EveStudioDocs

Chat

The Chat tab is where you talk to an agent and watch it work. It streams every step of a turn (tool calls, subagent delegations, reasoning, and approval prompts) against either your local dev server or the deployed agent, and gives you controls over the session itself.

Local vs Deployed

A Target toggle at the top picks what you're talking to:

  • Local: the agent's local dev server (eve dev). This is what you use while building; it reflects your unsaved-to-production changes. The dev server must be running: hit Start in the per-agent header (see The workspace) and wait for the status pill to read running.
  • Deployed: the production deployment on Vercel. Use it to smoke-test what your users actually reach. It needs a deployment to exist; see Deploy to Vercel.

The Deployed target bar

Switching to Deployed opens a bar under the toggle:

  • A URL field (https://<project>.vercel.app). Detect fills it from the linked project's latest production deployment.
  • A secret field, "paste Protection Bypass secret from Vercel". If the project has Deployment Protection on, generate the secret in Vercel → Project → Settings → Deployment Protection → Protection Bypass for Automation and paste it here. It's a system variable and is not included in env pull. The OIDC token the deployed agent's route auth expects is read from .env.local automatically (pull env in the Environment tab).
  • Save stores both. Test saves and then probes the deployment's authenticated info route. Success reads Reachable ✓. Deployed chat is ready. Failures name the cause: blocked by Deployment Protection (add the secret), route auth rejected the request (pull env to refresh the OIDC token), or the URL couldn't be reached.

A thread is pinned to the server that minted its session. If you switch a thread between Local and Deployed and send again, Studio starts a fresh session on the new target and drops a notice into the transcript: Chat target changed to <url>. Starting a new session there; the earlier session stays on its own server. The same happens if the local dev server comes back on a different port.

What streams

As the agent runs a turn, Chat renders each event live, in order:

  • Your message appears the moment you send it. On a cold production deployment the agent can take a while to accept the run, so the reply may lag; the message itself never does.
  • Turns: the assistant's messages as they're produced.
  • Tool calls: each call is a collapsible row with the tool name and a status badge; expand it to see the tool's input and output. Partial output streams in while a long tool runs.
  • Subagent delegations: when the agent hands work to a specialist child agent, a delegated to <name> row appears with the child session id, so you can follow the delegation.
  • Reasoning: a collapsed thinking… block showing the model's reasoning before it acts.
  • Approval prompts: when a tool is gated for human-in-the-loop, the turn pauses and asks you to Approve or Deny (or pick from the options the request defines) before it proceeds.
  • Sign-in prompts: when a Vercel Connect connection needs authorization, a Sign in to <service> card offers Open sign-in.
  • Compaction: Compacting context… and Context compacted markers show where Eve summarized older history.
  • Cancellations and failures: a cancelled turn shows Turn cancelled (not an error). A failed step, turn, or session shows Eve's error code and message; when the three fail in a row with the same message, Chat renders them as one row titled by the most severe stage rather than three.

Session controls

Next to the composer:

  • Cancel turn: shown while a turn is streaming, as a pill above the composer and in place of the send button. It posts to the session's cancel route; Eve confirms on the stream with turn.cancelled followed by session.waiting, and the session accepts the next message afterward. Whatever the turn already streamed stays in the thread.
  • The ⋯ menu (Session controls) offers three actions, none of which asks for confirmation:
    • Compact context: summarize history in place. Useful before a long task in a long thread.
    • Clear context: drop the model history but keep the session (its id, tools, and durable state).
    • Reset session: retire the session. Studio archives the current thread and opens a fresh one.

The composer is disabled while a turn is streaming; use Cancel turn if you want to interrupt and send something else.

The composer and usage bar

A composer at the bottom sends your message (Enter sends, Shift+Enter adds a line). Directly under it, a usage bar shows the model configured for the agent, a context-window fill bar, and the running cost and token counts for the thread, formatted like $0.0123 · 12,345↑ 678↓.

How sessions work

Each thread is a durable Eve session. When you send the first message, Eve returns as soon as its Workflow accepts the run and starts streaming. Studio waits for the session to reach session.waiting before it sends a follow-up; if a follow-up lands too early Eve answers 409 session_not_active, and Studio retries with backoff rather than starting a new session. If the retries are exhausted, Chat explains that the session is no longer active and points you to ⋯ → Reset session.

Eve exposes no history API, so Studio keeps the transcript itself, one file per thread. Reopen a thread and Studio replays the stored events, then resumes the live stream, deduplicating by each event's durable id.

Threads and archiving

Each conversation is a thread. Threads live under the active agent in the left rail; start a new one with New chat. Hover a thread to Archive it, and reopen it later from the Archived sessions modal. See The workspace for the rail controls.

If the agent won't reply

If you see a Not connected to Vercel (or Sign in to Vercel) banner instead of a response, the agent has no model credential yet. The model routes through the Vercel AI Gateway, so it needs one before it can run. Follow Connect to Vercel, then send your message again.

Next