Skip to main content

​
What the Agent Can Do

The chat agent has access to a broad tool palette. The exact tools available depend on the lane’s permission config and the session’s settings.

File operations

Read, write, and search files within the lane’s worktree. The agent sees your full directory structure and can reference any file by path.

Terminal commands

Run shell commands in the lane’s terminal session — build scripts, test runners, linters, package managers, or any CLI tool installed in your environment.

Git operations

Create and switch branches, stage and commit changes, view diffs, check status, and push to remote. The agent stays within the lane’s worktree.

Pull requests

Create PRs with auto-generated descriptions, update PR bodies, request reviewers, and apply labels — if GitHub integration is configured.

Linear integration

Create and update Linear issues, link issues to branches, and attach artifacts to issues.

Mission spawning

Escalate the current task into a new Mission, carrying over conversation context as the planner’s seed.

Context pack queries

Query the lane’s pack for architectural context, file maps, open PRs, recent activity, and human work digests.

Image attachments

Attach images (JPEG, PNG, GIF, WebP) to any message. Images are sent as inline base64 content blocks so the agent can see screenshots, diagrams, or UI mockups natively.

External MCP tools

If external MCP servers are configured and permitted for the session, the agent can call their tools. Toggle MCP tool groups in the chat header.
Provider-native tools (built into the model’s API, like Claude’s native file tool) are tracked separately from ADE-owned tools (served via MCP). The chat header MCP badge shows the count of ADE/external MCP tools; provider-native tools are always available based on the model.

​
Creating a PR from Chat

**Screenshot: PR creation flow in chat — show the agent’s message proposing a PR, the pre-filled PR form with auto-generated title and description, and the “Create PR” confirm button.
The most common workflow is: make changes in chat → have the agent create a PR.
1

Make your changes

Work with the agent to implement the feature, fix the bug, or refactor the code. Changes are committed to the lane’s branch as you go.
2

Ask the agent to create a PR

Say “create a PR” or “open a pull request.” The agent will:
  • Read the current branch name and recent commits
  • Query the lane pack for architectural context
  • Generate a PR title and description based on the changes
  • Pre-fill reviewer suggestions if you have GitHub team integration enabled
  • Apply relevant labels based on the branch name and changed files
3

Review and confirm

The agent surfaces the PR form as an interactive block in the chat. Edit the title, description, or reviewer list, then click Create PR. The PR is opened in GitHub and a link is attached as an artifact in the chat session.
If your project has a PR description template, put it in the lane’s context pack. The agent will use it automatically when generating PR descriptions.

​
Spawning a Mission from Chat

When a task is too large for a single chat session, escalate it to a Mission without losing context.
1

Recognize the scope

Signs that a task should become a Mission: it requires changes across more than 3–4 unrelated files, it needs parallel workers, it has an explicit validation phase, or you want a formal audit trail.
2

Say 'start a mission'

Type “start a mission for this” or “this should be a mission.” ADE opens the Mission creation dialog, pre-filled with:
  • A title derived from the current conversation
  • A description assembled from your messages in the session
  • The current lane pre-selected
3

Review and launch

Adjust the mission config as needed (add lanes, set budget cap, configure permissions), then click Start Mission. The mission opens in the Mission detail view. The chat session remains in History for reference.

​
Long-Running Tool Calls

Tool calls that take more than a few seconds are displayed as expandable blocks in the chat message list:
  • A spinner and elapsed timer indicate the call is in progress
  • Click the block to expand it and see partial output as it streams
  • Click the block again to collapse it once it completes
If a tool call is taking unexpectedly long (e.g., a test run has hung), press Escape to interrupt it. The agent receives an interruption signal and will report the partial results before asking for guidance.
If a tool call fails, the agent automatically explains the error and offers alternatives — you do not need to re-phrase your request. You can either accept the agent’s proposed workaround or provide explicit instructions.