0.13.0

Added

• wa mcp with no profile serves every profile, on stdio and with --http (cli-core 0.450.0). An AI app needs one
  connector for WhatsApp, not one per profile. Each tool takes a required profile (every profile here, no default)
  and the agent is told to ask when you did not name the account; the call runs in that profile's own
  wa <profile> mcp, so its permissions apply as before. WhatsApp allows one connection per account: run
  wa <profile> serve while another wa process may hold it. --public-url takes a path, https://<name>.ts.net/wa.
• An unlinked profile's error also offers WireCat Desktop. is not linked to WhatsApp now says: run wa <profile> setup in a terminal, or link from WireCat Desktop (Tools → WhatsApp → Manage). The skill no longer says chats delete|clear are left out of MCP: at level ask MCP runs them.
• wa skill install --for hermes gives Hermes Agent the skill, and wa skill status says what Hermes reads (cli-base 0.59.0). It adds ~/.agents/skills to Hermes's skills.external_dirs through hermes config set, keeping the folders already listed, or copies the file into Hermes's home when that fails. With no Hermes installed it fails with not_found and writes nothing. A bare wa skill install and --for all still write only the Claude Code and ~/.agents/skills copies. skill status --json gains hermes beside targets.
• A run started by OpenClaw or Hermes Agent names that agent (cli-core 0.437.0). runs show gives agent as openclaw for a command OpenClaw's agent ran, with from_chat: true when a chat message started it; a Hermes run carries its session, so a task claim from Hermes is held per session.
• The whatsapp-cli skill tells OpenClaw and ClawHub that it needs wa, and how to install it (metadata.openclaw: requires.bins and the npm package @wirecat/whatsapp-cli). With wa missing from PATH, OpenClaw marks the skill not ready and offers to install the package. Claude Code and Codex ignore the block.
• wa <profile> mcp server start|stop|status runs the HTTP MCP server in the background (cli-core 0.445.0). start takes --public-url, --port and --permission as mcp --http does and returns once the server listens; status gives the port, the address, the one-time login code and the apps that logged in; stop ends only the server it started.

Changed — may break callers

• wa needs cli-base 0.63.0 (cli-core 0.452.0). Nothing wa uses changed in it.

• A run record keeps the command line as typed (cli-core 0.439.0 to 0.447.0). runs show and run.json carry argv, the words after wa, with secrets [redacted] (a secret-named option, a password in a URL, a known token format, a random-looking string) and a phone number never kept. wa <profile> config set record_input false keeps [redacted] in place of every typed value that may be text. Recorded events still hold ids, never a chat's title or a message's text.

Fixed

• Search by meaning and --discover run on the native ONNX Runtime on Windows too (cli-core 0.451.0). After
  wa <profile> models download, text models and the reranker use the native library on Windows x64 and arm64, as
  they already did on Linux and macOS: 2–3× faster than WebAssembly. The download takes onnxruntime.dll from the
  onnxruntime-node npm package (113.5 MB, once, checked by sha256). Picture models (CLIP, SigLIP2) use the native
  library as well, wherever it is installed. Results are the same.

• Every command spends less time getting ready (cli-base 0.62.0, cli-core 0.449.0). The check of shared commands now
  describes each command once, not again for every group above it. It was about 60% of each command's setup; in tg a
  warm run(["--version"]) went from 49 to 29 ms. It catches the same changes as before.

• Search by meaning and search messages --discover's reranker run 2–3× faster on Linux and Apple-silicon Macs after the next wa models download (cli-core 0.448.0). The download installs a native ONNX Runtime, and the models run on it; without it they run as before. The reranker now reads only the start of the message a hit replies to (cli-core 0.447.0).