@mmmbuto/nexuscrew 0.8.45 → 0.8.47

package/CHANGELOG.md

@@ -4,6 +4,52 @@ All notable changes to NexusCrew are tracked here.
 
 ## Unreleased
 
+## 0.8.47 — 2026-07-31 — "Kimi First-Run Corrective"
+
+- The Kimi engines (`kimi.native` and the Claude Code "Kimi Code" provider)
+  no longer receive the cell prompt on the process command line. The prompt
+  is delivered to the interactive prompt only, at most once per process
+  generation, after a real readiness classification of the visible terminal.
+- Login screens, the Claude custom-API-key consent dialog and
+  onboarding/trust dialogs are treated as **not ready**: nothing is pasted
+  and no Enter is sent, so a bootstrap prompt can never be lost behind a
+  dialog, and the cell session is left alive for the operator. When the
+  terminal stays not ready past a bounded wait, the up response carries a
+  bounded `actionRequired` code with a constant-catalog recovery hint — for
+  "Kimi Code" the `/config` → "Use custom API key" path (never the Anthropic
+  `/login` flow), for `kimi.native` the CLI's own device-code login.
+- Prompt delivery is a single bracketed paste followed by a separate Enter on
+  the exact pane, with a random per-send buffer and a `0600` temporary file.
+  There is no automatic retry after any paste attempt (a partial composer
+  state can never be doubled); a supervised restart performs at most one new
+  classified delivery for the new generation. Legacy `send-keys` custom
+  engines keep their paste-without-Enter contract on the same safer
+  transport.
+- The PWA surfaces the recovery hint when an up action reports
+  `actionRequired` (App, session list and Fleet tab).
+
+## 0.8.46 — 2026-07-31 — "Native Kimi"
+
+- The Fleet gains a native Kimi Code CLI engine (`kimi.native`) that launches
+  the official `@moonshot-ai/kimi-code` binary directly in the cell's working
+  directory. Authentication and providers stay with the CLI (device-code
+  login, `config.toml`); NexusCrew never reads, stores or injects Kimi
+  credentials, and existing installations receive the engine through an
+  idempotent, collision-safe backfill.
+- The engine defaults to interactive (standard) permissions. The unsafe
+  policy maps to `--yolo`, which auto-approves regular tool calls while the
+  agent may still ask questions; the fully autonomous `--auto` mode is
+  deliberately not exposed.
+- Because the CLI documents no interactive prompt flag (`kimi -p` is
+  non-interactive and skips the TUI), the cell prompt is never placed on the
+  command line: it is delivered with bracketed paste once the session is
+  ready, including on supervised restarts.
+- The Claude Code "Kimi Code" provider (`claude.kimi-code`) is unchanged and
+  remains the managed K3 path with the 1M-context profile. The PWA presents
+  the two paths under distinct names ("Kimi Code CLI" vs "Kimi Code") with an
+  explanatory notice, and login-delegated engines no longer show a misleading
+  credential KEY section.
+
 ## 0.8.45 — 2026-07-30 — "Resilient Directory"
 
 - Fleet cell-directory reads no longer wait for external model catalog discovery.

package/README.md

@@ -66,7 +66,7 @@ install-script approval.
 | **Live terminals** | Attach to real tmux sessions through a real PTY, WebSocket and xterm.js. |
 | **Persistent workspaces** | Arrange sessions into decks with saved layouts, ordering, pins and per-cell drafts. |
 | **Multi-node Fleet** | See and control authorized cells across Linux, macOS and Android nodes. |
-| **AI-ready cells** | Launch Claude Code, Codex, Codex-VL, Pi, Agy or a trusted shell with explicit providers and policies. |
+| **AI-ready cells** | Launch Claude Code, Codex, Codex-VL, Pi, Agy, Kimi Code CLI or a trusted shell with explicit providers and policies. |
 | **Mobile-native control** | Scroll tmux history, use terminal keys, dictate prompts and move files from a phone. |
 | **Operator alerts** | Receive visual, push, browser speech and opt-in node-native Audio Share TTS. |