Multi-Model Planner
Create verified implementation plans using a council of AI models. Grok searches for ground truth, Qwen analyzes code, Kimi reasons step-by-step, GPT critiques gaps, and Gemini scores quality.
What is the Planner?
The Planner is a multi-model council that creates implementation plans through structured phases. Instead of asking one AI to make a plan (which may hallucinate), TachiBot runs your task through 9-13 specialized steps across 6 different AI models.
Each model has a specific role:
- Grok Search — Grounds truth with real-time web data. Prevents the plan from being based on outdated or invented information.
- Qwen Coder — Analyzes code, APIs, dependencies, data structures. When you provide code, it reviews your actual implementation.
- Kimi K2.5 (Reasoning) — Step-by-step systematic reasoning. Traces execution flows and maps exact change sequences.
- Kimi K2.5 (Decomposition) — Breaks the task into structured subtasks with IDs, dependencies, and acceptance criteria using Agent Swarm reasoning. Feeds dependency graph into synthesis.
- GPT-5.2 (Critic) — Adversarial review. Finds holes, gaps, failure modes, security vulnerabilities, and O-complexity issues.
- Qwen Reason (235B) — First synthesis. Merges all analyses into a draft plan with quality scores.
- Gemini 3 Pro (Final Judge) — Creates the actionable final plan with numbered steps, success criteria, complexity estimates, potential blockers, and confidence scores.
Optional phases (auto-detected):
- UX Analysis — Triggered for UI/frontend tasks. Traces user journeys, checks accessibility (WCAG 2.1 AA), scores usability.
- Responsive Design — Triggered for layout tasks. Evaluates mobile/tablet/desktop breakpoints, touch targets, typography scaling.
Why Use the Planner?
Why not just ask Claude to make a plan?
A single model can confidently give you a plan based on outdated libraries, deprecated APIs, or patterns it invented. The Planner grounds every plan in real search data, cross-checks with multiple models, and scores confidence before you write a single line of code.
Key benefits:
- Grounded in real data — Grok searches for current best practices and latest documentation before any analysis begins
- Multi-perspective analysis — 6 models with different strengths catch different types of errors
- Code-aware — Pass your actual code and get plans that work with your existing architecture
- Quality scored — Every plan comes with Code Quality, Security, Performance, and Confidence scores
- Checkpoint verification — Built-in 50% and 100% checkpoints to catch problems early
- Saved automatically — Plans persist to devlog for future reference
How to Use
Creating a Plan
The Planner uses a coordinator pattern: each call returns ONE tool to execute. You run it, pass the result back, and get the next instruction. This lets you see every step and interrupt if needed.
With Code Review
Pass your actual source code and the council will analyze your real implementation instead of guessing at your architecture.
Executing a Plan
Once a plan is created, use planner_runner to execute it step-by-step with built-in verification checkpoints.
Finding Saved Plans
Plan Output Format
What a plan includes
- Numbered steps with clear success criteria
- Complexity estimates (O notation)
- Potential blockers with mitigations
- 50% and 100% checkpoints
- Quality scores: Code Quality, Security, Performance, Confidence, Overall
Example quality scores section:
Tools Reference
The Planner consists of three tools that work together:
| Tool | Purpose | Key Parameters |
|---|
planner_maker | Create plans via multi-model council | task, mode, codeContext, prior |
planner_runner | Execute plans step-by-step | plan, mode, stepNum, checkpoint |
list_plans | Find recent plans | days (default: 7) |
Related
Explore more ways to build with TachiBot