06 · Work modes and model tiers
The previous page listed six controls. This one takes the two that matter most for both quality and cost: mode and model.
Three work modes
| Mode | What it does | When |
|---|---|---|
| Default (Agent) | Executes the instruction directly, may modify files | Goal is clear, one sentence is enough |
| Plan (计划) | Gathers missing information, produces a reviewable plan, waits for you | Multi-step, vague, or expensive to redo |
| Ask (仅问答) | Answers only; no execution, no file changes | Understanding, clarifying, pure consultation |
Plan mode
Plan means produce the approach first, then act. With it on, the model does not jump to a final result — it asks you for the missing pieces, hands back a complete plan, and starts only once you confirm.
The value is direct: correct the direction in thirty seconds, instead of discovering the thinking was wrong after the whole run finishes.
When to use it:
| Turn on Plan | Default is faster |
|---|---|
| The task needs several steps | A single question: a number, a definition, a translation |
| The requirement is vague, several solutions exist | A clear, small change |
| Key assumptions are still unsettled | You know exactly what you want and can say it in one line |
| The output is expensive, you do not want a wasted run | Light follow-ups within a conversation |
The flow is four steps:
- Turn on Plan in the
+menu. - Answer its questions. It asks about the key variables — timing, budget, preferences. Answer them rather than skipping: more input, better plan. If you skip everything it makes assumptions and labels them in the plan.
- Review the plan in the right-hand panel.
- Confirm. "Start" executes strictly against the agreed plan; "adjust" lets you say what to change (cut / narrow / add / reorder / restart) and it re-outputs. Nothing executes until you confirm.
Plan saves money
It looks like an extra round, but it is cheaper. A complex task that runs in the wrong direction wastes far more than one round of plan review. The more expensive the output, the better the trade.
Ask mode
No execution, no file changes, just answers. Two good uses: exploring an unfamiliar area before deciding what to do, and asking it to evaluate a plan when you absolutely do not want it editing files on the side.
Choosing a model tier
| Tier | For | Character |
|---|---|---|
| Fast | Simple, well-defined work: everyday questions, format conversion, small code edits, copy polishing | Quickest, cheapest |
| Balanced | Most work: writing documents, analysing data, multi-file code changes | Both decent and quick — use this when unsure |
| Maximum | Hard work: cross-module refactors, multi-step reasoning, analysis needing repeated verification | Slower, pricier, but more reliable when the task is genuinely difficult |
In practice: park on balanced. If a class of task consistently overthinks and drags, go down a tier. If it consistently misses details and needs three rounds of rework, go up.
Thinking effort
How much reasoning the model invests before answering. The available levels depend on what the model supports.
Higher effort means more analysis and self-checking — good for long chains of logic where errors compound. Lower effort answers directly and faster.
Usually leave it alone
The default already sits where most tasks want it. Change it only when you have actually observed overthinking or under-thinking.
Context window
Defaults to 200K, which suits nearly everything.
The billing model is the thing to understand: every turn sends the full current context to the model. A bigger context means more tokens per turn, and therefore more credits per turn.
So bigger is not better — going to 1M does not make the model smarter, it just multiplies what each turn costs.
Two situations justify raising it: genuinely long documents, and needing the model to hold a lot of history verbatim. Otherwise leave it at 200K.
Max mode
| Off (default) | On | |
|---|---|---|
| Context handling | Compressed automatically at sensible points | Preserved in full, never compressed |
| Detail retention | Some repeated context is dropped in long chats | Nothing is lost |
| Credit burn | Lower | Noticeably faster in long conversations |
Leave it off unless you genuinely need very long context. Running a dozens-of-turns conversation with Max on is one of the fastest ways to burn a month's credits.
The bundled model list
| Model | Notes | Capabilities |
|---|---|---|
| Hy4 preview | New-generation Hunyuan | Image input · thinking mode |
| Hy3 | Hunyuan reasoning model | Image input · thinking mode |
| GLM-5.3 | New-generation GLM, 1M context | Image input · thinking mode |
| GLM-5.3-Flash | Fast variant of GLM-5.3 | Thinking mode |
| GLM-5.2 | 1M context, strong on long-horizon work | Image input · thinking mode |
| GLM-5.1 | Well-rounded, good for daily use | Reasoning model |
| GLM-5v-Turbo | Natively multimodal | Image input · reasoning model |
| MiniMax-M3 | Natively multimodal, strong at code and agent work | Image input · reasoning model |
| MiniMax-m2.7 | Well-rounded, good value | Image input · reasoning model |
| Kimi-K3 | Strong on complex long-horizon autonomy, notable front-end ability | Image input · reasoning model |
| Kimi-K2.7-Code | Multimodal, tuned for programming | Image input · reasoning model |
| Kimi-K2.6 | Multimodal, good for daily work | Image input · reasoning model |
| Deepseek-V4-Flash | 1M context, speed first | Thinking mode |
| Deepseek-V4-Pro | 1M context | Thinking mode |
This list changes
It is a snapshot of the official docs at one point in time; the product moves fast, so trust what your client shows.
The selection logic outlasts the specific names: check two things — does it take image input, and is it a reasoning model. The first decides whether it can read screenshots; the second decides whether it survives multi-step work.
Switching model is the best debugging move
This appears again and again in the official FAQ and error-code docs: garbled replies, long silences, a bot that stops answering, error codes 1001 / 6003 / 6004 / 11133 / 14003 — the first response to all of them is switch model and retry.
These usually come from transient state on the model side, unrelated to your prompt, your network or your account. Pick a different one and run it again; that generally clears it.
Custom models
If you have your own model resources, connect them under Settings → Models through the GUI — no config file editing required.
Three routes:
- By provider — pick from the list (Tencent Cloud Token Plan, various Coding Plans, presets for the major vendors) and the URL, model list and capability flags fill in automatically; you only supply an API key
- Local deployment — connect Ollama. Once installed it listens on port 11434 and exposes an OpenAI-compatible interface. Good for data that must not leave the machine, for zero-cost experimentation, and for offline environments
- Custom — service not in the list; enter the URL, key and model name yourself
API keys stay local
Configuration including API keys is saved only in the local workbuddy/models.json and is not uploaded. WorkBuddy acts purely as a transport, forwarding your input to the model you configured.
Correspondingly, all costs from a custom model — tokens, subscriptions — are paid by you to that third party and do not go through credits. Which means: that spend never shows up on your credit statement, so watch the third-party account yourself.
Fuller setup notes are in MCP and plugins.
Next
Configuration done. 07 Run your first task.