Discuss an issue

An issue conversation lets an agent investigate the current code and post an answer on the issue without turning the question into an implementation assignment. Use it for questions such as where behavior comes from, whether a reported problem can be reproduced, or how a possible change might fit.

You need a configured conversation route and a running daemon. The dreamcatcher.toml that dreamcatcher init writes routes the agent-conversation label to the dream:scout skill from dream. That skill checks each claim against the code, and cites the code it relies on. Configure labels and harnesses shows how to write a route of your own.

Ask the first question

On GitHub:

1. Keep the issue open.
2. Assign it to the account authenticated through gh.
3. Add exactly one configured conversation label, such as agent-conversation.
4. Post the question as an issue comment from that same account.

The title and issue body provide context, but they do not start a conversation on their own. A comment is the request to answer. Comments from other accounts are not given to the agent as questions.

At its next update, dreamcatcher asks the configured harness to examine current origin/main and publishes the successful final answer as a marked issue comment. Its progress and tool activity remain in the local feed rather than becoming partial GitHub replies.

Monitor it on the web home page or in the terminal:

dreamcatcher view GH123
dreamcatcher feed GH123

Continue the discussion

Post another issue comment from the authenticated account. dreamcatcher resumes the same harness session with the new comments, so a follow-up can refer to the earlier answer while examining the latest origin/main.

The agent may read source, run reproductions and, when your question asks for it, make suitable issue changes such as filing a follow-up issue. An issue conversation investigates; it does not implement project changes or create and update a pull request. To request implementation, prepare an assignment instead.

Pause, resume or end a conversation

Each conversation belongs to the label that started it. dreamcatcher checks its issue at every update, including while a round is running. Removing that label, or removing the authenticated account as assignee, pauses the conversation, and a second configured conversation label holds it back in the same way: no new round starts, including one that would retry an errored round. A round already running may finish and publish its answer. The home page shows the conversation as held, says why beside the issue, and says what brings it back, such as "put the label back to resume".

Undo the change to resume the conversation. Only a label that you put back yourself resumes it. It continues in the same harness session, and comments posted while it was paused are given to it. It keeps its original recipe even if its route is removed from dreamcatcher.toml.

Closing the issue ends the conversation after any running round finishes, and the web home page lists it with your other ended work. dreamcatcher then removes its worktree and keeps its record. If the issue is reopened with its label, a new conversation starts and is given every comment you have posted on the issue.

Taking one conversation label off and putting another on pauses the first conversation and starts a new one under the new label, with a session of its own. The new conversation is given every comment you have posted on the issue. Each conversation keeps its own agent work identifier. With more than one piece of agent work at the issue, dreamcatcher view GH123 prints a command for each piece, with the identifier that selects it, and giving one of those identifiers in place of GH123 shows that one.

To stop just the round that is running, follow Stop and recover work.