pal-mcp-expert_skill

Overview

This skill provides expert guidance for using the Pal MCP Server (zen-pal-nas) to orchestrate multi-model AI workflows, manage tool-based pipelines, and troubleshoot configuration. It focuses on practical patterns for context revival, CLI subagents, model selection, timeouts, and environment configuration to keep multi-step processes reliable and auditable.

How this skill works

It inspects orchestration patterns, tool parameters, environment variables, and continuation_id usage to maintain context across models and sessions. The skill analyzes tool chains (chat, thinkdeep, planner, consensus, codereview, precommit, debug, clink), recommends model routing, and surfaces fixes for common misconfigurations and timeout problems.

When to use it

• Orchestrating multi-model workflows that need consistent context across sessions
• Using or debugging PAL tools like chat, consensus, clink, codereview, precommit, thinkdeep
• Recovering conversation state after Claude resets or compaction
• Optimizing model selection (DEFAULT_MODEL=auto or per-request overrides)
• Configuring environment variables, API keys, or timeouts for MCP

Best practices

• Always reuse the last continuation_id to preserve context between tools and sessions
• Use absolute file paths for code operations and working_directory for generated code
• Prefer DEFAULT_MODEL=auto for intelligent routing, override per request for edge cases
• Set tool_timeout_sec to 1200 seconds for multi-step tools (thinkdeep, codereview, consensus)
• Isolate heavy tasks with clink subagents to avoid polluting the main session context

Example use cases

• Run codereview with gemini-pro, capture continuation_id, then feed it to planner to generate a refactor plan
• Spawn a clink codex codereviewer to audit authentication code in a fresh context and return a concise report
• Use thinkdeep in max mode to investigate a complex race condition with iterative hypothesis testing
• Run consensus across gpt-5-pro (for) and gemini-pro (against) to decide architecture, then continue with clink to implement
• Enable vision-capable models to analyze a system diagram or stack trace screenshot during debug