[davior]

Summary

GVC's half of M3c (davior/gam#61 is GAM's). No suite change was needed: the suite already served web search and personas to apps and applied a persona_id in chat. Merging this changes nothing until AI_BACKEND=suite (the same switch as the assistant's model and the narration).

• web_search, a reading tool. For what the library can't say: a fact for the narration, who someone is, when something happened. The suite runs it, as the signed-in user, credited to the project. It is offered only where there is a suite to run it (tools.specs_for), so a model is never told about a tool it can't use. Results come back numbered, for the model to cite and capped at a page; a failed search is a result the model reads, not a failed turn; a request may make only eight. It never changes the video, and the prompt says to say where a fact came from, to put web facts in the script only if the person asked for facts, and never to treat the web as footage.
• Personas. A turn can be asked as one of the suite's personas, chosen in a picker above the message box that appears only when the suite has some. The suite applies the persona's system prompt first and its model for the task; this app only names it. The clip-picking pass does not use the persona (it is a smaller job with its own task).
  • Kept on the turn row (new nullable column, migration b6d29f0c4e17), so a turn a restart reruns is asked as the same persona. It belongs to the turn: the next message is asked as whatever the picker then says.
  • GET /api/assistant/personas returns only what a picker shows (id, name, description, avatar), never a system prompt, and is empty for any reason at all (none made, the suite down, no session token, not in suite mode), so a missing picker is never an error.
  • A persona deleted in the suite since it was chosen is dropped from the picker's state on the next load; one the suite refuses mid-turn ends that turn with the suite's own sentence.

Test plan

• cd backend && pytest -q: 821 passed (9 new: the search end to end (credited to GVC and to the project, numbered results, a step, the video untouched), a failed search read by the model, the tool offered only where there is a web, the search cap and an empty query, the picker without prompts, no personas when none/down/no session/local, a persona sent to the suite and kept per turn, a refused persona's sentence).
• cd frontend && npm run format:check && npm run lint && npm test && npm run build: 217 tests passed (4 new).
• CI green, including the Docker image build.
• A real search and a real persona against a real suite. Not run here: everything is against a faked suite. (To try it: make a persona on the suite's side first.)

Deployment

Merging and redeploying changes nothing until you use the assistant on the suite.

1. Before pulling: nothing; the suite needs no change. To use personas, create one in the suite (for Video Creator or for every app).
2. Configuration: nothing new.
3. Deploy: rebuilds the backend and the frontend; the server needs outbound access to github.com for the client:

   git pull origin main

   docker compose up --build -d

   docker compose ps

   curl -f http://localhost:${APP_PORT}/api/health

4. Migration: b6d29f0c4e17 (add_assistant_persona) runs on container start and adds one nullable column, persona_id, to assistant_turn. Existing turns read as "no persona". It downgrades by dropping the column; nothing needs backing up first.
5. Other apps: none.
6. Verify: with AI_BACKEND=suite, ask the assistant "write a line about the history of water wars". A step says it searched the web, and the reply says where a fact came from. With a persona made in the suite, a picker appears above the message box; ask as it, and the suite's Usage shows the call under GVC. With AI_BACKEND=local there is no picker and no web tool.
7. Roll back: the previous commit (the column is harmless if left).

---

---

Generated by Claude Code