Skip to content

Choose Engine

phinn edited this page Sep 18, 2026 · 1 revision

🌐 Language: English | 中文

Which engine? V1 / V2 / V3

KinetAios ships three generations of the built-in Kaios engine, switchable per session. All three run the same model with the same tools — what differs is how they work. Picking the right one can easily halve the time a task takes.

30-second cheat sheet

Your task Pick
Q&A, translation, writing, reading a file, fixing one file V1
Tight cost control, want to see every step V1
Multi-file code changes, refactors with verification, multi-step data processing V2
Data analysis, reports, cross-table comparisons V3
Not sure V3 (it grades the task difficulty itself)

V1 (Direct) — light and direct

How it works: a single-threaded ReAct loop; the model calls tools directly, every step visible in real time. Lightweight context strategy — fastest responses, lowest token cost.

Good for:

  • Q&A, explanations, translation, copywriting, emails
  • Reading a file, inspecting code, small fixes
  • Continuous one-question-one-answer conversations
  • Precise control over cost and flow

Not good for:

  • Refactors spanning many files (long conversations drop earlier findings)
  • Multi-step data processing (intermediate results live in conversation memory and get truncated)

Tip: when using V1 for data analysis, invoke /data-analysis to load the analysis discipline — it measurably reduces "mental-math" errors.


V2 — plan first, then execute

How it works: complex tasks enter a planning phase first (read-only exploration producing a step plan), then execute step by step, each step optionally carrying a verification command (typecheck / tests). Failures auto-retry (≤3 per step, ≤2 replans).

Good for:

  • Multi-file code changes and refactors
  • Dev tasks that must pass tests/lint after the change
  • Multi-step data processing, cross-file stats (intermediate results are forced to disk/memory, truncation-proof)
  • Tasks where you want to see the plan before execution starts

Not good for:

  • Simple Q&A, single-file tweaks (planning is wasted overhead — slower and pricier)
  • Large multi-file parallel work (V2 executes serially, step by step)

Tip: V2 auto-degrades to plain mode on trivial tasks — no "nuking mosquitoes" worry. But if you already know the task is simple, V1 is faster.


V3 — adaptive pipeline (recommended default)

How it works: a zero-cost rules-based router grades the task and picks one of three paths:

  • fast: reading files, docs, simple Q&A → single-round direct answer, zero overhead
  • std: bugfix, feature work, data analysis → multi-round tool execution
  • deep: cross-file refactors, architecture-level changes → auto-planned into a DAG; independent steps run in parallel, with built-in verification gates

Good for:

  • Data analysis / reports / statistics — V3 has the analysis discipline built in: schema before conclusions, compute via python/sqlite (never mental math), intermediate results on disk, conclusions must be sourced, key numbers cross-checked. When 「📊 Analysis mode」 shows in the status bar, it's active
  • Tasks with unknown complexity (let the router decide)
  • Complex refactors (deep's parallelism + verification gates are unique among the three)

Not good for:

  • Advanced workflows needing precise per-step control (V1 is more transparent)
  • Minimum-cost single Q&A at all costs (the fast path is cheap, but V1 is slightly cheaper)

Scenario comparison

Scenario Pick Why
"Look at this error for me" V1 or V3 Single-point problem, just be fast
"Move this feature from file A to B, run the tests after" V2 or V3 Code change with verification needs
"Compare these two CSVs and produce a report" V3 Analysis discipline + intermediate results on disk
"Migrate the whole project from JS to TS" V3 (deep) The only one with DAG parallelism + verification gates
"Translate this paragraph" V1 No tool orchestration needed
"Give me a plan first, I'll approve before you act" V2 The planning phase is literally "plan first"

Caveats

  • Switching engines clears cross-engine context: the three engines' history formats are incompatible — don't switch mid-task
  • All three share the same model config, MCP tools, and memory system — switching needs no reconfiguration
  • When data files (csv/xlsx/db) appear in the task description, V3 enters analysis mode automatically; on V1, pair it with /data-analysis

Clone this wiki locally