Skip to content

ROADMAP.md

Latest commit

 

History

History
226 lines (186 loc) · 10.3 KB

ROADMAP.md

File metadata and controls

226 lines (186 loc) · 10.3 KB
Raw

vibe-local Improvement Roadmap

Last updated: 2026-02-23

vibe-coder: Claude Code Replacement (v0.3)

Architecture change: Claude Code CLI + proxy replaced with vibe-coder.py

OLD: vibe-local.sh → claude CLI (proprietary) → proxy.py → Ollama
NEW: vibe-local.sh → vibe-coder.py (OSS) → Ollama direct

Benefits: No login, no Node.js, no proxy process, fully OSS, Python+Ollama only

vibe-coder.py Implementation Status

Phase Feature Status
1 Config (CLI args, config file, env vars) DONE
1 OllamaClient (direct /v1/chat/completions) DONE
1 BashTool (subprocess, timeout) DONE
1 ReadTool (cat -n format, offset/limit) DONE
1 WriteTool (absolute paths, mkdir -p) DONE
1 EditTool (unique check, replace_all) DONE
1 ToolRegistry + OpenAI function calling schemas DONE
1 PermissionMgr (safe/ask/deny) DONE
1 Session (in-memory, JSONL persistence) DONE
1 Agent loop (LLM → tool → result → loop) DONE
1 TUI (readline, ANSI colors, streaming) DONE
1 XML tool call fallback (Qwen models) DONE
2 GlobTool (os.walk + fnmatch) DONE
2 GrepTool (os.walk + re, context lines) DONE
2 WebFetchTool (urllib, HTML→text) DONE
2 WebSearchTool (DuckDuckGo HTML) DONE
2 NotebookEditTool (JSON .ipynb) DONE
2 System prompt with OS-specific hints DONE
3 Markdown rendering in TUI DONE
3 Multi-line input (""") DONE
3 Session persistence (JSONL) DONE
3 Session resume (--resume, --session-id) DONE
3 Context compaction DONE
4 Permission system (safe/ask/network tiers) DONE
4 -y flag (auto-approve) DONE
5 Subagent/Task tools TODO
5 Plan mode TODO

Launcher Updates

  • vibe-local.sh: Updated to use vibe-coder.py directly (no proxy)
  • vibe-local.ps1: Updated to use vibe-coder.py directly (no proxy)
  • install.sh: Node.js/Claude Code now optional
  • install.ps1: Node.js/Claude Code now optional

Current State (v0.2)

Working:

  • Dual-model routing (qwen3-coder:30b + qwen3:8b sidecar)
  • macOS environment injection (brew, /Users/, system_profiler)
  • Windows environment injection (winget, %USERPROFILE%, PowerShell)
  • Proxy auto-restart on code update (mtime detection)
  • Tool filtering (20 → 9 essential tools)
  • XML tool call fallback
  • Zero crashes, zero errors in proxy
  • Speed test via curl example works
  • HTML preferred over pygame for GUI
  • Windows support (install.ps1, vibe-local.ps1, vibe-local.cmd)
  • Init probe fast-path (instant response for connectivity checks)
  • tool_choice forwarding (Anthropic→OpenAI format conversion)
  • stop_sequences / top_p / top_k forwarding
  • Image/Vision support (Anthropic base64→OpenAI data URI)
  • Ollama tokenize for accurate token counting
  • Security: replay scripts 0o700, 50MB request size limit

P0 — Critical (Breaks User Experience)

P0-1: Response verbosity / TOOL FIRST not effective enough

Status: IN PROGRESS Impact: Every response has unnecessary explanation Evidence:

  • "回線速度を測定するには、専用のツールが必要です。まずは、そのツールをインストールしてみますか？" (should just brew install)
  • "はい、Gitにアクセスできます。現在の状態では..." (4 bullet points nobody asked for)
  • "Python3がインストールされており..." (3 lines of explanation for python3 --version)

Fix: Strengthen system prompt: "ZERO preamble text before tool calls. After tool result, reply in 1 sentence MAX. NEVER ask follow-up questions."

P0-2: Follow-up questions after every action

Status: IN PROGRESS Impact: Model asks "何か特定の操作が必要ですか？" after EVERY response Evidence: Appears 4 times in one session Fix: Add rule "NEVER end with a question. Wait for user's next instruction."

P0-3: Won't run commands for user

Status: IN PROGRESS Impact: Tells user to run commands instead of using Bash tool Evidence: "ゲームを実行するには、以下のコマンドをターミナルで実行してください" instead of Bash(python3 minesweeper.py) Fix: Add rule "ALWAYS execute commands yourself using Bash. NEVER tell user to run commands."

P0-4: curl URL quoting (zsh glob)

Status: IN PROGRESS Impact: URLs with ? fail on first try in zsh Evidence: no matches found: https://speed.cloudflare.com/__down?bytes=10000000 Fix: Update speed test example with quoted URL. Add rule "ALWAYS quote URLs containing ? or & in Bash commands."

P0-5: \\n literal in curl output

Status: IN PROGRESS Impact: curl -w format shows \\n as text instead of newline Evidence: 速度: 62497265 bytes/sec\\nダウンロード時間: 0.160007s\\n Fix: Fix speed test example to use $'\n' or avoid newlines entirely.

P0-6: WebSearch returns fabricated URLs

Status: DONE ✅ (verified 2026-02-22) Impact: WebSearch sidecar calls go to Ollama which can't search → model fabricates URLs Evidence: Sidecar request with tool_choice: {name: "web_search"} gets proxied to Ollama, returns empty, model invents sources Fix: Proxy intercepts WebSearch sidecar calls (tool_choice.name == "web_search"), performs real DuckDuckGo HTML search (html.duckduckgo.com), returns actual results as text. 8 results in ~1.3s. Offline fallback returns honest "not available" message. Key learning: Detection must check tool_choice.name only (not tools array), because Claude Code sends tools: [web_search_tool] (1 element) not tools: []. Response uses simple text format (not server_tool_use + web_search_tool_result which requires opaque encrypted_content). "Did 0 searches" counter is cosmetic only — real results are used.

P1 — High (Functional Issues)

P1-1: Auto-install dependencies

Status: TODO Impact: Won't install pygame before running, tells user to do it Evidence: "Pygameライブラリがインストールされている必要があります" → should pip3 install pygame first Fix: Add rule "Install missing dependencies with pip3/brew BEFORE running scripts."

P1-2: stdin limitation awareness

Status: TODO Impact: Runs scripts with input() via Bash → EOFError Evidence: minesweeper.py with input() → EOFError: EOF when reading a line Fix: Add rule "Scripts with input()/stdin CANNOT run in Bash. Use GUI (pygame/HTML) or rewrite to accept args."

P1-3: Don't give up on errors

Status: TODO Impact: Model suggests going back to browser version instead of fixing Python version Evidence: "代わりに、ブラウザ版のマインスイーパを使用してください" after one error Fix: Already partially in prompt ("If a tool fails, try alternative"), strengthen with "NEVER suggest user do something else. Fix the problem."

P1-4: Mixed language output

Status: TODO Impact: Chinese characters randomly appear in Japanese responses Evidence: "是非" (Chinese context), "或者" (Chinese conjunction) Fix: Add "Reply in the SAME language as the user. If user writes Japanese, reply in Japanese only."

P2 — Medium (System/Proxy Improvements)

P2-1: Sidecar cold start (18.5s)

Status: MITIGATED (init probe fast-path bypasses Ollama entirely for probe requests) Impact: First sidecar request takes 18s while model loads into VRAM Fix: Init probe fast-path added (A1). Warmup thread already exists in proxy startup.

P2-2: Proxy log overwrite

Status: TODO Impact: proxy.log loses previous session data on restart Fix: Use append mode (>>) in vibe-local.sh instead of overwrite (>)

P2-3: env_injected tracking in debug

Status: TODO Impact: Cannot audit environment injection from debug files Fix: Add env_injected field to req_meta log output. Estimated: 3 lines.

P2-4: Prompt token growth unbounded

Status: TODO Impact: In 94-request sessions, prompt grows from 8K to 12.6K tokens Risk: May exceed model context window in very long sessions Fix: Consider conversation truncation strategy (drop old tool results, keep system + last N turns)

P2-5: Write tool latency (30-46s)

Status: ACCEPTED (inherent to 30B model) Impact: Large file writes take 30-46 seconds Mitigation: Could chunk writes, but complexity vs benefit is low

P3 — Nice to Have

P3-1: Tool variety

Status: TODO Impact: Model only uses Bash/Write, rarely Read/Glob/Grep Fix: Better tool guide examples in prompt

P3-2: Native streaming for tool-use

Status: TODO Impact: Currently forces sync mode for tool requests (sync-as-SSE) Fix: Investigate if Ollama supports streaming with function calling

P3-3: Empty session directory cleanup

Status: TODO Impact: 6 empty dirs from rapid restarts Fix: Don't create session dir until first request, or cleanup on shutdown

P3-4: Legacy debug file migration

Status: TODO Impact: ~189 files in root proxy-debug/ from before session dirs Fix: One-time cleanup script

Windows Support (v0.2)

Status: DONE

  • install.ps1 / install.cmd: Windows installer with winget-based dependency installation
  • vibe-local.ps1 / vibe-local.cmd: Windows launcher (PowerShell + CMD wrapper)
  • Proxy cross-platform fixes: Windows log paths, chmod tolerance, Windows OS hints
  • install.sh detects MINGW/MSYS/CYGWIN and redirects to PowerShell installer
  • .gitattributes for correct line endings (LF for .py/.sh, CRLF for .ps1/.cmd)

Proxy Improvements (v0.2)

ID Feature Status
A1 Init probe fast-path DONE
A2 tool_choice forwarding DONE
A3 stop_sequences forwarding DONE
A4 top_p / top_k forwarding DONE
A5 Token count via Ollama tokenize DONE
A6 Image/Vision support DONE
A7 Windows cross-platform DONE
A8 Security (0o700, 50MB limit) DONE

Metrics to Track

Metric Session 1 (pre-fix) Session 2 (post-fix) Target
Text-only responses 53% ~40% <20%
Linux-isms 7 0 0
"I cannot" patterns 3 0 0
Follow-up questions N/A 4 0
Tool variety (unique) 3/9 3/9 5/9
Avg response length ~200 chars ~150 chars <80 chars
Sidecar cold start 18.5s 18.5s <2s