feat(skills): Add deep-tutor builtin skill inspired by PR #504

[sisyphus-dev-ai]

Summary

Converts the DeepTutor-inspired structural improvements from PR #504 into a builtin skill called deep-tutor. This allows users to opt-in to these disciplined workflow patterns when needed, rather than changing Sisyphus's default behavior.

Original Inspiration: DeepTutor Repository

Repository: https://github.com/HKUDS/DeepTutor

DeepTutor is an AI-powered personalized learning assistant with sophisticated multi-agent architecture that coordinates specialized agents (RephraseAgent, DecomposeAgent, ManagerAgent, ResearchAgent, NoteAgent, ReportingAgent) across three phases: Planning → Researching → Reporting.

The Five Patterns Adapted from DeepTutor

1. Phase-Based Tool Selection Strategy

DeepTutor Source: research_agent.py#L113-L189

DeepTutor's _generate_tool_phase_guidance() creates explicit phase-based tool selection:

• Phase 1: Basic Exploration (rag_hybrid, rag_naive)
• Phase 2: Deep Mining (paper_search, web_search)
• Phase 3: Completion (run_code)

Our Adaptation: Tool Selection Strategy by Phase

Early (Problem understanding) → grep, glob, read
Middle (Pattern discovery) → lsp_*, ast_grep
Late (Gap filling) → websearch, context7

With tool cost awareness: FREE → CHEAP → EXPENSIVE

2. Sufficiency Check Gate

DeepTutor Source: research_agent.py#L318-L371

DeepTutor's check_sufficiency() evaluates knowledge sufficiency before proceeding to next phase.

Our Adaptation: 5-checkpoint validation before implementation:

• Context (3+ sources?)
• Patterns (understand existing code?)
• Dependencies (all imports identified?)
• Edge Cases (considered?)
• Scope (clearly defined?)

Requires 80% pass rate to proceed.

3. Iteration Limits (Per-Task Guardrails)

DeepTutor Source: research_agent.py#L38-L42

DeepTutor implements max_iterations with early termination when sufficient knowledge is gathered.

Our Adaptation: Hard limits to prevent infinite loops:

• Max 3 fix attempts per task
• Max 5 same-file edits
• Max 2 consecutive failures
• ~15 min time cap per task
• Escalate to Oracle after limit exceeded

4. Request Type Classification

DeepTutor Source: Multi-agent architecture with specialized roles

DeepTutor assigns different agent types based on task requirements.

Our Adaptation: Request classification matrix with Min Parallel Calls:

• Conceptual: 3+ parallel calls
• Implementation: 4+ parallel calls
• Debugging: 4+ parallel calls
• Refactoring: 5+ parallel calls

NOT optional - under-exploration is a failure mode.

5. Dynamic Task Queue with State Management

DeepTutor Source: manager_agent.py

ManagerAgent manages DynamicTopicQueue with state transitions (pending → researching → completed).

Our Adaptation: Parallel execution strategy where explore/librarian agents fire simultaneously via background tasks, not sequentially.

What This Skill Provides

Users can invoke deep-tutor skill when they need structured, disciplined execution:

# Via skill tool
skill("deep-tutor")

# Via slash command
/deep-tutor

The skill enforces:

• ✅ Request classification with minimum parallel calls
• ✅ Cost-effective tool selection (FREE → CHEAP → EXPENSIVE)
• ✅ 5-checkpoint sufficiency validation (80% threshold)
• ✅ Iteration limits (max 3 attempts per task)

Design Decision

Instead of modifying Sisyphus's default behavior (which worried @code-yeongyu in PR #504), this implements these patterns as an opt-in builtin skill:

• ✅ Sisyphus remains flexible by default
• ✅ Users get DeepTutor discipline when they need it
• ✅ No breaking changes to existing workflows
• ✅ Best of both worlds

Files Changed

• src/features/builtin-skills/deep-tutor/SKILL.md - Full skill template with all patterns and references
• src/features/builtin-skills/skills.ts - Registered deep-tutor in builtin skills array

Testing

• ✅ Typecheck passes
• ✅ Build succeeds
• ✅ No breaking changes to existing behavior
• ✅ Proper DeepTutor citations and source code links

Related

Addresses #504 (by converting to builtin skill as requested by @code-yeongyu)

---

Co-authored-by: @gtg7784 (Taegeon Go)

Inspired by: DeepTutor - AI-powered personalized learning assistant

[greptile-apps]

Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.

[cubic-dev-ai]

No issues found across 2 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.