Implement agent-trace standard for event emission and observability

[nkyriakidis]

Describe the feature or problem you'd like to solve

Implement agent-trace standard for event emission and observability

Proposed solution

The GitHub Copilot CLI currently operates largely as a "black box"—users input natural language and receive a command, with limited visibility into the specific context usage or reasoning steps involved.

I propose adopting the agent-trace standard (https://github.com/cursor/agent-trace) to optionally emit structured lifecycle events.

Benefits:

• Interoperability: Allows third-party terminal tools, IDEs, and shells to visualize Copilot's "thinking" process (e.g., creating a visualization of the prompt construction and response parsing).
• Debugging: Helps power users and contributors understand why a specific command was suggested by inspecting the raw trace events (Input -> Reasoning -> Output).
• Auditing: Provides a standardized JSON format for enterprises to log exactly how shell commands are being generated and what context was retrieved.

Example prompts or workflows

Debugging a hallucination: Running gh copilot suggest "deploy to k8s" --trace would output the standard command suggestion to stdout, but write a agent-trace compatible JSON file to a temp directory or stderr. The user can inspect this JSON to see that the model failed to detect the local kubectl context.

IDE/Terminal Integration: A terminal emulator (like Warp or Ghostty) could wrap the Copilot CLI execution. By listening to the agent-trace events, the terminal could display a real-time status indicator ("Analyzing shell history...", "Generating command...") rather than a static spinner.

Audit Logging: A security-conscious developer sets an environment variable COPILOT_TRACE_DIR=/var/log/copilot. Every time they run gh copilot explain, a trace file is saved containing the exact prompt sent and the raw model completion, allowing for post-incident review if a dangerous command was executed.

Additional context

Reference: https://github.com/cursor/agent-trace

Implementation note: Since the CLI is often used in pipes or with eval, trace output should likely be directed to stderr, a specific file path, or a side-channel, ensuring it does not corrupt the standard output intended for execution.