Browse documentation

Repopack Chat

Connect a local agent harness, configure provider, and work in Steering and task sessions.

Repopack Chat connects a local agent harness to your web session through rp. It requires project.read and task.write on an active project.

Supported harnesses: Codex and OpenCode. Request others in the Repopack project.

Start here

  1. Set up your runner. On Linux (or a Linux VM), install rp, Codex or OpenCode V2, Bubblewrap, Git, and OpenSSH.
  2. Sign in. Run rp login on the runner (rp login --headless over SSH). Run rp agent configure, choose a harness, and follow its prompts. Provider credentials stay on the runner.
  3. Pair. In your browser, open a project and choose Agent workspace (Show chat on mobile). In Connect agent, choose Pair agent harness. Copy and run rp agent start on the runner, then enter its code in the same dialog.
  4. Send a message. Choose New Steering chat, pick a model, and press Enter to send. Use Shift+Enter for a new line.

Steering and task chats

  • Steering: Choose a named Steering thread to return to its history and draft, or New Steering chat to start fresh. Say “Take #42” to take that task and move the conversation there. If the take fails, the conversation stays in Steering.
  • Navigate from chat: Ask your agent to open a task or the next stack step. Choose an Open button to switch chats without taking the task or starting work.
  • Task chat: Choose Open in Chat on a task to reopen its last-viewed open tab. If there's no open session, you get a draft. The first message starts the session; opening the draft doesn't take the task. Clicking an offered local task mention opens its chat without sending a message. Other task-page links navigate to the task page.
  • Verify against another project: Ask the agent to prepare source from a named Repopack project at a full commit SHA. The runner provides a temporary archive for commits reachable from a branch or tag you can read. Download it once within five minutes; if it expires, request it again. No write access is granted and your checkout stays unchanged. This needs an updated runner.
  • Do a task: Asking Chat to do, finish, or complete a task means implement, verify, and submit it for review. Asking to revise means address the review and resubmit. Say “keep it local” to stop before pushing or submitting. Review submission does not merge or deploy the work.
  • Session tabs: A task has one row in the sidebar. Use + New session to start a separate conversation on the same task. Each tab has its own history and draft. A task spinner means a session is working. New sessions get a title from the first message, without a naming prompt. Unnamed tabs show just their short ID. Existing titles are kept, except for the old exact labels Build, Plan, Review, and Audit, which switch to short IDs when the new tabs roll out. Titles don’t change permissions or review assignments. To rename a saved conversation, open the Chat command picker with / and choose /rename. This doesn’t send a message to the agent.

Sending and recovery

  • Waiting for a reply? The blue dots beneath the composer show when the agent is working. Questions and permission requests still appear separately.

  • Command output: The $ card shows the command and a short output preview. Click it to expand the recorded invocation and output; click again to collapse.

  • Tool activity: While a command runs, a spinner replaces its $ prompt. Patches can also show a running spinner. Command output may update as it arrives. If a turn ends without a tool result, Outcome unconfirmed means Chat did not see that tool finish.

  • Codex replies: Reply text can appear as Codex writes it. Native history remains the record of the turn if the live preview is interrupted.

  • Offline? Reconnect the original runner with the same profile and workspace. Unsent drafts aren't sent automatically.

  • Not sure a message went through? Check Session details → View native history before sending it again. A timeout beside the message distinguishes a request the runner never picked up from one it claimed but did not acknowledge in time. The composer keeps your draft. You can edit it or send a message in another chat. When you return, the original text and its delivery warning stay together in the conversation. The task can reach Ready independently of that message receipt.

  • Needs attention? Chat stopped retrying an uncertain Codex turn. Inspect native history and local changes before continuing in a new session.

  • Need command output? Expand the row marked $. A spinner means the command is still running.

  • Looking for an old chat? Open Archived. Stopped sessions are read-only, and their history is available when the original runner is connected.

  • Delete a session: Select the saved tab, then use … → Delete session on the task row. Check the session in the confirmation: the original runner permanently erases it. Removing the local worktree is optional and discards uncommitted files and local-only commits. If the session moved to another task while the dialog was open, select it again before confirming.

Runner upgrade notices

If a paired runner’s discovery poll reports an unsupported rp version, its owner sees the runner name, reported version and required version in the web UI. Run rp update on that host and restart the runner. A supported discovery poll clears the notice. The notice describes the version check, not task progress.

Delegate from Steering

Ask “What commit is master at now?” to verify a release or merge. Chat can use resolve_branch_head to read the current full SHA from Repopack, including for another project where you have code-read access. This does not need SSH credentials in the Steering checkout. For source verification, use that SHA with prepare_project_source.

Create as many Steering threads as you need. Planning or creating a task leaves it available for later. Ask Steering to implement, review or audit a named task: the agent can use the delegate_task tool to start a background task chat and report a receipt. You can keep discussing the next task in Steering. Task links navigate; Implement task and Review this submission explicitly start work. Use the conversation menu to archive old Steering threads. Deletion in Archived removes the selected session; task conversations have their own histories.

Choose how task chats work

Use + New chat for another conversation on the task. Multiple conversations appear as named tabs; use /rename to change the selected conversation's name. The task title below the composer keeps its identity visible.

  • On submission: Wait, Review here in the current conversation, or Separate reviewer in a dedicated reviewer conversation.
  • On requested changes: Wait or Revise automatically. The task owner's runner can revise and resubmit when feedback arrives.

For one continuous conversation, select Review here and Revise automatically. For a separate reviewer, select Separate reviewer instead. Each submission receives at most one workflow review instruction. The reviewer checks the exact head and records a verdict. Repopack applies the project's ordinary review and readiness rules. View review in Steering or the task opens the conversation handling that submission. Update your runner for review: it reports its current checkout head on each poll. The review instruction waits until that head matches the submission.

These switches belong to your chat. Another person's revision preference does not block yours; only the current owner's runner can receive a revision turn. For reviews, the first request or automatic claim reserves that submission. Other reviewers see that a review was already requested; their runners do not receive a duplicate instruction.

Opening a task directly shows whether another reviewer has requested the current submission or you have already reviewed it. You can still read existing chats.

Set defaults for new conversations in Account → Agent, including a runner model (provider/model) and variant, and optional review-fork overrides. Models must be available on the connected runner. Chat uses Auto tool permissions; individual native permission questions still appear in the transcript.

If a runner's acknowledgement is lost, inspect its native history. Resume later submissions without resending releases future turns while preserving the original receipt. A declined instruction is not replayed for the same submission.

Separate reviewer creates a reviewer conversation on the first submission and returns to it on resubmission with the new exact head. Wait pauses future reviews and keeps that conversation available for the next round.