Agent chat
Open the Agent chat with ⌘1. It gives you agent chats per project, instead of separate Claude Code and Codex windows. Behind each chat is a headless Claude Code session acting as the orchestrator. It answers questions and makes small changes itself. For anything bigger it plans the work, checks with you, and hands it to specialists, whether Claude subagents or Codex jobs. To give the agent a team of roles, see The team.
Replies, jobs and permissions
Replies are rendered as Markdown. Tool calls appear as one-line entries you can expand. Each specialist job appears as a card with live status; open a card for the job's brief and result (rendered as Markdown), its conversation (Claude) or activity (Codex, needs kit 0.6.1), and its log.
Permission prompts, questions and the approval gates appear as buttons in the chat. The selector under the message box sets what the agent may do without asking: Ask before changes, Accept edits, Auto or Plan only.
The mode a new chat starts in, and the model, effort and engine, can be set for all projects in Settings → Agents or for one project in Project settings. To restrict Slurm submissions, git publishing, HPC changes or writes outside the project, set the rules in Settings → Agent policy or per project in Project settings.
Job progress
When a specialist reports its steps, the job card gets a collapsible progress block: an overall bar with stage N/M, a row per stage, a done/total bar for countable work (such as files), the time elapsed against the usual duration for that stage, and an estimated time left once Codeome has seen the stage before. A stage that is running but has not reported for 120 seconds gets a stall badge. Cards without progress look as before.
The agents report the steps; Codeome does not guess them from logs. Progress is read from the job's progress.json rather than saved in the chat history, and shows for running, paused and held jobs and for jobs that ended within the last hour. It needs agent kit 0.7.0, so update the kit in projects that have it installed (see Agent kit). Background commands get a card only through the log watcher below.
Test and build progress
For a long test run or build, the agent can start codeome-agent watch in the background. It tails the log and parses Vitest, pytest, tsc -b, Snakemake or a custom pattern. The chat then shows a card of its own, but only for a Bash or Monitor call that contains codeome-agent watch ... --task <key>; other background commands get no card. The card has the same block as a job card: a bar, the time left estimated from past runs, and a stall badge after 120 seconds without an update. The progress file decides the outcome, so a run with failed tests ends the card as failed even though the command itself exits 0. A finished card stays in the chat history as its last snapshot.
The same card follows a Snakemake run (steps, local or Slurm executor) and, on the HPC, Slurm jobs and arrays or a Snakemake log. The HPC forms go through the ssh connection Codeome already has open, so they need a connected HPC (Settings > HPC) and are for Claude chats and the Claude hpc-operator; Codex chats use only local logs. Nothing is installed on the HPC.
It needs agent kit 0.9.0 (0.8.0 for the local Vitest, pytest and tsc parsers); projects with an older kit see it as outdated. The parsers and options are in Agent kit.
Attachments and saved messages
The + to the left of the message box attaches files: they show as chips above the box and go to the agent as paths with your next message. The bolt to the right lists the messages you send often (Commit, push and deploy to start with); picking one sends it at once. The same menu saves what you have typed and edits the list.
Specialist personas
An empty chat lists specialists (Analyst, Coder, Manuscript writer) with example requests. Picking one starts the chat with that specialist's instructions in place of the default agent's; typing without picking works as before. The choice locks when you send the first message and shows on the chat tab, next to the engine, and in the chat history. The Manuscript writer always uses Claude, since the Google Docs connectors are Claude-only, so picking it selects Claude and locks the engine choice.
Several chats at once
Each chat is a tab above the conversation and runs its own session, so one can plan while another fixes a bug. ⇧⌘N (or the + in the tab bar) opens a new chat; the + in the chat header creates a new agent, for this project or for all of them; ⇧⌘] / ⇧⌘[ switch. Each tab shows its state: a spinner while the agent works, amber when it is waiting for you, green when a reply is ready that you haven't seen, red after an error. A badge counts the specialists that chat has running. The line under the tabs sums it up (for example 3 chats · 1 working · 1 needs you · 2 subagents · 1 Codex job). Click needs you or reply ready to jump to that chat. Project cards and the rail icon show the same states for every project.
Right-click a chat tab to see the conversation as a time tree, with job branches, plan boxes, timestamps and token counts.
Saved conversations
Conversations are saved (in ~/.codeome/chats/) and keep running while you work in another tab or project. Open tabs come back when you reopen the project. Closing a tab stops its session, and the ⋯ menu reopens the conversation later. The ⋯ menu can also move the current chat into a Claude Code terminal session.
The chat runs the first Claude command in Settings → Agents (for example claude --model opus), with the model chosen there.
Token use
The meter under the message box shows how full the chat's context window is; click it for what the chat has used so far (turns, tokens, and the figure at API list prices). Past 70% (Settings → Agents) it turns amber and the chat suggests starting a new one, since every turn sends the whole conversation again. Each finished turn is also kept as one line in ~/.codeome/stats/turns.jsonl, on this Mac only, and ⋯ → Session statistics sums them up by day, project, chat and by how full the context was when the turn started. Terminal sessions and Codex jobs are not counted.