copilot-runtime

Copilot CLI Runtime

Use this skill only inside copilot-cli:copilot-task and copilot-cli:copilot-review subagents.

Invocation Template

Always use a temporary file to pass the prompt — this avoids shell injection when file contents are embedded:

_prompt_file=$(mktemp /tmp/copilot-prompt-XXXXXX.txt)
cat > "$_prompt_file" << 'PROMPT_EOF'
<task_prompt>
PROMPT_EOF
copilot -p "$(cat "$_prompt_file")" \
  --model <model> \
  [--effort <effort>] \
  -s \
  --allow-all \
  [--add-dir <working_dir>]
rm -f "$_prompt_file"

Required Flags

Flag	Purpose
-s / --silent	Only output model response, suppress stats
--allow-all	Grant all permissions for non-interactive mode. Grants Copilot CLI full tool access — do not use in untrusted environments

Model Routing

Model	When	Strengths
gpt-5.4	Default for all real tasks	Edge cases, boundary conditions, error handling
claude-opus-4.6	Caller requests design/architecture work	System design, architecture decisions
gpt-5-mini	Debug only — never for real tasks	Plugin testing

Default --effort high for all real tasks.

Context Injection

⚠️ Privacy notice: Copilot CLI auto-loads ./CLAUDE.md and ./.claude/CLAUDE.md from the working directory and sends them to GitHub Copilot (Microsoft/OpenAI servers). If your CLAUDE.md contains internal system prompts, proprietary constraints, or sensitive business context, run from a clean directory or ensure your CLAUDE.md is safe to share externally.

For additional context (memory, file contents): embed directly in the prompt inside <context> tags.

Self-Containment Checklist

Before invoking, verify the prompt:

• Has clear completion criteria
• Contains all necessary file contents inline
• Has no ambiguity that would trigger clarifying questions
• Specifies output format explicitly
• Can be completed in a single invocation

Timeout Guidelines

Task Type	Timeout
Quick lookup	60000ms
File scanning / exploration	180000ms
Code review / analysis	180000ms
Large batch / architecture	300000ms