Agent Chat Overview

Conversational AI with real project access — read files, run commands, create PRs, spawn worker runs, and capture proof artifacts, all from a single chat interface.

What is Agent Chat?

Agent Chat is ADE’s conversational interface for AI agents wired into your development environment. The agent reads and writes real files, runs terminal commands, creates branches and PRs, queries your context pack, and can escalate to a Worker run when the task outgrows a single session. Every chat session is recorded with a full transcript, tool call log, and timing data.

Supported providers

Agent Chat works across multiple provider backends:

Provider	Backend	Auth
Claude	Anthropic SDK v2 sessions	CLI subscription or API key
Codex	OpenAI app-server JSON-RPC	CLI subscription or API key
Cursor	Cursor SDK	Cursor API key
OpenCode	OpenCode server (local process)	API key, OpenRouter, or local endpoint

ADE detects which providers are available and lets you switch between them per-session using the model selector in the chat header. Each provider has its own session lifecycle, timeout behavior, and interrupt semantics.

What Chat is for

Exploratory work, targeted edits, quick Q&A about your codebase, PR creation, test runs, and anything that resolves in a single focused session.

When to escalate to a Worker run

Tasks spanning multiple files or agents, work requiring a formal audit trail, or anything you expect to run for more than 15–20 minutes.

---

Chat Scoping

Chat in ADE is always scoped — it operates within a defined context boundary that determines what the agent can access and what its tool calls affect.

Lane Chat

Lane chat is the most common mode. It is opened from within a specific Lane view and anchors the agent to that lane’s context:

• File reads and writes are scoped to the lane’s git worktree
• Terminal commands run in the lane’s terminal session
• Git operations affect the lane’s branch
• Context is seeded from the lane’s pack (recent activity, file map, open PRs)

Lane chat is how you do day-to-day development work with an agent.

CTO Chat

The CTO chat operates at project scope. The CTO agent has access to:

• Project-wide context (all lanes, all branches, all pack snapshots)
• Linear sync — it can read and write issues
• Delegation — it can spawn worker agents or new worker runs
• Architecture decisions recorded in the project pack

Use CTO chat for cross-cutting questions, planning, and delegation rather than hands-on coding.

Worker run Chat

Worker run chat is embedded inside a Worker run’s detail view. It is a direct channel to the worker run’s orchestrator while the worker run is running. Use it to:

• Ask the orchestrator for a status summary
• Provide additional context mid-worker run
• Redirect a worker that has gone off course

Worker run chat does not start new work — it communicates with the orchestrator of an already-running worker run.

---

Starting a Chat

1. Open a Lane

   Navigate to the lane you want to work in from the left sidebar or the Workspace Graph.

2. Open the Chat pane

   Click the Chat button in the lane toolbar, or press Cmd+K. A new chat session opens with automatic context injection from the lane’s pack.

3. Choose a model (optional)

   The model selector in the chat header defaults to your configured preferred model. Click it to switch models for this session. ADE remembers your last-used model per lane.

4. Pick a permission preset (optional)

   For CLI-wrapped models, the composer surfaces a permission picker in the header. The presets vary per provider:

   Claude — standard Claude CLI permission modes (plan / edit / full-auto / default), each with its own allowlist.

   Codex — four presets map directly to Codex’s --approval-policy and --sandbox flags:

   • Default permissions — workspace-write sandbox, approval policy on-request. Edits and shell commands in the workspace run without asking; anything outside the workspace prompts for approval.
   • Plan mode — read-only sandbox, approval policy on-request. Agent can explore and answer questions; it needs approval to edit or run commands.
   • Full access — danger-full-access sandbox, approval policy never. No sandbox, no approvals — safe only in an externally sandboxed environment.
   • Custom (config.toml) — no ADE flags are passed. Codex honors ~/.codex/config.toml.

   There is no “Guarded edit” preset anymore; Default permissions is the standard middle ground.

5. Type your first message

   The agent receives your message along with the injected context. Tool calls begin immediately if the task requires them.

   You can also attach images (JPEG, PNG, GIF, or WebP) to your message using the attachment button in the composer. Images are sent inline as base64 content blocks, so the agent sees them natively — no file-read workaround required.

Start a new session from an existing session’s history by opening the History tab and clicking Resume. The previous transcript is loaded as context so the agent picks up where you left off.

---

Proof Capture

Chat agents can leave intentional evidence of their work by calling the ade proof CLI — screenshots, short recordings, or ingested browser traces. There is no per-session policy chip anymore; the agent decides when a step deserves a proof, runs the CLI, and the captured artifact is linked to the chat session automatically.

Captured proofs appear in the session artifact panel and can be promoted to the lane’s durable context, attached to a worker run, or published to a PR or Linear issue. See Proof Overview for the full workflow.

---

Session behavior

Auto-titling

ADE automatically generates descriptive session titles using AI. The primary configuration namespace is sessionIntelligence.titles:

• sessionIntelligence.titles.autoTitleEnabled — toggle auto-titling on or off
• sessionIntelligence.titles.reasoningEffort — depth of analysis for title generation
• sessionIntelligence.titles.refreshOnComplete — regenerate the title when the session ends
• sessionIntelligence.titles.modelId — override the model used for title generation

Legacy keys ai.chat.autoTitleEnabled and chat.autoTitleReasoningEffort / chat.autoTitleRefreshOnComplete still function as runtime fallbacks.

Interrupting agents

Press Escape or click the Stop button to interrupt a running agent. Behavior varies by provider:

• Claude — graceful interrupt; agent reports partial results
• Codex — issues a turn/interrupt for the active turn; partial output is preserved in the transcript
• Cursor — cancels the active SDK run

Steering and editing

You can steer an in-progress agent by sending a follow-up message before the current turn completes. ADE queues the steer message and delivers it at the next turn boundary. You can also edit a previous message to re-run the conversation from that point.

Context integration

Chat sessions receive lane, session, artifact, worker run, and explicit user context. For product or architecture knowledge, agents inspect the repository directly through ADE tools.

---

Explore more

Chat Capabilities

What the agent can do — file operations, git, PRs, worker runs, lane and action tools via the ade CLI, and more.

Context & Artifacts

Managing context windows, working with artifacts, chat history, and troubleshooting.