菜单
Skill for making programmatic LLM calls via Copilot CLI. Covers correct invocation patterns, model configuration, caching, and common error recovery. Use when any agent or script needs to call an LLM programmatically.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Skill for using Beads (bd) as the single source of truth for task tracking, dependency management, and agent coordination. Covers all bd commands needed for multi-agent orchestration.
Skill for working with artifact contracts — the file-level dependency system that prevents pipeline failures by ensuring tasks produce expected outputs and only start when required inputs exist.
Skill for generating scientific figures from experimental data. Ensures every figure referenced in LaTeX documents exists on disk before compilation, preventing blank PDFs and broken references.
Complete recipe for dispatching, monitoring, merging, and cleaning up worker agents. Read this skill BEFORE dispatching any worker. Covers build_worker_prompt, spawn-agent.sh, merge-agent.sh, checkpoint updates, and common failure modes.
Use when asked to reduce code complexity, simplify code, refactor safely, merge micro-files, remove dead code, shrink functions, reduce coupling, clean modules, or improve maintainability in src/voronoi/.
Use for adversarial review of Voronoi plans, findings, claims, manuscripts, and final deliverables. Covers reviewer-style objections, confounds, alternative explanations, evidence gaps, and falsification-oriented critique.
| name | copilot-cli-usage |
| description | Skill for making programmatic LLM calls via Copilot CLI. Covers correct invocation patterns, model configuration, caching, and common error recovery. Use when any agent or script needs to call an LLM programmatically. |
| user-invocable | true |
| disable-model-invocation | false |
Use this skill when scripts or agents need to make programmatic LLM calls.
# 1. Read configured model from workspace config
LLM_MODEL=$(jq -r '.worker_model // ""' .swarm-config.json 2>/dev/null)
MODEL_FLAG=""
if [[ -n "$LLM_MODEL" ]]; then MODEL_FLAG="--model $LLM_MODEL"; fi
# 2. Call Copilot with the prompt as a direct argument
copilot $MODEL_FLAG -p "<your prompt here>" -s --no-color --allow-all
| Rule | Why |
|---|---|
Prompt is a direct argument to -p | Stdin/pipe patterns produce empty responses |
NEVER echo "..." | copilot -p - | Pipe/stdin = broken |
ALWAYS include --model from config | Without it, copilot uses its default model |
| Cache by SHA-256 of prompt text | Avoid redundant API calls |
| Prompt is a single shell-quoted string | Prevents shell expansion issues |
# WRONG — pipe to stdin
echo "Analyze this data" | copilot -p -
# WRONG — missing model flag
copilot -p "Analyze this data" -s --no-color
# WRONG — unquoted prompt
copilot -p Analyze this data -s --no-color
# RIGHT
copilot $MODEL_FLAG -p "Analyze this data" -s --no-color --allow-all
Cache LLM responses to avoid redundant API calls:
PROMPT="Your prompt text here"
CACHE_KEY=$(echo -n "$PROMPT" | shasum -a 256 | cut -d' ' -f1)
CACHE_FILE=".cache/llm/${CACHE_KEY}.txt"
if [[ -f "$CACHE_FILE" ]]; then
RESPONSE=$(cat "$CACHE_FILE")
else
mkdir -p .cache/llm
RESPONSE=$(copilot $MODEL_FLAG -p "$PROMPT" -s --no-color --allow-all)
echo "$RESPONSE" > "$CACHE_FILE"
fi
| Error | Fix |
|---|---|
| Empty response | Check --model flag is set; verify prompt is passed as arg, not pipe |
| Auth failure | Check COPILOT_HOME and GH_HOST env vars; run copilot /login |
| Rate limit | Wait and retry with exponential backoff; check cache first |
| Timeout | Reduce prompt size; split into sub-prompts |
For experiments requiring multiple LLM calls:
.swarm/events.jsonl for audit trail.