@xingjisoft/neutrino 0.1.4 → 0.1.6

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@xingjisoft/neutrino",
-  "version": "0.1.4",
+  "version": "0.1.6",
   "description": "XINGJI Neutrino - a CLI coding agent in the terminal.",
   "type": "module",
   "license": "Apache-2.0",
@@ -10,15 +10,17 @@
   "main": "dist/app.js",
   "files": [
     "dist",
-    "assets"
+    "assets",
+    "web"
   ],
   "engines": {
     "node": ">=18.0.0"
   },
   "scripts": {
     "build": "tsc -p tsconfig.json",
-    "test": "npm run build && node scripts/test-paste.mjs && node scripts/test-mouse.mjs && node scripts/test-cursor.mjs && node scripts/test-selection.mjs && node scripts/test-copy.mjs && node scripts/test-scrollbar.mjs && node scripts/test-render.mjs && node scripts/test-drift.mjs && node scripts/test-sanitize.mjs && node scripts/test-markdown.mjs && node scripts/test-wrap.mjs && node scripts/test-fetch.mjs && node scripts/test-compact.mjs && node scripts/test-tools.mjs",
-    "prepare": "npm run build"
+    "test": "npm run build && node scripts/test-paste.mjs && node scripts/test-mouse.mjs && node scripts/test-cursor.mjs && node scripts/test-selection.mjs && node scripts/test-drag-scroll.mjs && node scripts/test-selection-performance.mjs && node scripts/test-copy.mjs && node scripts/test-scrollbar.mjs && node scripts/test-render.mjs && node scripts/test-drift.mjs && node scripts/test-sanitize.mjs && node scripts/test-markdown.mjs && node scripts/test-wrap.mjs && node scripts/test-fetch.mjs && node scripts/test-compact.mjs && node scripts/test-history-fold.mjs && node scripts/test-web-history.mjs && node scripts/test-turn-fold.mjs && node scripts/test-web-settings.mjs && node scripts/test-tools.mjs && node scripts/test-approval.mjs && node scripts/test-approval-review.mjs && node scripts/test-approval-ui.mjs && node scripts/test-web-approval-details.mjs && node scripts/test-approval-mode.mjs && node scripts/test-approval-context.mjs && node scripts/test-settings-render.mjs && node scripts/test-approval-flow.mjs && node scripts/test-web.mjs && node scripts/test-web-settings-http.mjs && node scripts/test-web-auth.mjs && node scripts/test-web-auth-ui.mjs && node scripts/test-interrupt.mjs && node scripts/test-terminal-lifecycle.mjs && node scripts/test-persistence.mjs && node scripts/test-resilience.mjs && node scripts/test-approval-timeout.mjs && node scripts/test-approval-hardening.mjs && node scripts/test-approval-extreme.mjs && node scripts/test-approval-reason.mjs",
+    "prepare": "npm run build",
+    "web": "node dist/bin.js web"
   },
   "keywords": [
     "cli",

package/web/README.md

@@ -0,0 +1,210 @@
+# Neutrino Web UI
+
+A small, dependency-free single page app that is served by the Neutrino CLI on
+a port you choose. It lets you:
+
+- browse every saved conversation grouped by working directory (like Codex),
+  with only the five latest transcripts per path shown and the rest collapsed;
+- watch a running terminal conversation stream in real time, with every
+  finished turn folded into one summary line that opens on a click (the turn
+  being answered stays open, and a paused turn is not folded either);
+- switch the provider, model and intensity for the next turn;
+- edit defaults, context compression and approval settings in a Settings dialog;
+- drive the running conversation: pause / resume the current turn and send new
+  messages (they queue while the agent is busy);
+- pick up a saved conversation in the browser and keep chatting, even when no
+  terminal is running.
+
+Nothing needs to be built — these are plain `.html`, `.css` and `.js` files.
+
+## Running it
+
+While using the terminal UI, mount the website on a port:
+
+```sh
+neutrino --web-port 8765
+```
+
+Or serve only the saved history, without a terminal:
+
+```sh
+neutrino web --port 8765
+# or
+npm run web
+```
+
+Then open <http://127.0.0.1:8765>.
+
+## Using it from another device
+
+The server binds to loopback by default. To open it from your phone or another
+machine, bind an interface others can reach:
+
+```sh
+neutrino web --port 8765 --lan               # every interface (0.0.0.0)
+neutrino web --port 8765 --host 100.64.0.7   # one specific address
+```
+
+Startup prints the addresses to open:
+
+```
+Neutrino web UI listening on 0.0.0.0:8765
+Open the address that matches your device:
+  http://192.168.1.211:8765/
+  http://127.0.0.1:8765/
+```
+
+Hypervisor and container host-only adapters (`vEthernet`, VMware, VirtualBox,
+Docker), link-local `169.254/16` and loopback addresses other than `127.0.0.1`
+are left out of that list on purpose: they cannot be reached from another
+device. CGNAT ranges such as Tailscale's `100.64/10` are kept, since a VPN is a
+common way to reach a machine.
+
+## Authentication
+
+The Web UI uses a permanent password, initially **`114514`**, on both localhost
+and network interfaces. Change it in **Settings → Web access** in the browser
+or **Settings → Web password** in the terminal. In the browser, enter the new
+password twice and save; leaving both fields blank preserves the current one.
+In the terminal, select the password row, press Enter, type or paste a new
+password, and press Enter to save (Escape cancels).
+
+Only a salted scrypt hash is stored in `~/.neutrino/config.json`. The password
+does not expire and is not reset when the server restarts. Settings never send
+back the password or its hash.
+
+A modal login form stays above the conversation UI. Successful login sets an
+HttpOnly, SameSite=Strict cookie, remembered for up to one year and shared by
+HTTP requests and the SSE stream. Session credentials are never put into URLs
+or browser localStorage. A password change invalidates existing login cookies; the browser
+that saves the password receives a replacement cookie and stays signed in.
+
+To sign in directly from a link, open `http://<host>:8765/?pwd=114514`.
+The browser automatically checks the supplied password through the login endpoint
+before loading conversations, remembers the login cookie, and removes `pwd`
+from the address bar while keeping other query parameters and the URL fragment.
+An incorrect password or a failed request shows the normal login dialog with
+an error, so you can enter the password manually. URL-encode passwords that
+contain special characters (for example, `+` becomes `%2B` and `&` becomes `%26`).
+
+For scripts, POST `{"password":"114514"}` to `/api/auth/login` and retain the
+returned cookie for subsequent requests. POST `/api/auth/logout` clears the
+browser cookie. Missing or invalid login cookies receive HTTP 401. Legacy token
+headers, token URLs, `--token` and `NEUTRINO_WEB_TOKEN` are no longer used.
+
+Static files and the login/logout routes stay public. `--no-auth` explicitly
+disables password login.
+
+## Models and settings
+
+Open a conversation to see the **Provider**, **Model** and **Thinking intensity**
+selectors. Choose a configured model and press **Apply to conversation**. The
+next turn uses that selection, and it is saved with the transcript. The same
+controls also update a conversation running in the terminal. Pause an active
+turn before applying a change; a running turn returns a clear conflict message.
+
+For saved history, **Continue with model** resumes the transcript with the chosen
+selection. Changing the selectors and then using **Continue here** or sending a
+message also uses the draft selection, so a transcript whose old model was
+removed can still be continued.
+
+The **⚙ Settings** button in the sidebar opens defaults for new conversations,
+the context compression threshold, the approval reviewer model, and the
+extreme-risk automatic approval option. Save settings explicitly; cancelling
+leaves the saved values intact. Conversation model changes are separate from
+these defaults. Models come from the same catalog as the terminal, and API keys
+stay on the server. Configure missing keys in the terminal model screen first.
+Intensity uses the same Low–Ultra sampling settings as the terminal.
+
+## History vs. live sessions
+
+A saved transcript on disk is a finished record: nothing is running, so the
+page shows it read-only (dashed `read-only` badge) and there is nothing to
+pause.
+
+To continue one, either
+
+- press **✎ Continue here** in the header, or
+- just start typing in the composer and press Enter.
+
+Both create a *browser-owned session*: the server resumes the transcript from
+disk, registers it with the live hub, and runs the agent loop headlessly with
+the model, provider, intensity, working directory and token counters recorded in
+the file. From that moment the page behaves exactly like a live terminal
+conversation — streaming output, tool calls, pause/resume, queued messages — and
+every turn is written back to the same transcript file.
+
+If the same transcript is already live (a terminal attached, or another tab
+resumed it), the page attaches to that session instead of starting a second one.
+Browser-owned sessions are dropped after 30 idle minutes, or immediately through
+`POST /api/live/:key/close`. Sessions owned by the terminal cannot be closed from
+the web.
+
+## How real-time works
+
+A `ChatScreen` registers itself with the in-process hub when it is opened. Every
+time it changes (a streamed token, a tool call, a pause, a save) it notifies the
+hub, which coalesces updates and pushes a snapshot over
+`GET /api/live/:key/stream` (SSE). The page keeps the open conversation in sync
+and re-renders it; the sidebar polls `/api/overview` for new sessions.
+
+Browser-owned sessions publish to the same hub, so the page code does not need
+to know whether a conversation is driven by the terminal or by the server.
+
+## API
+
+| Method | Path | Purpose |
+| --- | --- | --- |
+| GET | `/api/overview` | Sessions grouped by cwd + live sessions |
+| GET | `/api/sessions` | Flat list of saved sessions |
+| GET | `/api/settings` | Model catalog and public settings; no credentials |
+| POST | `/api/settings` | Save validated `{ selection?, settings? }` defaults |
+| GET | `/api/session/:id` | One saved transcript |
+| POST | `/api/session/:id/continue` | Resume a transcript with optional `{ selection }`, returns `{ key }` |
+| GET | `/api/live/:key` | Snapshot of a live session |
+| GET | `/api/live/:key/stream` | SSE stream of a live session |
+| POST | `/api/live/:key/pause` | Pause the running turn |
+| POST | `/api/live/:key/resume` | Resume a paused conversation |
+| POST | `/api/live/:key/message` | Send `{ "text": "…" }` |
+| POST | `/api/live/:key/model` | Apply `{ provider, model, intensity }` while idle or paused |
+| POST | `/api/live/:key/close` | Dispose a browser-owned session |
+| GET | `/api/health` | Liveness probe |
+| POST | `/api/auth/login` | Sign in with `{ "password": "…" }` |
+| POST | `/api/auth/logout` | Clear the login cookie |
+| GET / POST | `/api/settings` | Read settings or update them, including `settings.web_password` |
+
+Conversation and settings routes require the login cookie; see
+[Authentication](#authentication). `OPTIONS` is answered unauthenticated so
+preflights are not blocked, and it returns no data.
+
+### Approval requests
+
+When a running agent wants to do something risky (a delete-class command, an
+install, a path outside the working directory) the browser shows an orange
+**Request Approval** banner above the composer, showing the full command (or
+formatted tool arguments), working directory and risk reasons, plus two buttons.
+Long proposals remain complete and scroll inside the card. Automatic reviews
+show the same proposal details while they run.
+`Approve` runs the call, `Deny` returns a note to the agent.
+The reviewer uses the current turn's captured user request and earlier user
+messages; queued input belongs to the next turn. Expand a denied tool result
+to inspect the user requests that were reviewed.
+
+Use the **Request Approval / Auto Approval** switch in each conversation's
+controls to choose human approval or independent AI review. The choice is saved
+with the conversation and synced with the terminal. `自动审批` / `自动批准` in
+prompt text no longer changes this mode. Delete-class automatic review still
+requires **Settings → Extreme-risk auto-approval**. Each pending call receives two
+independent safety assessments. A round passes at an average safety score of
+60/100; task relevance has no weight or ordinary minimum. Only an extreme,
+obvious task mismatch independently confirmed with evidence by both reviewers
+(relevance at most 5/100 each) can veto on intent. Concrete serious hazards,
+explicit user prohibitions and failed reviews still prevent execution.
+
+It is the same gate the terminal uses: the prompt opt-outs `无需审批` and
+`我已知悉风险，同意给予完整权限，无需任何审批` still work, and
+`POST /api/live/:key/approve` with `{"approved": true|false}` answers the request
+programmatically (409 when nothing is pending).
+
+`POST /api/live/:key/approval-mode` with `{"mode":"request"|"auto"}` changes
+the conversation mode and returns its current snapshot (400 for invalid modes).