Refactor agent loop and AI initialization for better lifecycle management

[hua-bang]

Summary

This PR refactors the agent loop architecture and AI initialization to improve lifecycle management, add comprehensive instrumentation hooks, and simplify the tool execution flow. The main changes include:

• Lazy AI initialization: Moved from static tool setup to explicit initializeAI() call after skill registry is ready
• Structured loop results: Replaced string returns with LoopResult objects containing turn-by-turn telemetry
• Lifecycle hooks: Added comprehensive LoopHooks for observability (turn start/end, tool execution, text generation, errors)
• Removed unnecessary LLM calls: Eliminated the checkLoopFinish() function that made an extra LLM call every turn; now uses standard ReAct pattern (no tool calls = done)
• Improved CLI UX: Added slash command support for direct skill invocation (/skill-name), skill listing (/skills), and better formatted output with timing/metrics

Key Changes

AI & Loop Architecture

• ai.ts: Removed static BuiltinToolsMap initialization; added initializeAI(tools) function that must be called after skill registry setup
• loop.ts: Complete rewrite with structured LoopResult return type, LoopHooks for lifecycle events, and removal of the extra checkLoopFinish() LLM call
• core.ts: Updated initialization sequence: registry → build tools → initialize AI; added slash command parsing and skill invocation
• Removed generateObject() function (was unused and relied on tool-calling for structured output)

Tools & Skills

• tools/index.ts: Introduced buildTools() factory function; separated static tools from dynamic skill tool
• tools/skill.ts: Changed from static export to createSkillTool() factory; now reads from initialized registry with fuzzy search fallback
• prompt/system.ts: Added buildSkillsSection() to dynamically include loaded skills in system prompt

Type System

• typings/index.ts: Added LoopResult, TurnInfo, ToolCallInfo, LoopHooks, and LoopOptions interfaces for better type safety and observability
• Added abortSignal to Context for external loop cancellation

CLI Improvements

• Slash command support: /skill-name [args] invokes skills directly
• /skills command lists all loaded skills
• Better output formatting with turn count, tool call count, duration, and finish reason
• Per-turn and per-tool lifecycle hooks for detailed logging

Implementation Details

• The loop now follows the standard ReAct pattern: if the model returns text without tool calls, the task is complete
• Tool execution is sequential with hooks fired before/after each call for observability
• Error handling tracks consecutive errors and stops after MAX_ERROR_COUNT
• All timing is captured at turn and tool granularity for performance analysis
• Skill tool uses exact match first, then fuzzy search, with helpful error messages listing available skills

https://claude.ai/code/session_01Q5TnLskGfMzuS7VZLLXRCS