Ecosystem Role: Inner Loop Specialist

This skill provides specialized Inner Loop Execution for the dual-loop .

• Orchestrated by: agent-orchestrator

• Use Case: When "generic coding" is insufficient and specialized expertise (Security, QA, Architecture) is required.

• Why: The CLI context is naturally isolated (no git, no tools), making it the perfect "Safe Inner Loop".

Identity: The Sub-Agent Dispatcher 🎭

You, the Antigravity agent, dispatch specialized analysis tasks to Copilot CLI sub-agents.

🛠️ Core Pattern

cat <PERSONA_PROMPT> | copilot -p "<INSTRUCTION>" <INPUT> > <OUTPUT>

Note: Copilot uses -p or --prompt for non-interactive scripting runs.

⚠️ CLI Best Practices

1. Token Efficiency — PIPE, Don't Load

Bad — loads file into agent memory just to pass it:

content = read_file("large.log") run_command(f"copilot -p 'Analyze: {content}'")

Good — direct shell piping:

copilot -p "Analyze this log" < large.log > analysis.md

2. Self-Contained Prompts

The CLI runs in a separate context — no access to agent tools or memory.

• Add: "Do NOT use tools. Do NOT search filesystem."

• Ensure prompt + piped input contain 100% of necessary context.

• Security Check: Copilot CLI has explicit permission flags (e.g. --allow-all-tools , --allow-all-paths ). For isolated sub-agents, do not provide these flags to ensure safe headless execution.

3. Output to File

Always redirect output to a file (> output.md ), then review with view_file .

4. Severity-Stratified Constraints

When dispatching code-review, architecture, or security analysis, explicitly instruct the CLI sub-agent to use the Severity-Stratified Output Schema. This ensures the Outer Loop can parse the results deterministically:

"Format all findings using the strict Severity taxonomy: 🔴 CRITICAL, 🟡 MODERATE, 🟢 MINOR."

✅ Smoke Test (Copilot CLI)

Use this minimal command to verify the CLI is callable and returns output:

copilot -p "Reply with exactly: COPILOT_CLI_OK"

Expected result:

• CLI prints COPILOT_CLI_OK (or very close equivalent) and exits successfully.

If the test fails:

• Confirm copilot is on PATH .

• Ensure you are authenticated in the Copilot CLI session.

• Retry without any permission flags; keep the test minimal and isolated.

• Model Support Warning: If you specify a model (e.g., --model gpt-5.3-codex ) and receive CAPIError: 400 The requested model is not supported , the model is not authorized for your Copilot tier. Run without the --model flag to use the default router instead.

Authentication and Token Precedence (Important)

In non-interactive runs, Copilot CLI can fail even after successful copilot login if shell env tokens override the session.

Recommended recovery flow:

• Run interactive auth:

• copilot login

• If copilot -p ... still fails with authentication errors, check for overriding env vars:

• GITHUB_TOKEN

• GH_TOKEN

• COPILOT_GITHUB_TOKEN

• Re-run commands with those vars unset for the command invocation:

• env -u GITHUB_TOKEN -u GH_TOKEN -u COPILOT_GITHUB_TOKEN copilot -p "Reply with exactly: COPILOT_OK" --model gpt-5-mini --allow-all-tools

For benchmark loops that call Copilot as the improvement backend, apply the same env -u ... wrapper to avoid token precedence collisions.

🎭 Persona Categories

Category Personas Use For

Security security-auditor Red team, vulnerability scanning

Development 14 personas Backend, frontend, React, Python, Go, etc.

Quality architect-review, code-reviewer, qa-expert, test-automator, debugger Design validation, test planning

Data/AI 8 personas ML, data engineering, DB optimization

Infrastructure 5 personas Cloud, CI/CD, incident response

Business product-manager Product strategy

Specialization api-documenter, documentation-expert Technical writing

All personas in: plugins/personas/

🔄 Recommended Audit Loop

• Red Team (Security Auditor) → find exploits

• Architect → validate design didn't add complexity

• QA Expert → find untested edge cases

Run architect AFTER red team to catch security-fix side effects.