B4.run: Build TypeScript Agents

B4.run is aimed at TypeScript developers and AI agent builders for building TypeScript agents. Its core idea is that files are features: tools, planning, memory, skills, subagents, and evals are all organized as files. Key capabilities include: - Tools: each tool is one file with a default-exported function that receives parsed args and a read-only ctx (toolName, routeId, threadId, signal, params), and can be async. - Routes and agents: define agents with the agent() descriptor, supporting four route kinds, and generate a type module .b4/b4.generated.d.ts that maps each route pathname to that route's tool signatures. - Workspaces and sandbox execution: supports workspaces and sandbox execution. - Approval gates: can ask a person before each call to certain tools. - Subagents: the parent gets the child's final text back as the tool result. - Retry and reasoning: supports retry configuration (e.g. maxAttempts, baseDelay) and reasoning effort (e.g. effort: high). - Recording and replay: supports recording and replaying model responses. - Offline agent test: provides an offline agent test. Usage: scaffold an app with npm create b4-app@latest my-agent. Running b4 run looks up the route's <routeId>#<kind> assistant id, creates a thread for the call, sends the input through Agent Protocol, and prints the final state. --dry-run creates no model and performs no write.