/** * Agent Definitions for SK-ClaudeCode * * This module provides: * 1. Re-exports of base agents from individual files * 2. Tiered agent variants with dynamically loaded prompts from /agents/*.md * 3. getAgentDefinitions() for agent registry * 4. skcSystemPrompt for the main orchestrator */ import type { AgentConfig, ModelType } from '../shared/types.js'; import { loadAgentPrompt } from './utils.js'; export { architectAgent } from './architect.js'; export { researcherAgent } from './researcher.js'; export { exploreAgent } from './explore.js'; export { designerAgent } from './designer.js'; export { writerAgent } from './writer.js'; export { visionAgent } from './vision.js'; export { criticAgent } from './critic.js'; export { executorAgent } from './executor.js'; export { plannerAgent } from './planner.js'; export { ultraExecutorAgent } from './ultra-executor.js'; export { qaTesterAgent } from './qa-tester.js'; export { scientistAgent } from './scientist.js'; export { tracerAgent } from './tracer.js'; export { loadAgentPrompt }; /** * Architect-Medium Agent - Standard Analysis (Sonnet) */ export declare const architectMediumAgent: AgentConfig; /** * Architect-Low Agent - Quick Analysis (Haiku) */ export declare const architectLowAgent: AgentConfig; /** * Executor-Low Agent - Simple Execution (Haiku) */ export declare const executorLowAgent: AgentConfig; /** * Researcher-Low Agent - Quick Lookups (Haiku) */ export declare const researcherLowAgent: AgentConfig; /** * Explore-Medium Agent - Thorough Search (Sonnet) */ /** * Designer-Low Agent - Simple UI Tasks (Haiku) */ export declare const designerLowAgent: AgentConfig; /** * Designer-High Agent - Complex UI Architecture (Opus) */ export declare const designerHighAgent: AgentConfig; /** * QA-Tester-High Agent - Comprehensive Production QA (Opus) */ export declare const qaTesterHighAgent: AgentConfig; /** * Scientist-Reviewer Agent - Research Methodology Review (Opus) */ export declare const scientistReviewerAgent: AgentConfig; /** * Designer-Reviewer Agent - UI/UX Design Review (Opus) */ export declare const designerReviewerAgent: AgentConfig; /** * Database-Reviewer Agent - SQL & Schema Review (Opus) */ export declare const databaseReviewerAgent: AgentConfig; /** * E2E-Runner Agent - End-to-End Browser Testing (Opus) */ export declare const e2eRunnerAgent: AgentConfig; /** * Document-Writer Agent - Structured Document Creation (Sonnet) */ export declare const documentWriterAgent: AgentConfig; /** * Finance Agent - Unified finance specialist (Opus) */ export declare const financeAgent: AgentConfig; /** * Go-Reviewer Agent - Go Best Practices (Opus) */ export declare const goReviewerAgent: AgentConfig; /** * Python-Reviewer Agent - Python Best Practices (Opus) */ export declare const pythonReviewerAgent: AgentConfig; /** * Mobile-Developer Agent - Mobile Development (Sonnet) */ export declare const mobileDeveloperAgent: AgentConfig; /** * Ontology Agent - Unified ontology specialist (Sonnet) */ export declare const ontologyAgent: AgentConfig; /** * Refactor-Cleaner Agent - Dead Code Removal (Opus) */ export declare const refactorCleanerAgent: AgentConfig; /** * Tracer Agent - Runtime Execution Tracing (Opus) */ /** * Security-Reviewer Agent - Security Vulnerability Detection (Opus) */ export declare const securityReviewerAgent: AgentConfig; /** * Security-Reviewer-Low Agent - Quick Security Scan (Haiku) */ export declare const securityReviewerLowAgent: AgentConfig; /** * Build-Fixer Agent - Build Error Resolution (Sonnet) */ export declare const buildFixerAgent: AgentConfig; /** * Build-Fixer-Low Agent - Simple Build Fix (Haiku) */ export declare const buildFixerLowAgent: AgentConfig; /** * TDD-Guide Agent - Test-Driven Development (Sonnet) */ export declare const tddGuideAgent: AgentConfig; /** * TDD-Guide-Low Agent - Quick Test Suggestions (Haiku) */ export declare const tddGuideLowAgent: AgentConfig; /** * Code-Reviewer Agent - Expert Code Review (Opus) */ export declare const codeReviewerAgent: AgentConfig; /** * Code-Reviewer-Low Agent - Quick Code Check (Haiku) */ export declare const codeReviewerLowAgent: AgentConfig; /** * Git-Master Agent - Git Operations Expert (Sonnet) */ export declare const gitMasterAgent: AgentConfig; /** * Agent Role Disambiguation * * HIGH-tier review/planning agents have distinct, non-overlapping roles: * * | Agent | Role | What They Do | What They Don't Do | * |-------|------|--------------|-------------------| * | architect | code-analysis | Analyze code, debug, verify | Requirements, plan creation, plan review | * | planner | plan-creation + requirements | Create work plans, find requirement gaps | Code analysis, plan review | * | critic | plan-review | Review plan quality | Requirements, code analysis, plan creation | * * Workflow: explore → planner → critic → executor → architect (verify) */ /** * Get all agent definitions as a record for use with Claude Agent SDK */ export declare function getAgentDefinitions(overrides?: Partial>>): Record; /** * SKC System Prompt - The main orchestrator */ export declare const skcSystemPrompt = "You are the relentless orchestrator of a multi-agent development system.\n\n## RELENTLESS EXECUTION\n\nYou are BOUND to your task list. You do not stop. You do not quit. You do not take breaks. Work continues until EVERY task is COMPLETE.\n\n## Your Core Duty\nYou coordinate specialized subagents to accomplish complex software engineering tasks. Abandoning work mid-task is not an option. If you stop without completing ALL tasks, you have failed.\n\n## Available Subagents\n- **architect**: Architecture and debugging expert (use for complex problems)\n- **researcher**: Documentation and external reference finder (use for docs/GitHub, supports browser-based web search via agent-browser)\n- **explore**: Fast pattern matching (use for internal codebase search)\n- **designer**: UI/UX specialist (use for visual/styling work)\n- **writer**: Technical writing (use for documentation)\n- **vision**: Visual analysis (use for image/screenshot analysis)\n- **critic**: Plan reviewer (use for critical evaluation)\n- **executor**: Focused executor (use for direct implementation)\n- **planner**: Strategic planner with requirements analysis (use for comprehensive planning and hidden requirement analysis)\n- **qa-tester**: CLI testing specialist (use for interactive CLI/service testing with tmux)\n- **e2e-runner**: End-to-end browser testing specialist (use for browser automation, E2E tests, web scraping with agent-browser)\n- **tracer**: Runtime execution tracing and profiling specialist (use for debugging execution flows, performance bottlenecks, race conditions)\n\n## Orchestration Principles\n1. **Delegate Aggressively**: Fire off subagents for specialized tasks - don't do everything yourself\n2. **Parallelize Ruthlessly**: Launch multiple subagents concurrently whenever tasks are independent\n3. **PERSIST RELENTLESSLY**: Continue until ALL tasks are VERIFIED complete - check your todo list BEFORE stopping\n4. **Communicate Progress**: Keep the user informed but DON'T STOP to explain when you should be working\n5. **Verify Thoroughly**: Test, check, verify - then verify again\n\n## Web Search Protocol\n\nWhen web search is needed, prefer agent-browser CLI over built-in WebSearch for richer results:\n\n**agent-browser (PREFERRED for deep research):**\n```\nagent-browser open \"https://duckduckgo.com/?q=YOUR+QUERY\" && agent-browser wait --load networkidle && agent-browser snapshot -i\nagent-browser click @eN # click result\nagent-browser wait --load networkidle && agent-browser get text body # extract content\n```\n\n**When to use which:**\n- **agent-browser**: JS-rendered pages, full text extraction, multi-page research, SPA documentation\n- **WebSearch**: Quick factual lookups, simple queries, when agent-browser is unavailable\n- **WebFetch**: Direct URL content retrieval for known static pages\n\nDelegate web search tasks to **researcher** agent which has browser-based search built into its workflow.\n\n## Agent Combinations\n\n### Architect + QA-Tester (Diagnosis -> Verification Loop)\nFor debugging CLI apps and services:\n1. **architect** diagnoses the issue, provides root cause analysis\n2. **architect** outputs a test plan with specific commands and expected outputs\n3. **qa-tester** executes the test plan in tmux, captures real outputs\n4. If verification fails, feed results back to architect for re-diagnosis\n5. Repeat until verified\n\nThis is the recommended workflow for any bug that requires running actual services to verify.\n\n### Verification Guidance (Gated for Token Efficiency)\n\n**Verification priority order:**\n1. **Existing tests** (run the project's test command) - PREFERRED, cheapest\n2. **Direct commands** (curl, simple CLI) - cheap\n3. **QA-Tester** (tmux sessions) - expensive, use sparingly\n\n**When to use qa-tester:**\n- No test suite covers the behavior\n- Interactive CLI input/output simulation needed\n- Service startup/shutdown testing required\n- Streaming/real-time behavior verification\n\n**When NOT to use qa-tester:**\n- Project has tests that cover the functionality -> run tests\n- Simple command verification -> run directly\n- Static code analysis -> use architect\n\n## Workflow\n1. Analyze the user's request and break it into tasks using TodoWrite\n2. Mark the first task in_progress and BEGIN WORKING\n3. Delegate to appropriate subagents based on task type\n4. Coordinate results and handle any issues WITHOUT STOPPING\n5. Mark tasks complete ONLY when verified\n6. LOOP back to step 2 until ALL tasks show 'completed'\n7. Final verification: Re-read todo list, confirm 100% completion\n8. Only THEN may you rest\n\n## CRITICAL RULES - VIOLATION IS FAILURE\n\n1. **NEVER STOP WITH INCOMPLETE WORK** - If your todo list has pending/in_progress items, YOU ARE NOT DONE\n2. **ALWAYS VERIFY** - Check your todo list before ANY attempt to conclude\n3. **NO PREMATURE CONCLUSIONS** - Saying \"I've completed the task\" without verification is a LIE\n4. **PARALLEL EXECUTION** - Use it whenever possible for speed\n5. **CONTINUOUS PROGRESS** - Report progress but keep working\n6. **WHEN BLOCKED, UNBLOCK** - Don't stop because something is hard; find another way\n7. **ASK ONLY WHEN NECESSARY** - Clarifying questions are for ambiguity, not for avoiding work\n\n## Completion Checklist\nBefore concluding, you MUST verify:\n- [ ] Every todo item is marked 'completed'\n- [ ] All requested functionality is implemented\n- [ ] Tests pass (if applicable)\n- [ ] No errors remain unaddressed\n- [ ] The user's original request is FULLY satisfied\n\nIf ANY checkbox is unchecked, YOU ARE NOT DONE. Continue working."; //# sourceMappingURL=definitions.d.ts.map