Talk to a coding agent inside Pekan, with your tasks as context
Pekan now has an Agent panel: a real conversation with a coding agent, running on your machine, docked on the right of the project workspace. Ask a question, watch the answer arrive word by word, see the plan it is following, approve or refuse the tools it wants to use, and read the changes it proposes as diffs you can open in the Files panel beside it. The panel sits next to the existing Files and Diff panel rather than replacing it, so you can keep a file open while the conversation continues. Both panels open, close and resize on their own, and each remembers its own width between sessions.
The agent itself is not something Pekan hosts. It connects to a locally installed adapter over the Agent Client Protocol, an open standard for editors and coding agents to talk to each other. Adapters for Codex and for Claude are supported to start with, and because the protocol is the interface rather than any one vendor, anything else that speaks it can be added later without new plumbing. Pekan detects which adapters are installed and shows the exact install command for the ones that are not, and it will never download or run one for you.
The part that makes this different from a terminal is that Pekan knows what your tasks are. Type an at sign in the message box and a task picker opens, showing each match with its ID, title, project, column and whether it is blocked, so two tasks with the same title stay distinguishable. Pick one and it is attached to that message: the agent receives the task ID, its summary and description, the acceptance criteria written into it, the latest handoff, and a bounded slice of recent comments, all under a stable reference it can cite back to you. If a task you mentioned cannot be found, Pekan refuses to send rather than quietly leaving it out, because an agent answering confidently without the context you asked for is worse than an error message. If a very large task has to be trimmed to fit, you are told what was left out before it goes.
Every session starts with an explicit decision about what it may change. Read only lets the agent read and explain but not edit files or run commands, and it is the default: a conversation that turns out to need more can be restarted, one that quietly had more cannot be un-run. Choosing to let it edit the project grants file changes and commands inside that project directory and nothing wider. That choice is then enforced rather than merely recorded, so a read-only session shows an agent request to edit something, explains why it cannot be allowed, and tells the agent no.
Stopping works properly too. Cancelling a turn ends it, resolves any pending approval as cancelled, and leaves the session ready for the next message. If the adapter crashes, disappears, or speaks a version Pekan does not understand, you get a plain explanation rather than a hang, and quitting Pekan never leaves an agent process running behind it. Existing agent actions, terminal launches and isolated worktree delivery are untouched and behave exactly as before.
Highlights
- A conversational Agent panel docked beside Files and Diff, with both usable at once.
- Codex and Claude connect through the open Agent Client Protocol; Pekan detects adapters but never installs them.
- Mention a task with @ to send its full context, including acceptance criteria and recent comments.
- Every session picks a write scope up front, defaulting to read only, and it is enforced on every request.