v1.19.0

Added

• Handoffs can pick themselves up. When the context watch offers a handoff and you
  say yes, it's marked **Autostart:** yes. Type /clear or open a new session and Claude
  starts on it without you typing "carry on", and the session is titled after the
  handoff. It fires once per handoff and only within 12 hours; under claude -p a marked
  handoff becomes the first turn. Turn it off with CLAW_NO_HANDOFF_AUTOSTART=1.
• Rules for building Claude Code plugins, hooks and skills (claude-code domain, 6
  rules), on automatically in any project with a .claude-plugin/ manifest,
  hooks/hooks.json or a SKILL.md.
• /clawness:messaging rewrites landing and UI copy around your customer. It scans
  your existing copy, drafts a seven-part message brief (customer, problem, guide, plan,
  call to action, stakes, success), and changes nothing until you approve it.
• /clawness:security-audit runs your installed SAST tools. If Semgrep, Bandit,
  Gitleaks, Trivy or CodeQL is on PATH, the audit runs it, merges the SARIF into the
  findings ledger and has it adjudicated with everything else. Nothing is installed for
  you, and Semgrep runs with p/default --metrics=off rather than --config auto.
• Asking for SAST or a named scanner now gets the same treatment, including
  third-party security skills: Claude offers the install command and audits the skill
  before using it.
• No scanner installed? Clawness offers one, once. The first time you ask for a
  security scan in a project without Semgrep, Bandit, Gitleaks, Trivy or CodeQL, Claude
  offers a one-line Semgrep install. Decline and the audit runs on Clawness's own scan.
  Silence it with CLAW_NO_SAST_OFFER=1.
• Security questions about your current changes point to /security-review. Asking
  whether a change, diff or PR is secure now names Claude Code's built-in review as the
  quick option, alongside the whole-repo /clawness:security-audit.

Changed

• The plan gate now applies in auto mode. Claude Code starts sessions in auto mode
  by default, and the gate treated auto as pre-approved, so it had stopped firing. It
  now asks once; approve a plan and auto runs freely. Unattended claude -p runs that
  start in auto need --permission-mode acceptEdits or CLAW_NO_PLAN_GATE=1.
• Replies default to under 100 words. The always-on voice rule now also rules out
  caveats and options you didn't ask for, and defines jargon in a clause. Longer replies
  still come when you ask for detail or the content needs it.

Fixed

• Clawness's slash commands work in the first session after install. Skills now run
  the CLI bundled with the plugin directly, instead of waiting for a session-start hook
  to write a helper first. Older Claude Code versions still use that helper.
• The context warning no longer false-alarms on 1M-window models. Opus 4.7+,
  Sonnet 5+ and Fable sessions run with a 1M window by default, but Clawness assumed
  200k unless model in settings said [1m]. A session at 158k was told "79% full".
  The window is now read from the session's model; CLAUDE_CODE_DISABLE_1M_CONTEXT=1
  is honoured.