multi-llm

Multi-LLM Skill

Skill directory & path resolution — read this first.

Every orchestrator script, instruction file, prompt, schema, and template this skill uses is bundled inside the skill's own directory, whose absolute path is:

${CLAUDE_SKILL_DIR}

Claude Code expands ${CLAUDE_SKILL_DIR} to that absolute path inside this SKILL.md before you read it, so every command and path shown below is already fully resolved — run them as written.

The mode instruction files you open with the Read tool, and any "run this next" commands the orchestrators print to stdout, are not pre-expanded: they still contain the literal placeholder CLAUDE_SKILL_DIR (written as the shell-style variable ${...}). Whenever you encounter that placeholder in a file you read or in script output, substitute the absolute skill-directory path shown above before running the command or reading the file. The shell does not export this variable, so never run a command that still contains an unexpanded CLAUDE_SKILL_DIR.

Always double-quote the substituted skill path in shell commands — write --project "${CLAUDE_SKILL_DIR}" and "${CLAUDE_SKILL_DIR}/script.py", never the unquoted form. The path may contain spaces (e.g. Windows user profiles like C:\Users\John Smith\...), and an unquoted expansion word-splits and breaks the command.

A unified skill for multi-LLM plan automation. Supports eleven workflow modes plus a status command:

1. Review Plan (--review-plan): Review an implementation plan with multiple LLMs (default)
2. Apply Suggestions (--apply-suggestions): Apply validated suggestions from review to the plan
3. Generate Tasks (--generate-tasks): Generate detailed implementation tasks from a high-level plan
4. Review Tasks (--review-tasks): Review generated tasks with multiple LLMs
5. Apply Task Suggestions (--apply-task-suggestions): Apply validated task review suggestions to tasks.md
6. Implement (--implement): Execute implementation tasks from a plan
7. Review Code (--review-code): Review code changes against the plan
8. Apply Code Fixes (--apply-code-fixes): Apply validated fixes from code review
9. Full Workflow (--full): Run all modes in sequence
10. Status (--status): Show current workflow state and suggested next action
11. Ask (--ask): Ask each model a free-text question about a plan; aggregate answers into one markdown file
12. Init Config (--init): Set up a per-project provider config override at <git-root>/.multi-llm/providers.yaml (no plan path; routed via instructions/init-config.md). Fully automatic and zero-prompt: it auto-detects which provider CLIs are installed on PATH and writes a preconfigured override (uncommenting the detected providers' blocks and default_provider); --template-only skips detection and writes the inert commented stub. Flags: --dir PATH, --force, --gitignore, --template-only.