Thinking Mode

Three thinking tiers (Off / Standard / Deep) decide how much compute the model spends reasoning before it answers.

What it is

CrabCode collapses model "thinking intensity" into three user-facing tiers you switch by task difficulty:

Tier	Behavior
Off	The model answers without an explicit reasoning step; cheapest and fastest
Standard	Recommended default; covers most daily work
Deep	Gives the model more budget for long chain-of-thought; use for hard planning, cross-file refactors, root-cause work

Under the hood, when CrabCode routes through the acosmi.com / acosmi.ai gateway, it emits the right parameter for the tier you picked. Per-provider field differences are absorbed by the gateway — one tier on your side, every supported model lines up.

When you see this doc

The thinking-tier indicator at the top of the TUI
The thinking section of the /model menu
/help model notes

How to switch

In the TUI:

shell

/effort               # no args → show the current tier
/effort off           # English alias
/effort standard
/effort deep
/effort 关闭思考      # Chinese explicit
/effort 标准思考
/effort 深度思考
/effort auto          # clear the explicit setting; fall back to default
/effort               # no args → show the current tier
/effort off           # English alias
/effort standard
/effort deep
/effort 关闭思考      # Chinese explicit
/effort 标准思考
/effort 深度思考
/effort auto          # clear the explicit setting; fall back to default

You can also pin a default in settings.json (see settings) or bind keys via keybindings.

Which tier to pick

Situation	Suggested tier
Touch up a small function, complete a snippet, write a comment	Off / Standard
Interactive review, Q&A	Standard
Multi-file code generation, refactors	Deep
Security- or architecture-critical decisions, root-cause hunts	Deep
One-shot shell scripts, bulk lint fixes	Off

Default behavior

On start, CrabCode follows your settings.json plus your account default (typically Standard). /effort auto removes the local explicit override and restores that default.

Limits and caveats

Not every model accepts Deep — there's a capability gate; unsupported models silently degrade to Standard, no error
Thinking still costs tokens — Deep's reasoning steps count toward output tokens; /cost reflects this
Off is cheapest and weakest — fine for quick lookups but don't let it plan
Bypassing the acosmi gateway can break the mapping — CrabCode defaults to acosmi.com / acosmi.ai with sanitizer translation. Custom direct-provider setups rely on that provider doing the right thing