Chat Mode

Overview

When a .prompty file declares an input with kind: thread, pressing F5 opens an interactive chat webview instead of running a single execution. This gives you a full multi-turn conversation interface directly inside VS Code.

Frontmatter with thread input

inputSchema:
  properties:
    - name: question
      kind: string
      default: Hello
    - name: conversation
      kind: thread

The thread input tells the extension that this prompt is designed for conversation — the extension will manage the conversation history and pass it back into the prompt on every turn.

---

Chat Interface

The chat webview provides a dedicated set of UI elements for interactive conversations:

• Message list — displays the full conversation history with role-colored bubbles. System, user, and assistant messages are visually distinct so you can scan the conversation structure at a glance.

• Text input — type your message at the bottom of the panel and click Send or press Enter to submit.

• Loading indicator — a visual spinner appears while the extension waits for the LLM to respond, so you know the request is in progress.

• Tool call cards — when the model invokes tools, each call is shown as an expandable card displaying the function name, arguments, and the inline return value. You can expand or collapse individual cards to inspect details.

• End chat button — stops the current session and saves the complete multi-turn trace as a .tracy file. Click this when you’re done with the conversation.

---

How It Works

Each user turn re-runs the full prompt pipeline with the accumulated conversation history passed into the thread input. Here’s the step-by-step flow:

1. User types a message in the text input and presses Enter
2. Extension builds the full conversation history — all previous user and assistant messages, including any tool call results
3. Passes the history as the thread input alongside the user’s current message
4. Runs the full pipeline — render → parse → execute → process — just like a single-shot execution, but with the conversation context included
5. Appends the response to the conversation in the message list
6. If the model returns tool calls, each call is displayed as an expandable card and the tool function is executed automatically
7. Re-sends with tool results if needed — the extension loops until the model returns a final text response

This means every turn sees the entire conversation, giving the model full context for coherent multi-turn dialogue.

---

Tool Calling in Chat

When your .prompty file defines tools in the frontmatter, the chat panel handles tool calls automatically during the conversation. If the model decides to invoke a tool:

1. The tool call appears as an expandable card in the message list
2. The card shows the function name and arguments the model provided
3. The extension executes the tool and displays the return value inline
4. The result is appended to the conversation and sent back to the model
5. The model can then respond with a final answer or make additional tool calls

This loop continues until the model produces a standard text response. You can expand any tool call card to inspect the arguments and result at any time during or after the conversation.

Tip

Tool calling in chat mode works the same way as in single-shot execution — the only difference is that the conversation history grows with each turn. Define your tools in the frontmatter tools: section as you normally would.

---

Saving Traces

The final .tracy file captures the entire multi-turn session — every user message, assistant response, tool call, and tool result across all turns. Click End Chat to stop the session and save the trace.

The trace is written to the .runs/ directory in your workspace, just like single-shot execution traces. Open it to inspect timing, token usage, and the full conversation flow.

Tip

See Tracing for details on navigating the trace viewer. The Conversation tab is especially useful for chat traces — it reconstructs the full multi-turn dialogue with role labels and tool call details.

---

Note

Chat mode is activated by declaring a kind: thread input in your prompt — it is not controlled by apiType. The apiType field determines which LLM API to call (chat, embedding, image), while chat mode is a UI feature of the extension. A prompt can use apiType: chat (or omit it entirely) and still open the chat interface as long as it has a thread input.