The Conversation panel is a normal chat, plus the editor-specific pieces: workspace mentions, image attachments, code-block actions, and a live view of what the agent is doing.

Sending a prompt

Type in the composer and press Enter. Shift+Enter adds a newline. The Send button is disabled until there is text or at least one image.

While a response streams you can press the stop button to abort it. Tool calls appear as they run: file reads, searches, edits, shell commands, and MCP tools each get a collapsible block with their arguments and results. Edit calls include a diff. The pi-agent.autoExpandSmallEdits and pi-agent.autoExpandEditLineThreshold keys are declared in the manifest but are not read by the 3.0.0 runtime — see Settings.

Mentioning workspace files

Type @ and start typing a path. PiLot Studio searches the workspace and shows matching files. Use the arrow keys and Tab or Enter to insert the path.

An @mention is more than a label. When the prompt is sent, PiLot Studio reads the file and prepends its contents to the prompt, then removes the mention token from the visible instruction. This is how you give the agent a precise file without relying on it to search.

Limits in 3.0.0:

Files larger than 10 MB are refused.
A single file is inlined up to 50 KB; larger files are truncated with a notice.
Up to 64 mentions are processed per prompt.
Binary formats (images, archives, PDFs, Office documents, compiled artifacts, fonts, and similar) are not inlined.
Mentions are resolved inside the workspace; paths cannot escape it.

The Ctrl+Shift+A shortcut and the paperclip button open a file picker and insert the selected workspace paths as mentions.

Images

Drop an image onto the composer or paste it from the clipboard. PiLot Studio accepts PNG, JPG, GIF, and WebP, up to six images per prompt and 10 MB per image. Attached images show as thumbnails with a remove button.

Images are sent to the selected model as image content. The model must support image input; otherwise the provider will reject the request.

Auto Context

Auto Context is a per-session toggle in Preferences → Agent Behavior. When it is on and Light Mode is off, PiLot Studio prepends a short project preamble to each prompt:

Project Root: /path/to/workspace
Project Name: your-package
Project Version: 1.2.3

The preamble is stripped from automatic session titles. If the workspace has no readable package.json, the root is still included and the name/version fall back to Unknown.

Auto Context is different from @ mentions: it is one small project header, not the contents of your files. It is forced off in Light Mode because every token competes with a local model's weights and KV cache.

Slash commands

Type / at the start of the composer to open the command list. Built-in commands are handled by PiLot Studio; skills and prompt templates from your session resources are appended to the same list.

Command	Action
`/settings`	Open the Preferences tab.
`/model`	Open the Models tab.
`/providers`	Open the Providers tab.
`/tools`	Open the Capabilities tab.
`/packages`	Open the Packages tab.
`/skills`	Open the Skills tab.
`/new`	Start a new session.
`/tree`	Open the History tab.
`/name <name>`	Set the session display name.
`/compact [instructions]`	Compact the session context now.
`/export [.html\|.jsonl\|path]`	Export the session; the default is `.html`.
`/reload`	Reload PI extensions, skills, and prompt templates.
`/skill:<name>`	Run a skill exposed by a package or skill directory.
`/<template>`	Expand a prompt template from a package or project.

The autocomplete also matches descriptions, so typing export finds /export.

Markdown is a separate export format in the dialog; the /export command supports .html and .jsonl.

Code blocks

Code blocks in assistant replies have four actions:

Apply replaces the current selection, or inserts at the cursor when there is no selection. It requires an active editor.
Diff opens a VS Code diff between the current file (with the selection replaced, or the code inserted at the cursor) and the proposed change.
Open in Editor opens the code in a new untitled document with the correct language.
Copy copies the block.

Apply operates on the active editor. If the editor is read-only or the file changed underneath you, VS Code rejects the edit and PiLot Studio reports that instead of claiming success.

Message actions

Hover a message to see its actions.

Copy message copies the full text.
Fork here appears on user messages. It starts a new session branched from that point, preserving the original conversation. This is the supported way to explore a different follow-up without losing the current branch.
Timestamp toggles between relative and absolute time.

The PiLot Studio: Edit Last Message command opens an edit overlay for the most recent user message. Saving it replaces that message in the session and persists the change. It does not delete later messages or automatically re-run the prompt.

Searching the conversation

Press Ctrl+F (or use the search icon) to open the in-chat search bar. It searches user text, assistant text, thinking text, and tool output.

Enter or F3 moves to the next match.
Shift+F3 moves to the previous match.
The counter shows the current match and total.
Escape closes the bar.

The Ctrl+F keybinding is scoped to the PiLot chat view or an editor chat panel so it does not steal VS Code's normal find shortcut elsewhere.

Prompt templates

The Prompt Templates button opens a library of built-in starting points grouped into analysis, implementation, quality, and workflow categories. You can search by name, description, tag, or prompt text; recently used templates appear under Recent. Choosing one inserts its text into the composer for editing.

Separately, PI prompt templates discovered from packages or project files appear in the slash-command list and are expanded by PI when sent.

Mermaid diagrams

Fenced mermaid blocks render as diagrams in the transcript. Invalid diagrams fall back to an error state rather than breaking the rest of the message. Mermaid rendering is client-side; the diagram source remains part of the message.

Tokens and context

The token strip above the composer reports:

In — tokens sent to the model this session.
Out — tokens generated by the model.
Cache — cache-read tokens reported by the provider.
Total — the session total.

The context indicator shows the current context usage against the model's window. Click it to compact manually, or use /compact. Automatic compaction is a PI session setting; the indicator shows whether it is on. If the session is resumed with a fallback model, PI's fallback notice is surfaced in chat instead of silently switching.

What the agent can do

The default tool set is read, bash, edit, write, grep, find, and ls. Use the Capabilities tab to switch to a read-only review preset, disable tools, or build a custom allowlist. MCP tools are managed separately in the MCP Servers tab.

If a tool is disabled, the agent can still reason about code you paste or mention, but it cannot read or change files with that tool.