Every session in the Agent Log has a chat box, and the conversation outlives the run that started it. The chat is the informal half of the workflow — for the things that are too small for a refine and too specific for a task.

When to chat instead of clicking

• You want to understand before deciding. Why a native query rather than a derived one? Which tests cover the ranking? The agent answers from the conversation it already has — it knows the change, the task, the files it touched.
• The fix is one sentence. The index migration should be idempotent. Asking in chat beats commenting, refining and reviewing a second run when the change is small.
• You want the next task with a twist. Run task 3, but use the existing SearchQuery DTO instead of a new one. Chat can launch the task and carry the instruction.

What a chat turn can do

Reply

A plain answer. Nothing changes on disk; the review stays as it was.

Edit code — and keep one review

When your message asks for a change, the agent registers a refine with SpecBuddy before touching a file — a snapshot is taken, your pending comments are attached — and then makes the edits in the same turn. The turn ends in the ordinary review: the Cockpit shows the diff, and rollback undoes it. Nothing is duplicated; you don’t get a second session to reconcile with the first.

Edit the documents

Ask for changes to the proposal, the deltas or tasks.md, and the turn is handled the same way — as a refine, with a snapshot before and a review after. The task count in the Explorer follows if sections were added or removed.

Run a task

Run the next step. Do task 3. Run “Search API” again. SpecBuddy recognises a run request and launches it through the same pipeline a Cockpit click uses — snapshot, worktree, review at the end. Naming a task that is already done re-runs it. If the plan has no pending task, or the name matches nothing, the log says so instead of guessing.

The agent does not do this itself; it hands the request to SpecBuddy, which validates it and launches. So the guarantees hold — a chat-launched task is reviewed like any other.

Comments come along

Inline comments you have not sent yet are attached to your next message, with their file, lines and quoted text, so “fix the two things I marked” works. The pending list is shown by the box; click a comment to exclude it from this message.

Timing

• Idle session — your message starts a turn immediately.
• Turn in progress — Message queued — delivered when the current turn ends. Queued messages are sent together as the next turn, in order. They are dropped, not delivered later, if the session ends before then.
• No live session — the box says so; the next run starts a fresh conversation. Reopen a persisted log and you can read it, not continue it.

Where chat stops

• The workflow tools act only from your chat turns. A task run launched from the Cockpit cannot be turned into something else by the agent mid-run; it does the section it was given. Launching, refining and running are decisions the agent may request on your behalf in chat, never take during a task.
• Chat is not a substitute for the review. Every edit the agent makes in chat is still diffed, still snapshotted, still rolls back. If you notice you’ve been steering a task by chat for ten turns, the plan was wrong — refine tasks.md and run it clean.
• Switching agents ends the conversation. Change the profile in the composer and the next turn starts a new session that has not read the log above it. The log shows a seam where that happened.