vexi-cli 0.8.0 → 0.9.1

package/README.md

@@ -89,7 +89,7 @@ On first run, paste any API key. Vexi **auto-detects the provider** from the key
 
 Your key is stored **locally** in `~/.vexi/config.json` with owner-only file permissions (`chmod 600`).
 
-### 🆓 Start for free — no credit card needed
+### <img src="assets/icons/gift.svg" width="20" height="20" alt=""/> Start for free — no credit card needed
 
 **International:**
 
@@ -135,6 +135,8 @@ Vexi auto-detects your system language on first run.
 ```bash
 vexi                  # start a chat session in the current project
 vexi --lang es        # start in Spanish
+vexi -p "explain the auth flow"        # one-shot, non-interactive (scripts/CI)
+vexi -p "add a health endpoint" --yes  # same, but auto-run any shell commands proposed
 vexi config           # show config location + provider + model
 vexi config reset     # delete the stored API key
 vexi skill list       # show active skills
@@ -153,6 +155,10 @@ vexi undo                   # revert the last AI file edit
 vexi redo                   # re-apply a reverted edit
 vexi history                # list recent AI file edits with timestamps
 vexi clean                  # clear old snapshots (.vexi/snapshots/)
+vexi setup                  # configure any OpenAI-compatible endpoint by URL
+vexi update                 # update Vexi to the latest version
+vexi uninstall              # uninstall Vexi (keeps ~/.vexi config)
+vexi uninstall --purge      # uninstall + delete ~/.vexi
 ```
 
 Inside the chat:
@@ -165,10 +171,12 @@ Inside the chat:
 /undo     revert last AI file edit
 /redo     re-apply last undone edit
 /history  list recent AI file edits
+/push     stage, commit and push to git
+/usage    token & cost estimate for this session
 /exit     quit
 ```
 
-## ⚙️ Multi-language build support
+## <img src="assets/icons/settings.svg" width="22" height="22" alt=""/> Multi-language build support
 
 Vexi can build and run projects in **any language** — not just JavaScript. When the AI suggests commands, it wraps them in a shell block, Vexi asks for confirmation, then executes them automatically and feeds the output back to the AI.
 
@@ -184,7 +192,7 @@ Vexi can build and run projects in **any language** — not just JavaScript. Whe
 
 The project scanner automatically detects `.py`, `.java`, `.c`, `.cpp`, `.rs`, `.go` files and tells the AI what language your project uses before the first message.
 
-## 🧠 Project memory — Context Compression Engine
+## <img src="assets/icons/brain-circuit.svg" width="22" height="22" alt=""/> Project memory — Context Compression Engine
 
 Most AI coding tools forget earlier decisions once the conversation gets long.
 Vexi doesn't delete old messages — it **compresses** them:
@@ -196,7 +204,7 @@ Vexi doesn't delete old messages — it **compresses** them:
   your decisions **across sessions**, even in large projects.
 - Inspect it anytime with `/memory`.
 
-## 🗺️ Full project understanding
+## <img src="assets/icons/map.svg" width="22" height="22" alt=""/> Full project understanding
 
 On startup Vexi scans your whole project (not just the open file) and injects
 a compact map into every prompt: languages, frameworks, and architecture
@@ -206,7 +214,7 @@ Scanner safeguards: respects `.gitignore`, always skips `node_modules`,
 `.git`, `dist`, `build`, `coverage`, and ignores files larger than 500KB —
 so it never floods the context window.
 
-## 🎯 Custom Skills
+## <img src="assets/icons/target.svg" width="22" height="22" alt=""/> Custom Skills
 
 Teach Vexi *your* conventions with plain markdown files in `.vexi/skills/`:
 
@@ -225,7 +233,7 @@ vexi skill add https://github.com/user/react-best-practices
 vexi skill add ./docs/conventions.md
 ```
 
-## 🎬 Vexi Replay
+## <img src="assets/icons/clapperboard.svg" width="22" height="22" alt=""/> Vexi Replay
 
 Every chat session is automatically recorded to `.vexi/sessions/` (locally,
 nothing leaves your machine). Export any session as a **single standalone
@@ -242,7 +250,7 @@ and it ends with a session summary (duration, messages, model). An
 **Export video** button records the replay right in the browser
 (MediaRecorder — no ffmpeg, the CLI stays lightweight). Share it anywhere.
 
-## 🌍 Explain code in your native language
+## <img src="assets/icons/languages.svg" width="22" height="22" alt=""/> Explain code in your native language
 
 > The first AI tool that explains any code in your native language.
 
@@ -257,7 +265,7 @@ numbers → how the pieces fit together. Latin-script languages stream
 directly in the terminal; Arabic is written to `.md` + `.html` (dir="rtl")
 and opened in your browser, where it renders perfectly.
 
-## 🗺️ Visual code graph
+## <img src="assets/icons/network.svg" width="22" height="22" alt=""/> Visual code graph
 
 ```bash
 vexi graph --visual
@@ -269,7 +277,7 @@ search. Node heat shows how many files depend on each module, and clicking
 a node runs **impact analysis** — highlighting every file that breaks if
 you change it.
 
-## 🔌 MCP support
+## <img src="assets/icons/plug.svg" width="22" height="22" alt=""/> MCP support
 
 **Vexi as MCP client** — connect external tools and the AI can call them
 mid-conversation (works with every provider, no function-calling API needed):
@@ -301,7 +309,7 @@ official `@modelcontextprotocol/sdk`:
 Vexi **complements** Claude Code instead of competing: its project memory
 becomes a shared memory layer usable by any agent.
 
-## 🧠 Vexi Learn
+## <img src="assets/icons/graduation-cap.svg" width="22" height="22" alt=""/> Vexi Learn
 
 > The agent gets more *you* over time.
 
@@ -319,7 +327,39 @@ session, so you stop repeating yourself. Everything stays local — the only
 network call is to your own model provider, and you always preview before
 saving.
 
-## ↩️ Undo / Redo — instant rescue from any AI edit
+## <img src="assets/icons/link.svg" width="22" height="22" alt=""/> URL-based setup — any OpenAI-compatible endpoint
+
+```bash
+vexi setup
+```
+
+Paste any API endpoint URL and Vexi auto-detects the provider, fetches available models, and lets you pick one — no manual config editing required. Works with:
+
+- **Local models** — Ollama (`http://localhost:11434`), LM Studio, etc.
+- **Custom proxies** — your own OpenAI-compatible gateway
+- **Specialist providers** — Z.ai Coding Plan, Cloudflare Workers AI, and any `/v1/chat/completions`-compatible API
+
+```
+? Paste your endpoint URL › https://openrouter.ai/api/v1
+  Detected: OpenRouter
+? Paste your API key › sk-or-...
+  Fetching available models…
+? Choose a model › meta-llama/llama-3.3-70b-instruct:free
+  ✓ Saved — run `vexi` to start
+```
+
+## <img src="assets/icons/rocket.svg" width="22" height="22" alt=""/> Git push from chat — `/push`
+
+Stage, commit and push without leaving the session. The AI drafts the commit message from your staged diff:
+
+```
+/push                 # AI writes commit message, confirms, then pushes
+/push --only          # push current branch without a new commit
+```
+
+Auth pre-flight runs before the real push so you get a clear error if credentials are missing, not a mid-push failure.
+
+## <img src="assets/icons/undo-2.svg" width="22" height="22" alt=""/> Undo / Redo — instant rescue from any AI edit
 
 > Approve a change, see it break things, type `vexi undo`. Done.
 
@@ -342,6 +382,35 @@ Or use the in-chat shortcuts without leaving the session: `/undo`, `/redo`, `/hi
 
 This pairs naturally with the confirmation prompt: even if you approve a change that turns out to be wrong, one command gets you back.
 
+## <img src="assets/icons/file-check.svg" width="22" height="22" alt=""/> Reliable file edits
+
+Vexi edits files through structured tools — `read_file`, `write_file`, and
+`edit_file` (exact `old` → `new` substring replacement) — instead of fragile
+shell patching (`cat >`, `sed -i`). `edit_file` refuses a non-existent or
+ambiguous match rather than corrupting the file, and every write is snapshotted
+on the **exact** file about to change, so `/undo` is precise rather than
+regex-guessed.
+
+On providers that reliably support it (Anthropic, OpenAI, Groq, Gemini,
+Mistral, Cerebras, DeepSeek, Qwen, Kimi, GLM, OpenRouter), the tools run over
+**native function-calling**; everywhere else Vexi falls back to a
+provider-agnostic text protocol, so the same capabilities work with every
+provider.
+
+## <img src="assets/icons/coins.svg" width="22" height="22" alt=""/> Token & cost tracking
+
+Vexi accumulates the real token usage reported by your provider and estimates
+the dollar cost from a built-in price table:
+
+```
+/usage    → 12,340 tokens (8.1k in / 4.2k out) · ~$0.0187
+```
+
+Check it anytime with `/usage`, and a one-line summary prints when you exit
+(and to stderr in `-p` mode). Known free-tier models read as **$0**, and
+unpriced models still show token counts. Since you bring your own key, the
+`~` marks it as an estimate.
+
 ## Roadmap
 
 | Phase | Feature | Status |
@@ -353,21 +422,30 @@ This pairs naturally with the confirmation prompt: even if you approve a change
 | 5 | **Vexi Learn** — adapts to your personal coding style | ✅ done |
 | 6 | **Multi-language builds** — auto-executes pip, gcc, javac, cargo, gradle from chat | ✅ done |
 | 7 | **Undo / Redo** — instant one-command revert of any AI file edit, without touching git | ✅ done |
+| 8 | **`/push`** — stage, AI-draft commit message, and push to git from chat | ✅ done |
+| 9 | **`vexi setup`** — configure any OpenAI-compatible endpoint by pasting a URL | ✅ done |
+| 10 | **Self-maintaining** — `vexi update`, `vexi uninstall`, daily update-check notice | ✅ done |
+| 11 | **Reliable file edits** — structured read/write/edit tools with exact undo snapshots, native function-calling where the provider supports it | ✅ done |
+| 12 | **Token & cost tracking** (`/usage`) · remote model-default manifest (defaults refresh without a release) | ✅ done |
 
 ## Why Vexi?
 
 | | Vexi | OpenCode | Claude Code | Cursor |
 | --- | --- | --- | --- | --- |
 | Install | `npm i -g vexi-cli` | binary/script | `npm i -g` | desktop app |
-| BYOK (any provider) | ✅ 12 providers incl. Chinese AI | ✅ | ❌ Anthropic only | partial |
-| Works fully offline/local | ✅ no server, no account | ✅ | ❌ account | ❌ account |
-| Native-language code explanations | ✅ ar/es/pt/fr | ❌ | ❌ | ❌ |
+| BYOK (any provider) | ✅ 13 providers incl. Chinese AI + URL endpoints | ✅ | limited (defaults to Anthropic; Bedrock/Vertex/gateways possible) | partial |
+| No server, no account | ✅ no server, no account | ✅ | ❌ account | ❌ account |
+| Native-language code explanations (built-in) | ✅ ar/es/pt/fr | via model | via model | via model |
 | Session replay export | ✅ | ❌ | ❌ | ❌ |
 | Persistent project memory | ✅ | partial | partial | ✅ |
 | Learns your personal coding style | ✅ from your own sessions | ❌ | ❌ | partial |
-| MCP server mode (be a tool for other agents) | ✅ | ❌ | ❌ | ❌ |
-| Builds any language (Python, Java, C, Rust, Go) | ✅ | ❌ | ✅ | ✅ |
+| MCP server mode (be a tool for other agents) | ✅ | ❌ | ✅ `claude mcp serve` | ❌ |
+| Builds any language (Python, Java, C, Rust, Go) | ✅ | ✅ | ✅ | ✅ |
 | Instant undo/redo of AI edits (no git required) | ✅ per-file snapshots | ❌ | ❌ | ❌ |
+| Git push from chat (AI commit message) | ✅ `/push` | ❌ | ✅ | partial |
+| URL endpoint setup (Ollama, custom proxies) | ✅ `vexi setup` | ✅ | ❌ | ❌ |
+| Structured file-edit tools + exact undo | ✅ read/write/edit + per-file snapshots | partial | ✅ | ✅ |
+| Token & cost tracking (`/usage`) | ✅ per-session estimate | ❌ | ✅ | partial |
 | License | MIT | MIT | proprietary | proprietary |
 
 Vexi **complements** tools like Claude Code instead of competing: its project memory and multilingual explanations will be exposed over MCP so any agent can use them.
@@ -388,7 +466,10 @@ src/
 ├── cli.ts          CLI definition (commander)
 ├── agent.ts        chat loop + first-run onboarding
 ├── config.ts       ~/.vexi/config.json (atomic, chmod 600)
-├── providers/      key detection + streaming API clients
+├── version.ts      dynamic VERSION from package.json
+├── endpoint.ts     URL provider detection + model discovery (vexi setup)
+├── onboarding.ts   interactive URL-based setup wizard
+├── providers/      key detection + streaming API clients (13 providers)
 ├── scanner/        project mapper (.gitignore-aware, size-capped)
 ├── memory/         Context Compression Engine (.vexi/memory.json)
 ├── skills/         custom skills loader (.vexi/skills/*.md)
@@ -398,6 +479,8 @@ src/
 ├── mcp/            MCP client (tools in chat) + server mode
 ├── learn/          Vexi Learn — style mining from your own sessions
 ├── snapshots/      undo/redo engine — per-file backups before AI edits
+├── git/            /push command — stage, AI commit message, push
+├── update/         vexi update / uninstall + daily update-check
 ├── i18n/           5-language UI strings + RTL strategy
 ├── ui/             terminal branding (chalk, ora)
 └── utils/          atomic JSON writes, cross-platform open
@@ -412,19 +495,19 @@ To add support for a new key format, edit a single file: `src/providers/detect.t
 
 ---
 
-## 🌍 العربية
+## <img src="assets/icons/globe.svg" width="22" height="22" alt=""/> العربية
 
 **Vexi** — وكيل برمجة بالذكاء الاصطناعي مفتوح المصدر يعمل في الطرفية. ثبّته بأمر واحد (`npm install -g vexi-cli`)، الصق مفتاح API الخاص بك مرة واحدة، وابدأ فورًا. لا تسجيل، لا خادم، كل شيء يعمل محليًا على جهازك. يشرح Vexi أي كود بالعربية الفصحى (`vexi explain auth.ts --ar`) في ملفات HTML تدعم الاتجاه من اليمين لليسار بشكل مثالي، ويتعلّم أسلوبك البرمجي الشخصي من جلساتك السابقة (`vexi learn`).
 
-## 🌍 Español
+## <img src="assets/icons/globe.svg" width="22" height="22" alt=""/> Español
 
 **Vexi** es un agente de programación con IA, de código abierto, que vive en tu terminal. Instálalo con un solo comando (`npm install -g vexi-cli`), pega tu clave API una vez y empieza al instante. Sin registro, sin servidor: todo se ejecuta localmente. Vexi detecta tu proveedor automáticamente y habla tu idioma.
 
-## 🌍 Português
+## <img src="assets/icons/globe.svg" width="22" height="22" alt=""/> Português
 
 **Vexi** é um agente de programação com IA, de código aberto, que vive no seu terminal. Instale com um único comando (`npm install -g vexi-cli`), cole sua chave de API uma vez e comece imediatamente. Sem cadastro, sem servidor: tudo roda localmente. O Vexi detecta seu provedor automaticamente e fala o seu idioma.
 
-## 🌍 Français
+## <img src="assets/icons/globe.svg" width="22" height="22" alt=""/> Français
 
 **Vexi** est un agent de codage IA open source qui vit dans votre terminal. Installez-le en une seule commande (`npm install -g vexi-cli`), collez votre clé API une fois et commencez immédiatement. Pas de compte, pas de serveur : tout s'exécute localement. Vexi détecte automatiquement votre fournisseur et parle votre langue.
 
@@ -432,7 +515,7 @@ To add support for a new key format, edit a single file: `src/providers/detect.t
 
 <div align="center">
 
-**MIT License** · Made with ⚡ by the Vexi community
+**MIT License** · Made with <img src="assets/icons/zap.svg" width="16" height="16" alt="" style="vertical-align:text-bottom"/> by the Vexi community
 
 `npm install -g vexi-cli`