Flue vs Sandcastle vs Building It Yourself

하이라이트

• Flu uses Pi and Hono under the hood, featuring Vercel's Just Bash in-memory filesystem interpreter to run commands like cat package.json without a full terminal.

• Flu provides built-in Cloudflare R2 bucket mounting, enabling agent sessions to persist message history and state across days, weeks, or years.

• Flu allows sandboxing CLI tools like the GitHub CLI (gh) with authentication tokens so agents execute commands without exposing raw secrets.

• Sandcastle uses Docker for isolation and implements Ralph loops to ensure task completion, executing iterations through Claude Code until all steps finish.

• Neither Flu nor Sandcastle natively support personal Claude Max or Codex Max subscriptions, driving the choice to build custom TypeScript automation.

• Building a custom TypeScript orchestrator for Inbox Zero allows precise multi-step execution including Linear issue fetching, TDD implementation, /simplify cleanup, PR creation, bot comment resolution, and Slack notifications.

타임라인

1. 00:00 - 01:03

   Overview of Flu and Sandcastle Frameworks

   Flu and Sandcastle represent two recently released TypeScript libraries designed to structure AI agent workflows. Flu focuses on agent harness structures, while Sandcastle specializes in sandbox execution for coding agents. Neither framework directly supports user subscriptions like Claude Max or Codex Max, leading creators to build custom in-house automation pipelines instead.

   • Flu is an agent harness framework developed by the Astro team.
   • Sandcastle is a TypeScript library created by Matt Pocock for orchestrating AI coding agents inside isolated sandboxes.
   • Custom agent setups remain necessary when subscription compatibility with tools like Claude Max or Codex Max is required.
2. 01:03 - 04:10

   Architecture and Features of the Flu Agent Harness

   Flu routes agent requests via HTTP webhooks using Hono and utilizes Pi internally. It features Just Bash, an in-memory execution environment from Vercel that gives agents filesystem capabilities without requiring full container setups. File storage and message histories mount to Cloudflare R2 buckets for long-term persistence, while scoped CLI command wrappers protect sensitive environment variables.

   • Flu integrates Pi and Hono to expose webhook endpoints for model interactions.
   • Just Bash provides a pure TypeScript in-memory bash interpreter to run safe terminal commands.
   • State and history persist through Cloudflare R2 storage integration.
   • Restricted CLI execution allows tools like gh to run safely without leaking tokens to the LLM.
3. 04:10 - 08:02

   Evaluating Agent Workflows and Sandcastle Features

   An end-to-end coding workflow requires structured execution across distinct phases: issue ingestion, TDD implementation, code simplification, pull request creation, automated code review handling, and team notifications. Sandcastle addresses execution reliability by running Claude Code inside Docker containers and applying Ralph loops, which re-prompt the model repeatedly until all steps in a plan are completed. It also automates repo operations like Git worktree initialization and environment variable symlinking.

   • Ideal software engineering workflows require multi-stage control across issue retrieval, TDD, code review, and notification.
   • Sandcastle relies on Docker containers to isolate Claude Code execution.
   • Ralph loops enable repetitive execution until complex multi-step tasks reach completion.
   • Built-in templates support sequential reviewers, parallel planners, and automated Git worktree management.
4. 08:02 - 11:05

   Building a Custom TypeScript Orchestrator for Inbox Zero

   To maintain compatibility with Claude Max subscriptions and retain complete control over task execution, a custom 7-step TypeScript script manages the workflow for the open-source Inbox Zero repository. The automation creates isolated Git worktrees, syncs Linear tickets, runs classification prompts, applies TDD implementation, executes cleanup via /simplify, pushes PRs, processes secondary bot comments, and broadcasts status messages to Slack upon completion.

   • Custom orchestration scripts deliver exact step-by-step control without framework abstraction overhead.
   • The automated Inbox Zero pipeline manages Linear tasks, Git worktrees, pnpm dependencies, TDD cycles, PR creation, and Slack alerts.
   • Writing orchestration code manually ensures flexibility to adjust prompts and stages without framework lock-in.