Agents

Project agents

A project’s Agents tab can hold any number of named agents, each either Claude Code or Codex, chosen when you make it with New agent. The two are equals. They work in the same workspace, each with a session that goes on when you close the page.

• Details: each agent’s row in the sidebar has a small “i” button that opens its details: its type, model, reasoning level, permission mode, and when it was made.
• The chat shows the agent’s text live, its thinking, its tool use (folded, with the changes it made to files), its questions and permission prompts, attachments, the model, usage and a stop control.
• ProjectStart’s notes in a chat (the agent stopped, renamed or cloned, a setting changed, a login switched, a definition change decided, a sub-agent finished, and the like) show as small cards with an icon, their words and their time.
• Only what you typed is yours: whatever ProjectStart sends an agent by itself, or when you tap a button rather than type (a dismissed question, Ignore my last message, a task’s Start or Pause, a schedule’s run), shows as ProjectStart’s card, never as a message you typed. Your messages and your answers to questions show as yours.
• Agents talk to each other: agents find one another by name and send messages that show who sent them. Each agent knows its own project’s name, and agents of a linked project address it as name@project (see Linked projects).
• Reading other chats: an agent can read the chats of its own project’s other agents and of linked projects’ agents. From a linked project it gets only the messages’ text, never tool output. The agent whose chat is read isn’t told.
• Agents in the trash aren’t listed and can’t get messages. An agent can read their chats only when it is asked to, until they are deleted for good.
• You can read every chat, stop any agent (see Actions) and remove any agent.
• History records an agent’s creation, stops, renames, trash, restore and removal, and deleted files. Messages and answers aren’t History entries: each agent’s chat keeps them, with who answered and when.

Claude logins and switching

Settings › Agents keeps several Claude logins, in the order you added them.

• Each shows its usage (Current session, This week, and Fable this week) and whether its account’s extra usage is on. ProjectStart never changes an account’s settings.
• Agents use the login in use for their model family (default or Fable). A switch never changes an agent’s model.
• Under When the usage limit is reached, two switches: Use extra usage (off by default) and Switch to the next login (on by default; it needs at least two logins).
• At a limit, the login in use spends its extra usage first when that switch is on. Then ProjectStart switches to the next login with usage left and moves the running agents to it. With neither, the agents stop until usage is back.
• An agent a limit stopped is told to continue once, as soon as its model family has a login with usage again (after a switch, or when its window resets). Its chat gets a card saying usage is back.
• Use this login now switches by hand. Every switch shows a short notice naming both logins, and each agent moved gets a card in its chat naming the login it moved to.
• Near the limit: when a login in use has used over 80% of one of its usage windows, a usage chip shows in yellow (over 90%, in red) under the row of buttons at the top of the sidebar and in the chat header of each Claude agent on that login. It names the window, its share and when it resets, and says what happens at the limit. Tapping it opens the logins in Settings. Below 80% nothing shows.

Codex logins and switching

Codex logins work as Claude’s do.

• Add a Codex login adds one; the first becomes the one in use. Each can be removed or put in use by hand.
• Each login shows its 5-hour and weekly windows (the share used and when each resets, and its credits when Codex reports them), with when they were measured. They also show in a Codex agent’s “⋯” menu.
• Switch to the next login (on by default; it needs two logins) moves the running Codex agents to the next login with usage left when the one in use reaches its limit: an idle agent at once, a busy one when its turn ends, then told to continue. Otherwise the agents stop until the limit resets.
• The usage chip warns as for Claude: in each Codex agent’s chat header, and in the sidebar’s chip, which shows the highest share among the Claude and Codex logins in use.

Codex agents

• Making one: the New agent dialog chooses Claude Code or Codex. Codex is offered once Settings has a Codex login, in container projects and on remote hosts (on a remote host with an older runner, after Update runner). Otherwise it shows as unavailable, with the reason.
• In the chat: the same chat as a Claude agent’s: live text and a summary of its reasoning, commands with their output, file changes, its plan as a task list, permission prompts, its sub-agents, attachments, Stop, model and reasoning level, and usage per turn. A message sent while it works joins the turn under way.
• Permission modes: Manual (Codex asks for commands outside its sandbox and for changes), Accept edits, Auto (Codex’s own reviewer decides instead of you), Plan (read-only) and Bypass permissions (no sandbox, no prompts).
• Actions: every action works as for Claude agents.

Agent rules

Every agent gets ProjectStart’s chat rules at every start, wherever it runs. They come first, and nothing can turn them off; you can read them in Settings › Agent instructions. Besides them, agents get your own rules:

• Rule sets: Settings › Agent rules keeps any number of named sets. Each is either a Git repository (an HTTPS address on an allowed Git host, the credential that can read it, and an AGENTS.md at its root) or Markdown written in a text box. Each can be ticked for new projects (Tick it for new projects). They are optional: without any, agents run with the chat rules alone. A rule set a project uses can’t be deleted.
• A project’s rules: Create and the project’s Settings choose which rule sets the project uses. The project’s Settings also has a text box for the project’s own rules.
• Order: Claude and Codex read the same rules in the same order: the chat rules, then each chosen rule set in the order Settings lists them, then the project’s own rules.
• Checked: each start checks that the rules are in place before the agent runs; when they can’t be, the agent doesn’t start and says why.
• Changes: a written set or the project’s own rules reach each agent at its next start; agents running then carry on with what they have. A repository set changes only when you ask: Update rules on the set in Settings updates every project using it, and the project’s page updates that project’s sets. Each update shows its changes and risks first, and Roll back rules puts the previous version back.

Everything ProjectStart tells the agents is shown, read-only, in Settings › Agent instructions: the chat rules, what each agent is told at its start, its tools, and each message ProjectStart sends by itself, with when.

Compaction

A long conversation is compacted into a summary, so each turn reads less.

• A Claude agent compacts on its own when its context reaches 400K tokens; a Codex agent does so at the start of its next turn once it passes 400K.
• Compact conversation in the agent’s Actions compacts now. Asked during a turn, it runs once that turn ends. It isn’t offered while the agent isn’t running or a compaction is under way.
• While it runs, the agent’s status says it is compacting, and a card in the chat shows it: an automatic one is titled Automatic compaction. At the end the card shows the conversation’s size before and after.
• History records who asked for a compaction and when.

When an agent is cut off

An agent’s runner can end without anyone stopping it: its connection is lost, or it crashes.

• During a turn: the agent is started again at once, resuming its session, and told that its turn was cut off, with the sub-agents it lost, so it carries on. The chat shows a note saying so.
• While idle: the agent isn’t started again; the note says it starts again with your next message.
• An agent is started again this way at most once in 5 minutes. If it ends again sooner, it stays stopped, with a note saying why, until a message starts it.
• None of this happens after you stop the agent, after it is moved to the trash or deleted, or when it couldn’t start at all.

ProjectStart itself can be updated while agents work: an update never cuts an agent’s turn, its sub-agents, or a prompt or question waiting for you.