Use UI

Say “use ui” to request a local browser interview, or use /tdt-ui. You can answer in chat/TUI instead. No stack or additional model subscription is needed. The active agent must wait/read responses; a browser cannot wake a stopped agent. In Codex, use $tdt-ui or the skill picker. Describe the interview you want; the skill prepares the page, reads your answers and handles follow-up questions. Ask it to resume an interrupted interview, close a session or clean up retained responses when you no longer need them.

Technical reference: session commands and page examples

The skill uses the commands below. You only need these details when operating UI directly or authoring a custom page.

tdt ui start --no-open
tdt ui present SESSION business.json
tdt ui wait SESSION --after 0 --timeout 20
tdt ui ack SESSION 1
tdt ui present SESSION follow-up.json
tdt ui close SESSION
tdt ui cleanup SESSION

Use the real session ID returned by start and open its URL. Keep the URL private. The wait command above is one iteration: while the interview is waiting and no events arrive, the agent repeats it with the same cursor. A 20-second timeout ends that command, not the interview or the agent turn. This applies after every follow-up page too. The agent should keep waiting until submission, cancellation, chat fallback or an explicit pause, with occasional progress updates. Finished or cancelled status ends the loop. Host interruption or execution limits can still stop the agent; in that case send a chat message to resume and read the retained response. Browser submission alone does not restart a stopped host. Read the full public contract at .tdt/contracts/ui.md. Resume after a crash with tdt ui start --session SESSION --no-open and open the new URL. Read retained answers with tdt ui read SESSION --after 0. A stopped or expired server does not remove answers. Cleanup deletes that session's retained answers; close only stops the service. Choose a shorter --idle when appropriate.

Business plan, with zero stacks

Save this as business.json, present it, and use the submitted audience and goal to ask about the first offer and price or write the first business-plan paragraph.

{"version":1,"title":"Shape your business","description":"Start with the people you want to help.","steps":[{"title":"People and purpose","fields":[{"id":"audience","type":"text","label":"Who will you help?","required":true},{"id":"goal","type":"multiline","label":"What problem will you solve?","required":true},{"id":"stage","type":"single","label":"Where are you today?","options":["An idea","First customers","Growing"],"alternative":true}]}]}

Multiple questions, one answer payload

A page can contain many fields across grouped steps. Review displays all answers; one explicit submit returns them together as event.answers, keyed by stable field IDs. For the business page above, the answers portion could be:

{"audience":"Neighbourhood bakeries","goal":"Reduce unsold bread","stage":"An idea"}

The event also includes session, round, event and submission IDs, the action, creation time, and optional custom data. The agent should use the answers together to advance the work. Read/wait includes the original page under prompts so follow-ups do not lose the meaning of an earlier answer.

Showing code and Markdown

In standard page description, field help, and answer previews, wrap code or literal Markdown in triple-backtick fences, with an optional language such as python, json or markdown. UI shows an inset monospace block, preserves indentation and allows horizontal scrolling. Use single backticks for inline code. Other Markdown stays literal; text is never interpreted as HTML.

For example, a JSON description can be:

"description": "Review this command:\n\n```sh\ntdt doctor\n```"

Custom pages use <pre class="tdt-code"><code></code></pre> and set the code element's textContent to the source. The shared styles are available inside the sandbox too. Do not inject user text through innerHTML.

Custom interaction

A stack can list a JSON template containing custom_html in its ordinary templates array; it needs no hooks or manifest changes. This two-dimensional priority pad submits structured custom data, then the agent can interpret the selected urgency/impact and ask a follow-up.

{"version": 1, "title": "Choose a priority", "custom_html": "<main><h2>Where does this idea sit?</h2><p>Click the pad, or focus it and use arrow keys. Left/right changes urgency; up/down changes impact.</p><button id='pad' style='width:100%;height:180px;background:var(--inset)' aria-label='Priority pad'>Select priority</button><p id='selection' role='status'>No selection</p><button id='send' disabled>Submit priority</button></main><script>const pad=document.getElementById('pad'),out=document.getElementById('selection'),send=document.getElementById('send');let point;function show(){out.textContent=JSON.stringify(point);send.disabled=false}pad.onclick=e=>{if(e.detail===0){point=point||{urgency:50,impact:50}}else{const r=pad.getBoundingClientRect();point={urgency:Math.round(100*(e.clientX-r.left)/r.width),impact:Math.round(100*(1-(e.clientY-r.top)/r.height))}}show()};pad.onkeydown=e=>{if(!['ArrowUp','ArrowDown','ArrowLeft','ArrowRight'].includes(e.key))return;e.preventDefault();point=point||{urgency:50,impact:50};const k=e.key==='ArrowLeft'||e.key==='ArrowRight'?'urgency':'impact';point[k]=Math.max(0,Math.min(100,point[k]+(['ArrowRight','ArrowUp'].includes(e.key)?5:-5)));show()};send.onclick=()=>tdt.submit(point);</script>"}

The pad supports keyboard selection and exposes its selected values as status text. Supply comparable keyboard alternatives for your own custom interactions. Custom code is optional executable content: keep it small and inspect it first. Installing a stack never starts its interview or executes its custom templates.

When switching to chat, your agent retains submitted answers. Unsubmitted edits remain only in the browser tab. Submit those explicitly if you want to carry them over; showing a default or changing a control does not authorize an action. Responses are local interview state, not approved brain knowledge. Avoid entering secrets. After completion, close the session and clean up when no longer needed.