scriptschnell

scriptschnell

This project is actively in initial development. Use at your own risk!

scriptschnell is a TUI that generates code from prompts using LLMs.

On first run, it will download tinygo (if not in PATH).

Frontends:

• TUI
• CLI
• Web (local)
• ACP

Features

• Multiple providers supported
• Native search engine support
• Golang WASM sandbox (compiled with TinyGo, which is downloaded on startup)
• Auto-continue for long-running sessions
• Auto-compaction during generation for longer sessions
• Separate LLM models (all must support tool calls via API):
  • Orchestrator: writes the actual code
  • Summarization: summarizes and makes small choices
  • Planning: Plans implementation (single shot or planning steps with one-by-one loops)
  • Safety: Used for safety-critical tasks, falls back to summarize model if not configured
• Read only contexts like external documentation, library sources, etc.
• Experimental agent client protocol support

TODO

• Infinite text loop detection (Gemini loves to repeat itself)
• Better authorization UI
• More e2e tests using other programming languages than Go
• Custom system prompt
• System prompt optimized for current codebase
• Better /init command for generating AGENTS.md
• Fix write diff
• agent-client-protocol for better editor integration
• Better UI experience for parallel_tools call
• Encourage LLMs to use the parallel_tools call (only gemini seems to do this)
• Limit the auto-continue judge so it doesn't get stuck in a loop
• Unify the model/provider detection (context window size/model specific workarounds)
• Fix auto compaction
• Get rid of magic literals in the codebase
• General UI improvements around tool calls
• Todo task are not showing up in the UI
• On configured web search, add text to the system prompt for more up-to-date results
• Speed up startup time
• Support user configured mcp's with auto-selecting only relevant ones
• Only cache directory listings in working directory
• Implement persistent authorization on a per-project basis
• Unify tui command programming
• Filter models by provider in models menu
• Fix models menu (weird scrolling behavior, layout issues)
• On configuring the first provider, choose a default model
• Fix mistral model performance
• Use token usage response from provider if available
• Output credit usage per session (currently just openrouter, implement cost estimation)
• Weird kimi k2 context window problems
• Deadlock problem when model used invalid tool call (the session cannot continue)
• Multi session support in the tui
• Anthropic Prompt Caching
• Auto session saving
• Server mode
• Split config in runtime config and user config
• Fix agent client protocol output (zed editor)
• Implement proper shell sandbox
• Continue implementing the eval tooling
• After auto compaction tell the model that this happened and it should verify that requested changes were fully implemented, if thinks that these were fully implemented

Goals

• Universal codebase LLM editing experience (cli, tui, code editor, web browser)
• Compatible with many LLM providers
• Promoting usage of open weight models
• Improved reliability of results

Maybes

• Implement a small bash parser so bash on Windows works?
• Support for easy containerization, e.g. devcontainer?
• Expose external MCP servers as OpenAPI endpoints in golang sandbox?

Agentic Workflow

The orchestrator coordinates LLM calls, tool execution, and adaptive control loops. The diagram below shows how the orchestration model, safety evaluator and model, tool registry, summarization model, and auto-continue judge work together.

flowchart TD
    user[User Prompt] --> session[Session State]
    session --> builder[Prompt Builder]
    builder --> orchestrator[Orchestrator Loop]
    orchestrator -->|Completion requests| orchLLM[Orchestration Model Client]
    orchLLM -->|Assistant text + tool calls| orchestrator

    orchestrator -->|Planning decision| summarizer
    summarizer -->|"Planning config (Run? MCPs?)"| orchestrator

    orchestrator -->|Pre-loop Planning| planning[Planning Agent]
    planning -->|Plan generation| planLLM[Planning Model]
    planning -->|Investigation| investigator
    planning -->|Record plan| session

    orchestrator -->|Dispatch tool call| registry[Tool Registry]
    registry -->|Check for secrets| secretDetector[Secret Detector]
    secretDetector -->|Secrets found?| secretAuth{Has secrets?}
    secretAuth -->|Yes| secretAwareAuth[Secret Aware Authorizer]
    secretAwareAuth -->|Judge with secrets| authActor[Authorization Actor]
    authActor -->|Authorization decision| summarizer
    summarizer -->|Allow / Deny| secretAwareAuth
    secretAwareAuth -->|Authorization decision| registry
    secretAuth -->|No| registry
    registry --> tools["Tool Actors<br/>(read_file, edit_file, shell, sandbox, MCP, ...)"]
    tools -->|Tool outputs| contentGate{External content?}
    contentGate -->|No| orchestrator
    contentGate -->|Yes| safetyEval[Safety Evaluator]
    safetyEval -->|Safety prompt| safetyLLM[(Safety Model)]
    safetyLLM -->|Safe / Unsafe| safetyEval
    safetyEval -.->|Fallback if unset| summarizer
    safetyEval -->|Allow / Block| orchestrator
    orchestrator -->|Record message| session
    orchestrator -->|Stream status| ui[TUI / CLI Output]

    orchestrator -->|Configured servers| mcpMgr[MCP Manager]
    mcpMgr --> mcpSources["External MCP Servers<br/>(command, openapi, openai)"]
    mcpMgr -->|Register tools| registry

    orchestrator -->|Optional tool selection| toolFilter[Optional Tool Filter]
    toolFilter -->|Summarization request| summarizer[(Summarization Model)]
    summarizer -->|Relevant tools| toolFilter
    toolFilter -->|Enable subset| registry

    orchestrator -->|Token usage check| autoCompact{Over context limit?}
    autoCompact -->|Summarize earlier messages| summarizer
    summarizer -->|Summary message| session

    orchestrator -->|Truncated or empty reply| autoContinue{Auto-continue Judge}
    autoContinue -->|Digest recent conversation| summarizer
    autoContinue -->|Continue / stop| orchestrator

    orchestrator -->|LLM error| errorJudge[Error Judge Actor]
    errorJudge -->|Retry analysis| summarizer
    errorJudge -->|Retry / abort| orchestrator

    orchestrator -->|Investigate codebase request| investigator[Codebase Investigator Agent]
    investigator -->|Limited tool registry| invRegistry[Investigation Tool Registry]
    invRegistry --> invTools["Read-only Tools<br/>(read_file, search_files, search_file_content, parallel_tools)"]
    invTools -->|Tool outputs| investigator
    investigator -->|Investigation queries| summarizer
    summarizer -->|Tool calls + reasoning| investigator
    investigator -->|Investigation results| orchestrator

Providers

Supported providers:

• OpenAI
• Anthropic
• Google Gemini
• Mistral
• OpenRouter
• Cerebras
• Groq
• Kimi (Moonshot AI)
• Z.AI (GLM)
• Ollama
• OpenAI-compatible

Recommended since 2025-12-10:

• Orchestrator model: Kimi K2.5 or ZAI GLM 4.7
• Summarization model: Qwen3 Coder Flash
• Planning model: Kimi K2.5
• Safety model: gpt-oss-safeguard-20b