Chetna Editor

An AI coding assistant for VS Code and Cursor — local-first agent workflows inside your editor.

Chetna indexes your workspace, retrieves only what matters for each turn, and executes changes through a governed tool loop with diff review, permission prompts, rule constraints, and full auditability. It runs as a native extension: no fork, no separate app.

Current release: 0.5.1 · VS Code ^1.85 · Publisher DistractionsForge

Features

Agent chat & editing

Multi-tab sessions (general, refactor, debug, docs, Unity, DevOps)
Native tool protocol (OpenAI / Claude / Gemini / LM Studio) plus XML fallback
Collapsible Thought / Plan / tool panels and streaming markdown replies
Clickable file paths in chat — open files at line/column
Applied edits with collapsible Diff review panel (Review / Keep / Revert per file)
Coding pipeline: verify → gap analysis → todos → generate (approval gate + Done state when finished)
Compile-fix loop with circuit breaker, diagnostic arbitration, and post-generation lint
In-chat permission elevation — Allow once / session / Deny when a tool needs a higher mode

Context & memory

4-layer symbol index — files → classes → functions → details (.chetna/layers/)
Codebase chunk RAG — semantic retrieval over function/class slices
Merkle incremental sync — re-index only what changed
Context window presets — 16k / 32k (default) / 48k / 64k / 128k / 256k from the composer toolbar
Episodic memory — prior turns, grep hits, and session ledger summaries
db_context_search — unified search across index, chat history, memory, and process maps
Process maps — auto-captured workflows for faster repeat tasks

Tooling

Shell, grep, glob, git, web fetch/search, LSP, Python, Unity (when detected), MCP passthrough, and a permission model (read-only → workspace-write → danger-full-access).

Composer toolbar: Active file, Git diff, LLM provider, model, and Context pickers — settings persist per workspace.

Install

Marketplace

Search Chetna Editor by DistractionsForge, or install the extension ID DistractionsForge.chetna-editor.

From VSIX

VS Code

code --install-extension chetna-editor-0.5.1.vsix

Cursor

Cmd/Ctrl+Shift+P → Extensions: Install from VSIX…, or:

cursor --install-extension chetna-editor-0.5.1.vsix

Reload the window after install.

Run

Open your project folder in VS Code or Cursor.
Run Chetna: Index Workspace (Cmd/Ctrl+Shift+P).
Open Chetna: Open Chat (Cmd/Ctrl+Shift+C).
Pick a provider/model (and Context size if needed), then run Chetna: Test LLM Connection once.

Use Active file / Git diff toggles as needed. When tools need more permission than your ceiling, approve or deny from the in-chat card.

LLM providers

Provider	Configuration
Auto (default)	LM Studio → OpenAI → Claude → Gemini → local context-only
LM Studio	Any OpenAI-compatible endpoint via `chetna.lmStudio.baseUrl` (local, LAN, tunnel, or remote)
OpenAI	`chetna.openaiApiKey`, `chetna.openaiModel`
Gemini	`chetna.gemini.apiKey`, `chetna.gemini.model` (default `gemini-2.5-flash`)
Claude	`chetna.claude.apiKey`, `chetna.claude.model`

{
  "chetna.llmProvider": "auto",
  "chetna.lmStudio.baseUrl": "http://192.168.1.50:1234/v1",
  "chetna.openaiModel": "gpt-4o-mini",
  "chetna.gemini.model": "gemini-2.5-flash",
  "chetna.context.preset": "32k"
}

chetna.lmStudio.baseUrl accepts any OpenAI-compatible endpoint. You can omit /v1 when the host has no path (e.g. http://192.168.1.50:1234).

Configuration

{
  "chetna.llmProvider": "auto",
  "chetna.tools.permissionMode": "workspace-write",
  "chetna.tools.promptOnElevation": true,
  "chetna.context.preset": "32k",
  "chetna.context.codebaseRag": true,
  "chetna.context.codebaseTopK": 8,
  "chetna.editing.requireDiffReview": false,
  "chetna.composer.autoStage": true,
  "chetna.agent.exploreBeforeEdit": true,
  "chetna.webSearch.provider": "duckduckgo"
}

Setting	Purpose
`chetna.context.preset`	Context window preset (default 32k) — sets model window + injection ceilings
`chetna.context.maxTokens`	Ceiling for injected context (RAG/rules/attachments); updated by preset
`chetna.context.historyMaxTokens`	Ceiling for chat history packing; updated by preset
`chetna.tools.promptOnElevation`	In-chat Allow once / session / Deny when a tool needs a higher mode
`chetna.composer.autoStage`	Apply edits immediately and show Diff review for Keep/Revert
`chetna.editing.requireDiffReview`	Force diff review even when auto-stage is off
`chetna.agent.orchestrator`	Planner → executor → summarizer for complex tasks
`chetna.compile.maxFixCycles`	Compile-fix loop safety limit

Cursor MCP integration

Run Chetna: Setup Cursor Integration (MCP) to register a stdio server in .cursor/mcp.json. Cursor Agent can then call the same toolchain as the sidebar:

chetna_search_codebase · chetna_grep · chetna_glob_search · chetna_git_* · chetna_rag_query · chetna_execute_tool · Unity/Python tools · chetna_capabilities

Commands

Command	Description
Chetna: Open Chat	Sidebar agent panel
Chetna: Index Workspace	Full index + merkle sync + codebase chunks
Chetna: Validate Index	Health check for stale or missing paths
Chetna: Test LLM Connection	Verify active provider
Chetna: Grep Workspace	Ripgrep-style search
Chetna: Show Layer Stats	Index, RAG, memory, and merkle stats
Chetna: Review Pending Edits	Diff review for applied changes (Keep/Revert)
Chetna: Setup Cursor Integration	Write MCP config

Data & privacy

Index, chat history, memory, and process maps live under .chetna/ in your workspace (local JSON).
API keys are stored in VS Code settings / Secret Storage — not sent anywhere except your chosen LLM provider.
No telemetry is bundled in the extension core; network use is limited to LLM APIs and tools you invoke (web search, fetch, etc.).