[DeyangChan]

What this changes

Adds /vstack:proofread. It passes the user-facing copy an agent is writing to ChatGPT in the user's own browser. The agent then uses the reply as a reference to improve its copy.

• The run is automatic. It uses the connected browser, and stops only for secrets, a signed-out browser, or a failed step.
• A hash confirms the copy reached the page unchanged.
• Working files go under .vstack/local/proofread/, through lib/workdir.mjs.
• contracts/host.md gains an optional user_browser op. The Claude Code adapter maps it to Claude in Chrome.
• The Codex and Grok adapters mark it Not available, because it has not been tested there. The skill says so and stops.

The version goes to 6.9.0, because this is new behaviour.

How it was tested

On Claude Code, with Claude in Chrome, in a signed-in Chrome profile:

• The flow ran end to end on a sample document, through open, insert, send, capture, and edit.
• A long generated document was tested separately on the attachment route. The reply kept all 65 sections, links, and code spans.
• node plugins/vstack/skills/proofread/tests/proofread.mjs covers the script and the host rules.

It has not been run on Codex or Grok. ChatGPT's page is not covered by CI, because that needs a real sign-in.

Checklist

• Nothing writes outside the user's project, and nothing transmits anywhere, unless that is the skill's stated purpose. Sending the copy to ChatGPT is its stated purpose.
• Per-machine state resolves through lib/workdir.mjs and lands under .vstack/local/<tool>/.
• Changed something under plugins/, and raised version in both host manifests with a matching CHANGELOG.md entry. Merging this publishes it.
• The security scans that run locally pass, and no finding was silenced.
• Installed the plugin from this branch and drove the affected skill end to end. The install was rehearsed, and it lists the skill. The skill's scripts were then driven by hand against ChatGPT, rather than through an invocation of the installed skill.

🤖 Generated with Claude Code