[ddehilster]

This lets a page that lists real documents open the studio on one of them, in the analyzer that reads it, ready to run and edit. The first user is TEG's NLP console: its Workbench has a button on every résumé.

The link names the analyzer: ?repo=owner/name&ref=branch&folder=path&handoff=1. It opens that analyzer from GitHub the way Open from GitHub does. If the person isn't signed in, the link is kept in sessionStorage across the sign-in round trip.

The text is never in the link. With handoff=1:

1. The studio posts nlp-studio:ready to window.opener.
2. It accepts {type: "nlp-studio:input", name, text} only if all of these hold:
   • it comes from the opener;
   • it comes from an origin listed in NLP_STUDIO_OPENERS (reported by /api/health; none by default);
   • it is at most 2 MB.
3. The name is reduced to safe characters, as input/<name>.txt.

A handed-in text is never kept. It is an input file while the page is open, and it is what a run reads. It is never:

• a browser draft;
• among the changed files;
• committed;
• in the Download zip.

The file list says so beside it. A reload (signing in is one) loses it, and the studio says ready again so the opener can hand it over again.

Changes

• src/handoff.ts: parsing the link, keeping it across sign-in, the name rule, and message acceptance.
• main.ts:
  • openLink;
  • handed-in inputs, excluded from drafts, changedPaths, the zip and the file list's normal Input rows.
• server/app.py: NLP_STUDIO_OPENERS → /api/health openers.
• docs/ARCHITECTURE.md: a new section. deploy/env.example: the new variable.

Tests

• tsc --noEmit is clean.
• vitest: 52 pass, 8 of them new for the hand-off (link parsing, sign-in round trip, the name rule, refusal from other windows and origins, the size limit).
• Server unittest: 68 pass.
• Not yet tried in a browser against a running studio.

🤖 Generated with Claude Code