T3 Code + Pydantic Logfire demo

Prepare the T3 Code + Pydantic Logfire title demo for me, ready to record. Keep the intentional bad-title starting state until I start the investigation. This is a demo-only branch; leave its PR unmerged.

Source: https://github.com/pingdotgg/t3code/pull/13671
Repository: https://github.com/SunkenInTime/t3code.git
Branch: demo/logfire-title-handoff
Pinned commit: b83de9d84a86a687642f3514094ae2480d816a13

1. Use a new clone in an unused directory. Check out the pinned commit on a new local branch. Preserve my existing work and T3 instances. Read AGENTS.md, demos/logfire-titles/README.md and docs/operations/logfire-demo.md there. This setup prompt authorizes preparing the demo, one three-case baseline evaluation, and a brief visual check of the isolated recording desktop. Stop before sending the investigation prompt.
2. Check Git, Node 24, vp, uv and authenticated Codex access to gpt-6-luna and gpt-6-astra. Report missing prerequisites or model access; keep the configured models. Run vp i in the clone.
3. Ask which Logfire project and region to use. Create .env.local from docs/operations/logfire-demo.env.example if absent. Let me supply my project's write token privately in that local file. Use the region-matching trace endpoint. This demo sends its constructed chats, model outputs, investigation messages and tool arguments/results to that project. Keep credentials and pairing URLs out of public artifacts and commits.
4. Check the existing Codex MCP configuration, reuse a matching Logfire connection, or add the regional Logfire MCP URL. For US: codex mcp add logfire --url https://logfire-us.pydantic.dev/mcp. Use codex mcp login logfire if needed and let me complete OAuth. Confirm the MCP can read the same project receiving telemetry.
5. Start node demos/logfire-titles/start.mjs and retain its process ID and log path. Read actual ports from its output. Build with vp run build:desktop, then start node demos/logfire-titles/desktop.mjs and retain that launcher PID and log. The recording desktop uses ports 14242/6202 and isolated .t3/recording-desktop state. If those ports belong to another running instance, report that conflict; preserve that instance. Wait for "Recording desktop ready". Read .t3/recording-desktop/recording.json for the investigation URL.
6. With the recording desktop running, run uv run demos/logfire-titles/evaluate.py --name baseline once. It makes one real Luna call for each of the three cases, exports to my Logfire project, and updates the recording sidebar. Save its results and evaluation URL. The reference baseline passes 1/3 subject checks; report the actual result if mine differs. Confirm there are three complete results and zero task errors. Use Logfire MCP to verify all three recent title spans contain the actual inputs and outputs; this is the evidence the investigator will need.
7. Visually check the recording desktop: exactly three example chats and one empty investigator, Astra / Medium / Full access selected, project bound to this new clone. Open "Fix Stale Label Palette Cache" and confirm the message starts with an offline-search request. If regenerated wording differs, open the offline-search case by its ID, logfire-title-initial-offline-search. Leave the investigator empty. Preserve the intentional regression.

Finish with the checkout path, running process IDs, the recording investigation URL, my Logfire Live and evaluation links, the actual baseline score, and the recording prompt from demos/logfire-titles/investigate.md. Include my Logfire project name in that prompt if needed. Return any pairing URL privately without consuming it. Setup is complete when the desktop starting state and hosted baseline evidence are both verified.