Whittl Glossary¶
Short definitions for the terms used across these docs. If something is unclear in a longer page, it probably has a one-liner here.
Agent Mode¶
Opt-in unbounded tool-use loop for capable models. The AI gets 20-50 rounds to iterate on a problem autonomously: read code, edit, test, fix, test again. Distinct from standard generation which is capped at a handful of auto-fix rounds. See Agent Mode.
Archive¶
A frozen snapshot of your project at a specific point in time. Whittl creates one automatically before each generation. Stored under archives/vN_<timestamp>/ inside your project folder. Restored from the History panel. See Project History.
Auto-fix¶
The iterative error-correction system. When a generation crashes or auto-detected errors fire, the AI gets another turn (or several) to fix the problem. Capped at 5 rounds by default, with an oscillation guard that bails if the same error signature recurs. See Auto-fix Rules.
Autofix rule¶
A static pattern in Whittl's Layer that catches common code-generation mistakes before you ever see them — hallucinated APIs, missing imports, Qt enum typos. 75+ rules ship in v2.3. Part of the Whittl Layer.
Backend¶
An AI provider Whittl can talk to. Five supported: Claude, Gemini, DeepSeek, OpenRouter, Ollama. See Choosing a Backend.
Capability chips¶
Badges on each model in the Models dialog showing what it supports: [Tools] [Thinks] [Long] [Vision]. Helps pick a model without looking up specs.
Commons¶
Whittl Commons — planned community-sourced bundle of autofix rules Whittl downloads on startup, like virus definitions for code generation. Ships in v2.5.
Explain Mode¶
The Explain toggle in the editor toolbar. When on, the AI adds inline #: comment annotations throughout the current file explaining what each non-trivial block does. Toggle off strips the comments. Cached per-file, so re-toggling is instant if the code hasn't changed. See Explain Mode & Inline AI Help.
Explain Selection¶
Right-click context menu action on highlighted code. AI explains the selection in chat without modifying the file. See Explain Mode & Inline AI Help.
Expand (prompt expansion)¶
The Expand checkbox in the chat panel. When checked, the AI first rewrites your short prompt into a detailed spec, then generates from the detailed version. Turns "calculator with history" into a 10-line spec before writing code. See Prompt Expansion.
Fix Selection with AI¶
Right-click context menu action on highlighted code. AI rewrites only the selected range — no full-file regeneration. Proposed replacement shown as a diff before apply. See Explain Mode & Inline AI Help.
General Python¶
One of Whittl's four targets. For non-GUI code: CLI tools, scripts, plugins, automation. No framework imposed.
Layer (Whittl Layer)¶
The knowledge layer between you and the model — autofix rules + skills + validators + anti-pattern catchers. What makes cheap models produce better code through Whittl than through a bare API call. See the dev log for positioning.
Planner¶
The component that decides what actions to take before a generation: which files to send, which skills to include, which autofix rules are relevant. Can hit rule-based shortcuts or invoke an AI classifier. Part of Smart Routing.
Project target¶
One of four framework choices per project: Desktop (PySide6), Mobile (Flet), Basic (CustomTkinter), General Python. Set when creating a project; changing it after generally means regenerating. See Choosing a Target.
Prompt caching¶
Anthropic's and Google's feature that discounts repeated context in multi-turn conversations. Whittl opts in automatically. Roughly 87% cost reduction on multi-round Claude sessions. See Claude.
Session memory¶
Across tool-use rounds, the AI's scratch state — what it edited, what errors came back, which files it looked at. Whittl retains this across a generation cycle so Agent Mode iterations don't restart cold each round.
Skills¶
Markdown files at ~/.whittl/skills/*.md that teach the AI reusable patterns — "always call setSizePolicy after creating a QWidget in a layout," "Flet uses ft.Page not ft.App," etc. Injected into the system prompt when relevant. GUI editor at Preferences → Skills. See Skills System.
SKILL.md¶
The cross-tool skill file format. Whittl, Claude Code, and OpenCode all read SKILL.md files. Making a skill for one makes it work in the others. See Skills System.
Smart Routing¶
AST-based file-relevance scanning. When your project has many files, Whittl only sends files the AI actually needs for the current request. The status bar shows "3 of 27 files sent" when routing activates. See Smart File Routing.
Stop button¶
The red button in the chat panel that cancels the active generation or auto-fix cycle. Persists across rounds — hitting Stop during round 3 of 5 cancels queued rounds 4 and 5 too.
Target¶
Shorthand for Project Target. See above.
Test Run¶
F5 in the editor. Launches your project in a subprocess, captures stdout/stderr, feeds crashes back to the auto-fix cycle. Non-blocking — you can keep editing while it runs. See Debugging a Crash.
Think mode¶
The Think checkbox in the chat panel. Activates extended reasoning on supporting models (DeepSeek R1, Claude extended-thinking, OpenAI o1-style). Costs more tokens; higher quality on hard problems. See Think Mode.
Tool executor¶
Whittl's mid-generation tool-use dispatcher. Handles AI requests for edit_code, read_file, run_bash, question, etc. Different loops for Claude (multi-turn, agentic) vs other backends (single-turn, action-capped).
Whittl Layer¶
See Layer.