메뉴
Copilot CLI sub-agent system for dispatching tasks and persona-based analysis to GitHub Copilot models. Use for task delegation (agent reads/writes files directly), security audits, architecture reviews, or any work requiring a fresh model context.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
(Industry standard: Parallel Agent) Primary Use Case: Work that can be partitioned into independent sub-tasks running concurrently across multiple agents. Parallel multi-agent execution pattern. Use when: work can be partitioned into independent tasks that N agents can execute simultaneously across worktrees. Includes routing (sequential vs parallel), merge verification, and correction loops.
(Industry standard: Sequential Agent / Agent as a Tool) Primary Use Case: Delegating a well-defined task to a worker agent, verifying its execution, and repeating if necessary. Inner/outer agent delegation pattern. Use when: work needs to be delegated from a strategic controller (Outer Loop) to a tactical executor (Inner Loop) via strategy packets, with verification and correction loops.
(Industry standard: Loop Agent / Single Agent) Primary Use Case: Self-contained research, content generation, and exploration where no inner delegation is required. Self-directed research and knowledge capture loop. Use when: starting a session (Orientation), performing research (Synthesis), or closing a session (Seal, Persist, Retrospective). Ensures knowledge survives across isolated agent sessions.
(Industry standard: Routing Agent / Orchestrator Pattern) Primary Use Case: Analyzing an ambiguous trigger and routing it to one of the specific specialized implementations. Routes triggers to the appropriate agent-loop pattern. Use when: assessing a task, research need, or work assignment and deciding whether to run a simple learning loop, red team review, dual-loop delegation, or parallel swarm. Manages shared closure (seal, persist, retrospective, self-improvement).
(Industry standard: Review and Critique Pattern) Primary Use Case: Iterative generation paired with adversarial review, continuing until an 'Approved' verdict is reached. Orchestrated adversarial review loop. Use when: research, designs, architectures, or decisions need to be reviewed by red team agents (human, browser, or CLI). Iterates in rounds of research → bundle → review → feedback until approved.
(Industry standard: Meta-Learning System / Automated Autoresearch) Primary Use Case: Continuous, self-improving orchestration of an agentic system over multiple sessions. Use when: building a continuous improvement layer that autonomously identifies workflow friction, postulates hypotheses, and tests improved instructions/coding skills against an objective headless benchmark before merging and persisting.
| name | copilot-cli-agent |
| plugin | cli-agents |
| description | Copilot CLI sub-agent system for dispatching tasks and persona-based analysis to GitHub Copilot models. Use for task delegation (agent reads/writes files directly), security audits, architecture reviews, or any work requiring a fresh model context. |
| allowed-tools | Bash, Read, Write |
You, the Antigravity agent, dispatch specialized analysis tasks to Copilot CLI sub-agents.
[!IMPORTANT] Default model:
gpt-5-mini(included — no credit cost). As of June 1, 2026, GitHub moves to AI Credits (per-token billing), butgpt-5-miniis an included model that does not consume credits. Higher-multiplier models (e.g.,claude-sonnet-4.6) do consume credits — see 💰 AI Credits & Cost Discipline before using them.
To ensure Copilot CLI behaves as a specialized persona rather than a generic responder, always embed the persona and source material directly into the prompt flag (-p).
copilot -p "$(cat agents/persona.md)
---SOURCE CODE---
$(cat target.py)
---INSTRUCTION---
Perform a full code review. Use severity levels: 🔴 CRITICAL, 🟡 MODERATE, 🟢 MINOR.
You are operating as an isolated sub-agent.
Do NOT use tools. Do NOT access filesystem." > review.md
run_agent.py (Cross-Platform)For reusable sub-agent execution, use the provided Python orchestrator which handles temp file assembly and prompt concatenation reliably across Windows, macOS, and Linux.
# Signature:
python ./scripts/run_agent.py <PERSONA_FILE> <INPUT_FILE> <OUTPUT_FILE> "<INSTRUCTION>" [MODEL] [isolated]
# ^ ^
# optional optional (default: false)
Task dispatch (default — agent has full filesystem access via --yolo):
# Agent reads/writes files directly. Pass the task prompt as INPUT_FILE.
python plugins/cli-agents/scripts/run_agent.py \
/dev/null \
tasks/todo/copilot_prompt_0025.md \
temp/copilot_output_0025.md \
"Implement all changes specified in the prompt." \
claude-sonnet-4.6
Isolated analysis (no filesystem tools — text output only):
# Pass isolated=true as 6th arg. Agent generates text output only.
python plugins/cli-agents/scripts/run_agent.py \
agents/security-auditor.md target.py security.md \
"Find vulnerabilities." gpt-5-mini true
run_agent.py)| Inputs present | Assembled prompt |
|---|---|
| persona + input | persona / ---SOURCE--- input / ---INSTRUCTION--- instruction |
| input only (task dispatch) | input / ---INSTRUCTION--- instruction |
| instruction only (heartbeat) | instruction |
Passing /dev/null for persona or input skips that block cleanly.
gpt-5-mini (included — no credit cost)# No model arg = gpt-5-mini (included model, does not consume credits)
python ./scripts/run_agent.py agents/security-auditor.md target.py security.md \
"Find vulnerabilities."
claude-sonnet-4.6 (9x multiplier — batch everything)# Pass model name as the 5th argument to override the default
python ./scripts/run_agent.py /dev/null /tmp/copilot_prompt.md /tmp/copilot_output.md \
"Generate all files exactly as specified using ===FILE:=== delimiters." \
claude-sonnet-4.6
[!NOTE] When to use
claude-sonnet-4.6: Complex multi-file generation, nuanced content requiring reasoning, tasks where output quality matters more than cost. See 💰 AI Credits & Cost Discipline for batching rules before calling.
| Model | Identifier | Credit cost |
|---|---|---|
| GPT-5 mini | gpt-5-mini | Included (no credits) |
| GPT-4.1 | gpt-4.1 | Included (deprecation upcoming) |
| GPT-4o | gpt-4o | Included |
| Claude Haiku 4.5 | claude-haiku-4.5 | Low |
| GPT-5.4 | gpt-5.4 | 6x |
| GPT-5.4 mini | gpt-5.4-mini | 6x |
| Claude Sonnet 4.5 | claude-sonnet-4.5 | 9x |
| Claude Sonnet 4.6 | claude-sonnet-4.6 | 9x |
| Gemini 3.5 Flash | gemini-3.5-flash | (verify current) |
| Claude Opus 4.7 | claude-opus-4.7 | 27x |
| Claude Opus 4.8 | claude-opus-4.8 | 27x |
[!WARNING] Model identifiers use dots not dashes —
claude-sonnet-4.6notclaude-sonnet-4-6. Identifiers change with Copilot CLI updates — runcopilot -i "list models"or check the interactive model selector to confirm current names before any high-multiplier run.Deprecated (do not use):
claude-sonnet-4(deprecated May 7),gpt-5.2/gpt-5.2-codex(deprecated). GPT-4.1 deprecation upcoming.
agents/)| Persona | Use For |
|---|---|
security-auditor.md | Red team, vulnerability scanning, threat modeling |
refactor-expert.md | Optimizing code for readability, performance, and DRY |
architect-review.md | Assessing system design, modularity, and complexity |
Using cat code.py | copilot -p "review this" is unreliable. The CLI often prioritizes the prompt flag and ignores the piped input. Always embed the code inside the command string as shown in the Core Pattern.
Large prompt expansions (e.g., $(cat ...) > 10KB) can silently fail when run in the background (&).
run_agent.py).wc -l.Always add these instructions to your dispatch prompt to prevent the sub-agent from attempting to use external tools. Strictly use gpt-5-mini as the default model.
To dramatically improve review results, add:
"Think step-by-step internally, but output only final results. Be strict and critical. Do not be polite."
[!CAUTION] As of June 1, 2026, GitHub Copilot moves to AI Credits (per-token billing). 1 Credit = $0.01. Credits do NOT roll over monthly. When credits are exhausted, Copilot stops — there is no fallback model. Set your additional-spend cap to $0 in GitHub billing settings to hard-stop at your monthly allotment.
gpt-5-mini,gpt-4.1, andgpt-4oremain included models (no credit cost). Annual plan subscribers stay on PRU-based pricing until their plan expires, then roll to the new Credit model. Code completions and Next Edit Suggestions never consume credits.
| Tier | Credits/month | Overage |
|---|---|---|
| Copilot Pro ($10/mo) | 1,000 | Optional, per credit |
| Copilot Pro+ ($39/mo) | 3,900 | Optional, per credit |
| Business ($19/user/mo) | $19 worth (pooled) | Optional |
| Model | Cost | Use when |
|---|---|---|
gpt-5-mini, gpt-4.1, gpt-4o | Included | Default — always prefer |
claude-haiku-4.5 | Low credits | When Claude quality needed cheaply |
gpt-5.4, gpt-5.4-mini | Moderate credits | Moderate quality step-up |
claude-sonnet-4.6 | High credits | Complex reasoning, multi-file generation |
claude-opus-4.7, claude-opus-4.8 | Highest credits | Critical tasks only |
===FILE: [relative/path/to/file]===
[complete file content]
===ENDFILE===
===FILE:=== markers in your prompt — confirm the same count appears in output before parsing.gpt-5-mini. Run the heartbeat against the cheapest model only — it verifies connectivity without burning credits.&) high-multiplier calls. Large prompts can silently produce empty output in background processes. Run foreground and verify with wc -l (expect 200+ lines for multi-file output).# Write the full multi-file prompt to a temp file first
cat > /tmp/copilot_prompt.md << 'PROMPT_EOF'
[Your complete, dense, multi-file generation prompt]
PROMPT_EOF
# Dispatch ONE request — all output in a single call
python ./scripts/run_agent.py \
/dev/null \
/tmp/copilot_prompt.md \
/tmp/copilot_output.md \
"Generate all files exactly as specified using ===FILE:=== delimiters." \
claude-sonnet-4.6
# Verify output is substantial before parsing
wc -l /tmp/copilot_output.md # expect 200+ lines for multi-file output
# Confirm all expected FILE markers are present before assuming success
grep -c '===FILE:' /tmp/copilot_output.md # should equal your expected file count
Before initiating major orchestrations or long-running iterative loops (e.g., Triple-Loop), you MUST perform a zero-shot heartbeat check to verify the host CLI has end-to-end connectivity and correct model defaults.
python .agents/skills/copilot-cli-agent/scripts/run_agent.py \
/dev/null /dev/null ./HEARTBEAT_MD.md \
"HEARTBEAT CHECK: Respond with 'HEARTBEAT_OK' only."
# Verification Logic:
[ -s ./HEARTBEAT_MD.md ] && grep -q "HEARTBEAT_OK" ./HEARTBEAT_MD.md && echo "HEARTBEAT_OK" || echo "HEARTBEAT_FAIL"
Logging Requirement: The result of this heartbeat (Success or Failure) MUST be explicitly written to the session log before proceeding. If it fails, halt execution and report the error details (e.g., 401 Unauthorized, 429 Rate Limit, or Network Error).
python ./scripts/run_agent.py agents/refactor-expert.md target.py output.md "Refactor this code."
Examine output.md. It should contain ONLY the refactored code and a brief 3-bullet summary.
claude-sonnet-4.6 works; claude-sonnet-4-6 returns "model not available". Always use dot notation for Claude version numbers in Copilot CLI.copilot --yolo --model <id> -p "HEARTBEAT_OK" first — if it echoes back any response, the identifier is valid. Do not assume identifiers from docs or memory are current.run_agent.py passes the 5th argument directly to --model. If the identifier is wrong, the script exits with a non-zero code and produces no output file. Check exit code and output file size before claiming success.&) a premium model call. Run foreground and verify with wc -l output.md — expect 200+ lines for multi-file output.gpt-5-mini, not a credit-consuming model. It's an included model — connectivity checks cost nothing.permissions.disableBypassPermissionsMode setting (v1.0.55+) — can lock out --yolo/allow-all mode. If headless dispatch stops working, check this setting hasn't been set to true by an org admin..agents/ level. Nested layouts work./mcp and /context — use this to audit which MCP tools are consuming credits.