#35 Support GitHub Copilot CLI as a fleet member LLM backend

[Anonymous]

Originally created by: kumaakh

Description

Add support for GitHub Copilot CLI (gh copilot) as an alternative LLM backend for fleet members.

Motivation

GitHub Copilot is widely adopted in enterprise environments where developers already have Copilot licenses through their GitHub org. Supporting it as a fleet backend avoids requiring a separate LLM subscription and leverages existing access. Microsoft is investing heavily in Copilot's agentic capabilities.

Scope

MCP Server (src/)

• Member registration should allow specifying Copilot as the LLM provider
• execute_prompt should route to gh copilot CLI
• provision_auth should handle Copilot auth (tied to gh auth login)
• OS command builders need Copilot CLI equivalents
• Error classification needs Copilot-specific patterns

PM Skill (skills/pm/)

• Model selection logic needs Copilot equivalents (if model tiers exist)
• Templates reference Claude-specific behavior — need Copilot variants

Research needed

• Copilot CLI's headless/non-interactive mode capabilities
• Session/context persistence (equivalent to Claude's --resume)
• Structured output support (equivalent to --output-format json)
• Available model tiers and selection flags
• Rate limits and token caps
• Tool use / agentic capabilities in CLI mode

Dependencies

• Depends on provider abstraction layer from [#26] / [#27]

[Anonymous]

Originally posted by: kumaakh

Technical direction: Before any implementation, the key research question is whether gh copilot supports a headless/non-interactive -p equivalent. The current execute_prompt architecture assumes a CLI flag for passing prompts programmatically (like claude -p or gemini -p).

Research checklist:

• gh copilot suggest --help / gh copilot explain --help — do these support piped/scripted input?
• Session/context persistence — equivalent to Claude's --session-id / --resume?
• Structured output — equivalent to --output-format json?
• Tool use / agentic execution — does gh copilot support multi-turn tool loops or is it single-shot?

If gh copilot is single-shot only, it is not viable as a fleet member backend — the doer-reviewer loop requires multi-turn agentic execution. In that case, the issue should be updated to 'blocked pending Copilot agentic CLI support.'

If it is viable: add CopilotStrategy in src/providers/copilot.ts following the same pattern as src/providers/claude.ts. provision_auth would check gh auth status and prompt gh auth login if needed.