Workspace Section

The «Workspace» section — agent behavior and token saving.

Workspace mode (general settings)

Agent style:
⚖️ Hybrid — the agent applies edits itself, runs the build and fixes errors until success.
🤖 Autonomous agent (Strict) — reads files, applies edits and compiles on its own.
💬 Chat-oriented — changes are discussed; code on request.
Reasoning & reply language — follows the UI language: Russian UI — the model thinks in Russian, English — in English.
Auto-apply edits — the agent writes files to disk immediately; otherwise it asks for confirmation per change.
Full working mode — the agent writes/edits C++ of any depth, creates folders, compiles the plugin and fixes build errors until success. When off — the agent reports after edits, and you run the build.
Limit agent iterations — when off, the agent works without an iteration limit (no loop warnings).

Plugin memory — the agent memorizes facts about the project (modules, engine version, preferences) and reuses them in later tasks. The memory budget is hard-capped (~500 tokens). The «Forget this folder's memory» button.
Solved-errors base — successful compile fixes are memorized (up to 40 records); when a build fails the agent gets a proven solution. The «Clear base» button.

Auto context compression — when a dialog takes more than 60% of the limit, the history is summarized automatically.
Model context limit (tokens) — enter the real limit of your model (32k / 128k / 200k…); used by the token indicator and auto-compression.

Token counter — total token usage, task count and cache share in the status bar.