GitHub Copilot CLI session tracing

About

GitHub Copilot can export every agent interaction as an OpenTelemetry trace: one invoke_agent span for the user message, one chat span per model call, and one execute_tool span per tool call. Point that export at Future AGI and each Copilot agent turn shows up as a trace in Observe, with model names, token counts, tool names, and timings. Prompts, responses, and tool arguments are not exported unless you turn content capture on.

This guide covers Copilot CLI 1.0.98 only, using its native OTLP/HTTP exporter. VS Code Copilot Chat is a separate surface and has no Future AGI setup instructions on this page.

See GitHub’s CLI OpenTelemetry reference for the exporter settings.

What you get

One Copilot agent turn becomes one trace. A Copilot CLI turn that lists a folder and reads a file looks like this in Future AGI:

invoke_agent                       conversation id, total tokens, turn count
├── chat gpt-6-luna                model, input/output/cached tokens, finish reason
├── execute_tool glob              tool name, tool call id
├── execute_tool bash bash         tool name, tool call id, exit code
├── chat gpt-6-luna
├── execute_tool view
└── chat gpt-6-luna

Prerequisites

• A Future AGI account and your API key and secret key (Settings > API Keys).
• GitHub Copilot access on your GitHub account. Use an account entitled to Copilot CLI.
• Install the CLI version used by this guide: npm install -g @github/copilot@1.0.98, then copilot login once.

Set up the Copilot CLI

copilot() {
  OTEL_EXPORTER_OTLP_ENDPOINT="https://api.futureagi.com/tracer" \
  OTEL_EXPORTER_OTLP_HEADERS="X-Api-Key=${FI_API_KEY},X-Secret-Key=${FI_SECRET_KEY}" \
  OTEL_RESOURCE_ATTRIBUTES="project_name=my-copilot-sessions,project_type=observe" \
  command copilot --secret-env-vars=FI_API_KEY,FI_SECRET_KEY,OTEL_EXPORTER_OTLP_HEADERS "$@"
}

cd ~/code/my-project
copilot

The repository example examples/coding-agents/github-copilot includes the same function in copilot-otel.sh.

Privacy

By default Copilot exports mostly metadata: model names, token counts, durations, tool names, and IDs. Prompt and tool message bodies are off by default. However, some vendor fields remain free text even when content capture is off, including hook command paths and raw failure text.

Other fields you should know about:

processors:
  transform/copilot-redact:
    error_mode: ignore
    trace_statements:
      - context: span
        statements:
          - delete_key(attributes, "copilot_chat.hook_command")
          - set(status.message, "") where status.message != ""
      - context: spanevent
        statements:
          - delete_key(attributes, "github.copilot.hook.error_message")
          - delete_key(attributes, "exception.message")
          - delete_key(attributes, "exception.stacktrace")

To capture prompt and tool content for one CLI session, run OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=true copilot through the wrapper. Opt in only in a trusted environment.

With content on, spans add gen_ai.input.messages, gen_ai.output.messages, gen_ai.system_instructions, gen_ai.tool.definitions, gen_ai.tool.call.arguments, and gen_ai.tool.call.result. In Future AGI they appear in the span’s Attributes, not in the trace list’s Input and Output columns. They can contain source code, file contents, secrets that appear in your files, and local paths such as your working directory and skill file locations.

What is not available

• Inline completions (ghost text), next edit suggestions, and Copilot code review do not produce these traces.
• The Copilot cloud agent that works on GitHub and opens pull requests is not covered. Its sessions run on GitHub, not on your machine.
• Session grouping. Future AGI’s Sessions view groups traces by the session.id span attribute. If spans contain gen_ai.conversation.id but not session.id, each user message is a separate trace and Sessions stays empty. Filter on gen_ai.conversation.id to follow one conversation.
• Cost in currency. github.copilot.cost is a billing multiplier and github.copilot.nano_aiu counts AI units. Neither is a dollar amount, and Future AGI does not convert them.
• Metrics and events. Copilot also exports OTel metrics. The Future AGI collector accepts traces only, so metric exports are rejected with HTTP 404 (Copilot logs HTTP export failed with status code: 404 at exit). Traces are not affected.
• Span type for invoke_agent. Future AGI classifies chat spans as LLM and execute_tool spans as tools; the invoke_agent root shows type unknown.
• Standalone custom_notification traces. Copilot CLI can export a one-span custom_notification trace for shell completion, idle, or permission notifications. Future AGI shows it as type unknown; it is safe to ignore or filter.
• The local file and SQLite exporters (COPILOT_OTEL_FILE_EXPORTER_PATH) are debugging aids, not the setup path.

This page is about Copilot as an application. It is not a wrapper for GitHub Models or for model calls your own code makes; use a traceAI integration for those.

Troubleshooting

Next steps

• Observe overview to filter, search, and evaluate traces.
• traceAI integrations for tracing your own LLM application code.