You are designing "filecode" — a from-scratch recreation and...

Prompt

You are designing "filecode" — a from-scratch recreation and improvement of Claude Code (Anthropic's terminal coding agent), for a solo builder (indie dev / early-startup founder) building this as both his own daily driver and a future product for indie devs, early-stage, and mid-market startups. CORE ARCHITECTURE: DUAL (FILE, CODE) SYSTEM Design an agent with two first-class operating modes that share the same session/context/tool infrastructure but diverge in behavior: - FILE mode: large-file and document handling — reading/editing/organizing huge files, PDFs, multi-file refactors, repo-wide search/organize, non-code artifact generation. - CODE mode: writing, editing, debugging, and shipping code — project creation, multi-file coding tasks, test running, git operations. Specify how mode-switching works (automatic inference vs explicit /mode command), and what state is shared vs isolated between the two. HARD REQUIREMENTS TO DESIGN FOR: 1. Token efficiency — this is the #1 differentiator vs Claude Code. Specify: context compaction strategy, prompt caching usage, a layered context hierarchy (e.g. Global → Workspace → Project → Task → Agent → Tool context) with rules for what gets loaded/evicted at each layer, diff-based edits instead of full-file rewrites, and an AST-aware repo map so the agent doesn't re-read whole files it already understands. 2. Multi-provider support — cloud AND local. Must support Anthropic, OpenAI-compatible endpoints, and local inference (Ollama/llama.cpp-style), with per-task or per-mode model routing (e.g. cheap/local model for trivial edits, frontier model for hard reasoning) and a pluggable provider abstraction so new providers can be added without core changes. 3. Full agent architecture — specify: the core agent loop (ReAct-style or otherwise), sub-agent/orchestrator pattern for parallelizable work, a permission engine (tiered: auto-approve / ask / deny by risk), sandboxing model for shell/file execution, git snapshot + undo, streaming output, and rate-limit/retry handling. 4. Sessions — persistent sessions with resumability, and MULTIPLE CONCURRENT SESSIONS running in parallel (e.g. separate sessions per project or per task) without state collisions. Specify how session state is stored (SQLite or similar) and isolated. 5. UI + UX parity/superiority — terminal UI (specify TUI framework choice and rationale), slash commands (list the essential set: /mode, /model, /session, /undo, /cost, /plan, etc.), plan-mode vs execution-mode separation, and streaming/interactive feedback that feels faster than Claude Code. 6. Privacy — local-first option: ability to run entirely against local models with zero cloud calls, clear boundary on what telemetry (if any) exists, and secrets/config handling that never leaves the machine unless explicitly cloud-routed. 7. Speed — cold-start time, incremental context loading, and caching strategy to beat Claude Code's perceived latency. OUTPUT FORMAT: Produce a complete technical specification covering: - Tech stack recommendation with justification (language, runtime, TUI lib, DB, schema validation, AI SDK) - Full system architecture diagram (described in text/ASCII) showing agent loop, sub-agents, provider layer, context layers, permission engine, and session store - The core system prompt for the agent itself (the prompt that would drive its behavior at runtime) - Data models (session, task, context layers) as concrete schemas - A phased build roadmap (MVP → v1 → product-ready) with the smallest possible MVP that is still genuinely useful daily - A list of concrete ways this beats Claude Code (name the specific mechanism, not just "it's better") Constraints: prefer open-source, self-hostable components over paid SaaS dependencies. Optimize for a single developer building and maintaining this. Be concrete and buildable, not aspirational — every recommendation should be something that could be scaffolded this week.