Verified implementation plans from a council of models. Grok searches, DeepSeek and Kimi analyze, GPT-6.1 Sol critiques, Qwen3.8 Max drafts, Gemini 4 Argon judges.
What is the Planner?
Instead of asking one model for a plan, TachiBot runs your task through 7 core steps across 6 labs — up to 12 when the optional debate, UX and responsive phases switch on.
Grok 4.7 · searchgrok_search — grounds the plan in current docs and best practices
Between steps each result is distilled to a short summary so later models don't lose the middle of a long context.
Optional phases
debate: true
Adds a pro/con round before the critique: Grok 4.7 argues for the approach, Gemini 4 Argon argues against.
ux: true
Kimi K3 traces the user flow, Gemini 4 Argon scores it against WCAG 2.1 AA. Auto-detected for UI tasks.
responsive: true
Kimi K3 checks breakpoints, touch targets and reflow. Auto-detected for layout tasks.
i
Why not just ask Claude to make a plan?
A single model can confidently plan around outdated libraries, deprecated APIs, or patterns it invented. The Planner grounds every plan in search data, cross-checks it with models from different labs, and scores confidence first.
Key benefits
Grounded in real data
Current best practices and docs before any analysis.
Multi-perspective
Six labs catch different kinds of errors.
Switchable
Set TACHI_PLANNER_TIER=standard for a cheaper lineup (Qwen3-Coder-Next does the code analysis).
Code-aware
Pass files or code; plans fit your architecture.
Goal-checked
Pass a goal and every checkpoint tests the work against it.
Quality scored
Code quality, security, performance and confidence.
Saved automatically
Plans are written to dokoro as they grow, step by step.
How to use
A coordinator pattern: each call returns one tool to execute. Run it, pass the result back, get the next instruction — so you see every step and can interrupt.
planner.tsts
1planner_maker({
2task:"Add real-time collaboration to the dashboard",
3goal:"Works offline, no data loss on reconnect",
4mode:"start"
5})
6// Returns: instruction to run grok_search
7
8planner_maker({
9task:"Add real-time collaboration to the dashboard",
10mode:"continue",
11step:2,
12prior:{search:"[grok_search output]"}
13})
14// Repeat until isComplete: true
Run the plan
planner_runner executes the plan and stops at fixed checkpoints. Each one is checked by a different model, fed with evidence from your git diff, test results and changed files. list_plans shows recent plans.
0% · Gemini 4 Argonsix deduction questions — are we building the right thing?
10% · Grok 4.7early drift detection — kill a wrong approach before it spreads
25% · GPT-6 Astrastrategy validation — can amend the plan
50% · Qwen3.8 Maxprogress and goal alignment, from a different reasoning style
80% · Kimi K3decomposes the remaining work so nothing is left dangling
100% · GPT-6 Astra + Gemini 4 Argontwo-judge final verdict against the evidence
UI and layout plans also get UX and responsive reviews at checkpoints.