Skip to content

Chat

Chat is the rail’s last area. ffgit runs Codex’s own app-server, or Claude Code in its stream-json mode, so the agent answers in a chat inside ffgit; a terminal stays a choice. The agents keep your own login and settings, and ffgit keeps none of the chat.

The list on the left shows the workspace’s chats in one of three views:

  • Recent — every chat, pinned first, then newest first under Today, Yesterday, Previous 7 days and so on.
  • Projects — chat projects: a name and a set of folders. Import… makes them from Claude Desktop’s presets and, once you allow it, the Codex app’s projects; nothing is written to either app.
  • Folders — each chat under its first folder.

Chips filter by Needs you, Working and Ready; the sort menu orders by Needs attention, Last activity, Created or Manual and filters by agent and unread. A pen after a title marks a chat ffgit runs, where you can write; other chats open read only until Chat in ffgit or Resume in ffgit. A Claude Code session another app has open continues as a copy, and the original stays as it was.

A chat’s menu, ordered as in Claude Desktop, has Pin, Mark read, Rename…, Fork, Settle and the snoozes, Archive, Copy…, Stop… and Delete…. Settle takes a chat out of the lists, a snooze hides it for an hour, until tomorrow or until its usage limit resets, and Archive keeps it under Archived chats until you unarchive it.

  • Search chats above the list narrows it as you type to the chats whose name or alias, agent, folders or worktree names hold every word you typed, in any order. It searches within the chips and filters you chose; Sort and filter › Archived chats searches the archived ones. Under the results, Search messages for “…”… opens Jump to… with the same words.
  • Jump to… (Cmd/Ctrl+K) finds a chat in any of your projects by words in any order: its name or alias, agent, folder or worktree, its pull or merge request’s number (#12, !12 or pr 12) or the issue it started from (issue #12). Chats that need you come first, then the ones you picked before; archived chats are found too.
  • Search chats’ messages too… shows in Jump to… once you type three characters. It is off until you turn it on, asked once for every project, and taken back in Privacy…. Jump then also lists the Claude Code and Codex chats whose messages — yours and the agent’s, not tool calls or their output — hold every word you typed, in any order, with the words around the newest match (In messages: …). It reads the transcripts of the newest 200 chats it lists, at most the last 32 MB of each, while you type, shows up to 50 of them and keeps nothing. The chat opens at that match with Find in chat, a read-only chat too.
  • Find in chat (Cmd/Ctrl+F in the chat, or the search button in its header) works in every chat, read-only ones included. It finds the rows that hold every word, in any order: messages, reasoning while it is shown, tool calls and their output, plans, questions, approvals and the files a turn changed. It starts at the newest match, counts them (3 of 3) and outlines the current one’s row; ↩︎ goes down, ⇧↩︎ up, and Esc closes it.
  • A read-only chat shows its latest messages first. Search earlier in the find bar reads back through the transcript until a message holds the words, up to 32 MB a press.
  • In a chat that runs in Terminal, ⌘F (Ctrl+Shift+F on Windows and Linux) finds text in the terminal, scrollback included.

New opens a draft in the chat pane: What should we build in …? over the message box, naming the project and the folders the chat goes to. A click on the first folder picks another place to start. Under the box you choose the agent and login, Chat or Terminal, and this worktree or a new one. Drafts stay listed under Drafts, each with what you typed, so you can switch to other chats and come back.

Answers stream in as Markdown, with code in its language’s colors. Tool calls between replies fold into one row that says what they did (Ran 2 commands, read section.png and edited cap.py); open, each step reads as a timeline with its output. File edits are diff cards, a finished turn lists the files it changed in the worktree, and a click opens each file’s diff.

  • Approvals wait on a card that only your click answers. Codex: run once, allow for this session or don’t run. Claude Code: allow, allow for this session (never written to a settings file), or deny with a reason Claude reads.
  • Questions show with numbered choices and a field for your own answer.
  • Plans show whole in Markdown. Claude Code’s plan waits for Approve plan, Approve and use Auto, Approve and accept edits or Keep planning…; a Codex plan offers Implement plan….
  • The agent’s checklist — Claude’s task list or Codex’s plan — sits above the message box.

On a wide window the chat’s status sits beside it: where it runs, how full its context is, the account’s 5-hour and weekly limits, background tasks, the web servers its agent started, the files it produced and what changed in its worktree.

  • / lists the agent’s commands and skills, @ the worktree’s files, ↑ brings back earlier prompts.
  • Attach with +, ⌘V or a drop: images, text files, a diff hunk, a marked region of an image, or a golden test failure with its expected, actual and difference images.
  • The permission mode, model, effort and Default / Plan sit in the box. ffgit never starts or switches a chat into Bypass permissions.
  • Messages you send while a turn runs wait in a queue above the box, where you reorder, edit or cancel them. ⌘↩︎ or Send now sends one into the running turn instead.
  • Templates puts one of your task templates in the message. On macOS the microphone starts the system’s own dictation.
  • Pasted text loses invisible characters that could hide instructions, and the box says how many.

Edit on an earlier message of a Claude Code chat sends it again, as you edit it, on a copy of the chat from before that message; Restore files to before this message can put the worktree back first, after a recovery copy. Fork from here under a reply opens a new chat that goes on from that reply.

⌘↩︎ in New chat starts the task and keeps the draft, so it can go to another agent or model, each in a new worktree. Review attempts… in a chat’s menu then lays up to four attempts side by side from one baseline you pick: the files and lines each changed, its commits, its branch’s CI, and the checks the agent says it ran, shown as the agent’s claim. Choose only marks one; nothing is merged, pushed or removed.

Continue in Codex (or Continue in Claude Code) opens New chat for the other agent with the chat’s first request, its latest messages and command outcomes in the message box, for you to read, edit and send.

  • The chat’s status shows the pull or merge request of its branch: its state, CI, required checks, a conflict and others’ threads, with Send to chat, which drafts a message about them for you.
  • Watch and wake reads the request every two minutes and sends the agent a message when a check fails, the required checks pass, someone comments or reviews, or the branch starts to conflict — until the request closes, you stop it, or ten wakes in a row bring only comments.
  • Start chat from issue… in repository actions puts a GitHub or GitLab issue into New chat’s message box; nothing starts until you send it.
  • Resolve with agent… on the conflict bar drafts a task for the conflicts left. See Conflicts.
  • Review notes from the diff or image review go to a running chat with Send to chat. See Review notes.

Agent accounts… in the main menu adds other logins of Claude Code and Codex. ffgit makes a folder for each and opens the agent’s own login in Terminal with CLAUDE_CONFIG_DIR or CODEX_HOME pointing at it. New chat’s agent menu lists each login (Claude Code · Work) with the account it is signed in to; their chats list with yours, named by their login, and resume with it.

  • New chat shows the 5-hour and weekly limits of the login it starts on, with the share left and when each resets.
  • A chat a usage limit stopped shows Resume at reset: once allowed, ffgit sends a message to continue a minute after the reset, and your own message cancels it.
  • Usage… in the main menu (off until you turn it on) shows the tokens Claude Code and Codex used by model, with the API-equivalent cost. It reads only token counts, models and times from their session files. Limits by login shows each login’s limits side by side.

New chat’s ffgit integration menu offers Orchestrator: a chat whose agent can start full Claude Code or Codex chats, send them messages, wait until one finishes, fails or asks, stop them, and read their state, last reply, changed files and commits.

  • It sees only the chats it started, and at most four of them are open at once (1–8 in Coding agents (MCP)…).
  • Every chat it starts is a chat you can open, write to and keep. Chat and Agents list them under it.
  • With Wake on, it gets a note when one of its chats finishes, fails or waits. With Questions on, it can answer its chats’ questions and plans, while each card stays open to you.

Groups of chats start from a template — Write and cross-review, Plan and render or your own — each role in its worktree with its agent, model and mode. A loop group relays each turn’s reply and changes to the other chat until its stop phrase, three rounds or Stop loop; a reviewer’s path:line findings become review notes.

Gemini CLI, OpenCode, Cursor, Grok and the ACP Registry’s other agents that install a command are listed in New chat when installed, and open as a chat over the Agent Client Protocol on their own login and settings. ffgit installs none of them.