@cordfuse/llmux 0.13.5 → 0.13.7

package/README.md

@@ -13,23 +13,26 @@ shared state. Drive any of them from a terminal (`llmux session attach`),
 a REST or WebSocket API, or a phone browser over Tailscale. Sessions
 survive daemon restarts; attach is raw-TTY (`Ctrl+]` to detach).
 
-### `claude -p` is metered. llmux isn't.
+### Headless ≠ `claude -p`
 
-Every major agent CLI ships a one-shot mode — `claude -p "prompt"` and
-the codex / gemini / qwen / amp equivalents. The catch: even when you
-authed via OAuth on a flat-rate subscription (Claude Pro/Max, ChatGPT
-Plus, Gemini Advanced, …), `-p`-style calls now route to the **metered
-API billing bucket** — real per-token dollars on top of your monthly
-subscription. Script a few hundred `-p` calls a day and the subscription
-stops mattering.
+The obvious way to script Claude Code is `claude -p "prompt"` (and similar
+non-interactive modes in codex, gemini, etc.). Each call spawns a fresh
+short-lived child — no shared conversation, no in-session OAuth, no
+`/commands`, no persistent tool state, no MCP context.
 
 llmux drives the **interactive** agent process — the same TUI a human
-launches — over `tmux send-keys`. Interactive use stays on the
-subscription billing path: same OAuth, same flat rate, no per-token API
-surcharge. Spawn `claude` once, fire prompts at the same live agent
-forever from any client (CLI, REST, WebSocket, web). The agent runs
-unmodified and doesn't know it's being driven headlessly. Tool state and
-conversation persist across prompts; resume from any client.
+launches — over `tmux send-keys`. Spawn `claude` once, fire prompts at the
+same live agent forever from any client (CLI, REST, WebSocket, web).
+The agent runs unmodified and doesn't know it's being driven headlessly.
+Tool state persists across prompts. Conversations are resumable from any
+client.
+
+That same surface is the substrate higher-level orchestration sits on top
+of — spec-driven development (SDD) pipelines, multi-agent chains,
+scheduled jobs, evals harnessed against live agents — all just `llmux
+session prompt <name> "..."` calls. No special agent build, no headless
+SDK, no mode flag. If a human can drive the agent, llmux can drive it
+the same way.
 
 ### OAuth from your phone, on a headless box
 
@@ -46,7 +49,7 @@ toolbar (Esc / Tab / Ctrl / arrows / shell chars), watch tool calls
 stream in. No "mobile app" — it's the same daemon serving a real
 terminal over a WebSocket.
 
-> **Status:** v0.13.5 — daemon + CLI client consolidated into one binary
+> **Status:** v0.13.7 — daemon + CLI client consolidated into one binary
 > (`llmux`). Auth, tokens, mobile picker, conversation resume, Claude Code
 > history adapter shipped. See [CHANGELOG.md](./CHANGELOG.md).
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cordfuse/llmux",
-  "version": "0.13.5",
+  "version": "0.13.7",
   "description": "tmux-based AI agent dispatcher — REST/WS daemon + CLI client in one binary",
   "license": "MIT",
   "type": "module",