:
      receivers: [otlp]
      processors: [batch]
      exporters: [debug]
    logs:
      receivers: [otlp]
      processors: [batch]
      exporters: [debug]
EOF

2. Run the Collector

You can run the collector using docker or using the otelcol-contrib binary directly.

Option 1: Use Docker

This is the simplest method if you have Docker installed.

  1. Run the Collector:

    docker run --rm --name otel-collector-local \
  -p 4317:4317 \
  -v "$(pwd)/.gemini/otel/collector-local.yaml":/etc/otelcol-contrib/config.yaml \
  otel/opentelemetry-collector-contrib:latest

  2. Stop the Collector:

    docker stop otel-collector-local

Option 2: Use otelcol-contrib

Use this method if you prefer not to use Docker.

  1. Run the Collector: Once installed, run the collector with the configuration file you created earlier:

    ./otelcol-contrib --config="$(pwd)/.gemini/otel/collector-local.yaml"

  2. Stop the Collector: Press Ctrl+C in the terminal where the collector is running.

Google Cloud

This setup sends all telemetry to Google Cloud for robust, long-term analysis.

1. Prerequisites

  • A Google Cloud Project ID.
  • APIs Enabled: Cloud Trace, Cloud Monitoring, Cloud Logging.
  • Authentication: A Service Account with the roles Cloud Trace Agent, Monitoring Metric Writer, and Logs Writer. Ensure your environment is authenticated (e.g., via gcloud auth application-default login or a service account key file).

2. Set environment variables

Set the GOOGLE_CLOUD_PROJECT, GOOGLE_CLOUD_LOCATION, and GOOGLE_GENAI_USE_VERTEXAI environment variables:

GOOGLE_CLOUD_PROJECT="YOUR_PROJECT_ID"
GOOGLE_CLOUD_LOCATION="YOUR_PROJECT_LOCATION" # e.g., us-central1
GOOGLE_GENAI_USE_VERTEXAI=true

3. Create a Configuration File

Create .gemini/otel/collector-gcp.yaml:

cat <<EOF > .gemini/otel/collector-gcp.yaml
receivers:
  otlp:
    protocols:
      grpc:
        endpoint: "0.0.0.0:4317"

processors:
  batch:
    timeout: 1s

exporters:
  googlecloud:
    project: "${GOOGLE_CLOUD_PROJECT}"
    metric:
      prefix: "custom.googleapis.com/gemini_cli"
    log:
      default_log_name: "gemini_cli"
  debug:
    verbosity: detailed

service:
  pipelines:
    traces:
      receivers: [otlp]
      exporters: [googlecloud]
    metrics:
      receivers: [otlp]
      exporters: [googlecloud]
    logs:
      receivers: [otlp]
      exporters: [googlecloud]
EOF

4. Run the Collector

You can run the collector for Google Cloud using either Docker or a locally installed otelcol binary.

_Option 1: Use Docker _

This method encapsulates the collector and its dependencies within a container.

  1. Run the Collector: Choose the command that matches your authentication method.

    • If using Application Default Credentials (gcloud auth application-default login):

      docker run --rm --name otel-collector-gcp \
  -p 4317:4317 \
  --user "$(id -u):$(id -g)" \
  -v "$HOME/.config/gcloud/application_default_credentials.json":/etc/gcp/credentials.json:ro \
  -e "GOOGLE_APPLICATION_CREDENTIALS=/etc/gcp/credentials.json" \
  -v "$(pwd)/.gemini/otel/collector-gcp.yaml":/etc/otelcol-contrib/config.yaml \
  otel/opentelemetry-collector-contrib:latest --config /etc/otelcol-contrib/config.yaml

    • If using a Service Account Key File:

      docker run --rm --name otel-collector-gcp \
  -p 4317:4317 \
  -v "/path/to/your/sa-key.json":/etc/gcp/sa-key.json:ro \
  -e "GOOGLE_APPLICATION_CREDENTIALS=/etc/gcp/sa-key.json" \
  -v "$(pwd)/.gemini/otel/collector-gcp.yaml":/etc/otelcol-contrib/config.yaml \
  otel/opentelemetry-collector-contrib:latest --config /etc/otelcol-contrib/config.yaml

  2. Check Status: Your telemetry data will now appear in Google Cloud Trace, Monitoring, and Logging.

  3. Stop the Collector:

    docker stop otel-collector-gcp

Option 2: Use otelcol-contrib

Use this method if you prefer not to use Docker.

  1. Run the Collector:

    ./otelcol-contrib --config="file:$(pwd)/.gemini/otel/collector-gcp.yaml"

  2. Check Status: Your telemetry data will now appear in Google Cloud Trace, Monitoring, and Logging.

  3. Stop the Collector: Press Ctrl+C in the terminal where the collector is running.

Data Reference: Logs & Metrics

A sessionId is included as a common attribute on all logs and metrics.

Logs

These are timestamped records of specific events.

  • gemini_cli.config: Fired once at startup with the CLI's configuration.

    • Attributes:
      • model (string)
      • embedding_model (string)
      • sandbox_enabled (boolean)
      • core_tools_enabled (string)
      • approval_mode (string)
      • api_key_enabled (boolean)
      • vertex_ai_enabled (boolean)
      • code_assist_enabled (boolean)
      • log_user_prompts_enabled (boolean)
      • file_filtering_respect_git_ignore (boolean)
      • debug_mode (boolean)
      • mcp_servers (string)

  • gemini_cli.user_prompt: Fired when a user submits a prompt.

    • Attributes:
      • prompt_length
      • prompt (except if log_user_prompts_enabled is false)

  • gemini_cli.tool_call: Fired for every function call.

    • Attributes:
      • function_name
      • function_args
      • duration_ms
      • success (boolean)
      • decision (string: "accept", "reject", or "modify", optional)
      • error (optional)
      • error_type (optional)

  • gemini_cli.api_request: Fired when making a request to the Gemini API.

    • Attributes:
      • model
      • request_text (optional)

  • gemini_cli.api_error: Fired if the API request fails.

    • Attributes:
      • model
      • error
      • error_type
      • status_code
      • duration_ms

  • gemini_cli.api_response: Fired upon receiving a response from the Gemini API.

    • Attributes:
      • model
      • status_code
      • duration_ms
      • error (optional)
      • input_token_count
      • output_token_count
      • cached_content_token_count
      • thoughts_token_count
      • tool_token_count
      • response_text (optional)

Metrics

These are numerical measurements of behavior over time.

  • gemini_cli.session.count (Counter, Int): Incremented once per CLI startup.

  • gemini_cli.tool.call.count (Counter, Int): Counts tool calls.

    • Attributes:
      • function_name
      • success (boolean)
      • decision (string: "accept", "reject", or "modify", optional)

  • gemini_cli.tool.call.latency (Histogram, ms): Measures tool call latency.

    • Attributes:
      • function_name
      • decision (string: "accept", "reject", or "modify", optional)

  • gemini_cli.api.request.count (Counter, Int): Counts all API requests.

    • Attributes:
      • model
      • status_code
      • error_type (optional)

  • gemini_cli.api.request.latency (Histogram, ms): Measures API request latency.

    • Attributes:
      • model

  • gemini_cli.token.usage (Counter, Int): Counts the number of tokens used.

    • Attributes:
      • model
      • type (string: "input", "output", "thought", "cache", or "tool")