[Feature Request]: Adjustable thinking strength & context length for DeepSeek and other reasoning models

[lidao2026-Coding]

Is your feature request related to a problem?

Yes. When using DeepSeek (and other reasoning-capable models) through DeepCode, I cannot adjust the thinking/reasoning strength (effort) or the context length / max context window from the Desktop GUI (nor clearly from the CLI). These are the most basic model controls for a coding agent, and their absence is very disappointing for daily use.

Concretely:

• Reasoning effort: The Desktop composer has a "Thinking" selector, but for DeepSeek (and similar providers) there is no exposed/working way to choose the thinking level (low/medium/high) — the option either doesn't appear or has no effect on the actual request.
• Context length: There is no UI to control the max context window / token budget used per turn (e.g. 32K / 64K / 128K), which matters on DeepSeek's 64K/128K contexts when working on large repos.

Describe the solution you'd like

1. Per-model thinking strength control: For providers that support reasoning effort (DeepSeek thinking, OpenAI-style reasoning_effort, Anthropic extended thinking, etc.), expose a working selector in the Desktop composer (and CLI /effort) that actually translates into the provider request. Show it for every model that advertises the capability.
2. Context/token budget control: Add a per-session (or per-turn) setting for max context window (token cap), with sensible presets, so users can tune memory usage for large repositories.

Describe alternatives you've considered

• Exposing raw "advanced model parameters" in Settings (JSON or form fields) as a fallback for power users.
• Documenting the exact CLI flags that map to these controls so at least CLI users can set them.

Additional context

• Provider: DeepSeek (openai-compatible endpoint), via deepcode provider set deepseek ....
• Model example: deepseek-chat / deepseek-reasoner.
• Version: DeepCode Desktop 0.1.0 (v2.0.0 runtime). README advertises "Thinking 档位" (/effort) support, but it is not functional/exposed for DeepSeek today.

[Zongwei9888]

Thanks for the detailed report — this one turned out to be more than a missing option, and it led us to several fixes. Addressed in #161 and #162.

What was happening. The code that decides which thinking levels a model offers had drifted out of sync with our model catalog. It still recognised older Claude names while the catalog had moved on, so a number of models were advertising no reasoning controls at all — DeepSeek among them, but also Claude's entire current line-up, Qwen and Grok. The picker builds its options from that value, which is exactly why you saw an empty control. Your description was spot on.

What you'll see now:

model	before	after
deepseek-chat	Auto	Auto / Off
claude-sonnet-5	Auto	Auto / Off / Low / Medium / High / Max
glm-4.6, MiniMax-M2	Auto	Auto / Off

On the low/medium/high you asked for specifically: DeepSeek's API offers thinking as an on/off switch rather than graded levels, so unfortunately there aren't three levels for us to expose — we'd only be adding options that don't do anything. deepseek-reasoner always reasons and can't be turned off, so it has no choice to offer either. If graded effort is important for your workflow, Claude and the GPT-5 family do publish real levels and both work properly now.

Your report also surfaced two other issues, both fixed: GLM and MiniMax were falling back to a 128K context (MiniMax M3 is actually 1M), and GLM requests were dropping the thinking setting entirely because the provider config was missing a field.

On context length: the automatic per-model limits now work correctly, but the manual control you asked for isn't there yet. There's some plumbing to sort out underneath first. We'll get to it — keeping this issue open to track it.