Agent actions

Press A on a commit in the graph to hand it to an AI coding agent: ask it something about the commit, or pick a ready-made request such as Explain branch or Resolve conflicts. tt-git starts the agent in the repository with a prompt that already says which commit, branch and files you mean.

The agent is Claude Code (claude) unless you set another in [agent]. It runs interactively, so you can keep talking to it once it has answered.

Asking an agent

Select a commit in the graph.
Press A, or pick Ask agent in the command palette. The Ask an agent about box opens on Custom message.
Type what you want, or press tab to step through the ready-made actions. The line under the input says what the selected one does. Text you type with an action selected is added to its prompt as extra instructions.
Press enter. esc closes the box without starting anything.

The Ask an agent dialog with its ready-made actions

Custom message needs some text; enter does nothing while the input is empty.

The ready-made actions

Action	Offered	Asks the agent to
Resolve conflicts	only while a merge, rebase, cherry-pick or revert is stopped	read the git log of both sides, resolve every conflicted file and stage it, without committing or continuing
Explain commit	always	explain what the selected commit changes and why
Explain branch	always	explain what the current branch adds compared to the base branch and how the feature works
Review commit	always	look for bugs and risky changes in the selected commit

While a merge or rebase is stopped on conflicts, Resolve conflicts comes first and is already selected when the box opens, so A then enter starts it. When it’s done, check the result in the status pane and press C to continue, as after resolving by hand. See Merge, rebase and cherry-pick.

Where the agent runs

Inside herdr, tt-git opens a new herdr tab named agent: <action> in the repository and starts the agent there. herdr starts agents it knows (claude, codex, gemini, opencode and others) itself and pastes the prompt in, so the tab shows herdr’s agent status. The help bar says agent started in herdr tab "agent: Explain branch".
Anywhere else, the agent takes the details pane’s place, with the focus on it. Type to it as in any terminal. ctrl+\ or f12 gives the focus back to tt-git, and 4 goes back to the agent. When the agent exits, tt-git refreshes, since it may have staged or committed files; any key then closes the pane.

Only one agent runs in the details pane at a time. A while one is running focuses it and says an agent is already running. In herdr each request gets its own tab.

If tt-git can’t reach herdr, it opens the agent in the details pane instead and the help bar says why. If the agent can’t be started at all (for example claude isn’t installed), an error names the program and points at [agent] command.

What the agent is told

The prompt is the system prompt followed by the action’s prompt (or your message), with these placeholders filled in:

Placeholder	Becomes
`{repo}`	the repository’s root folder
`{commit}`, `{subject}`	the selected commit’s full hash and subject
`{branch}`	the current branch (the commit hash on a detached HEAD)
`{base}`	the base branch, as on the Merged and unmerged tabs
`{upstream}`	the current branch’s upstream
`{state}`	`merge`, `rebase`, `cherry-pick` or `revert` while one is stopped, else empty
`{conflicts}`	the conflicted files, one per line
`{merge_head}`	what’s being merged, rebased or picked, by name
`{log}`	`git log --oneline <base>..<branch>`, at most 50 commits

A {name} tt-git doesn’t know is left as it is, so braces in a prompt need no escaping. The prompts point the agent at commits rather than pasting whole diffs: the agent runs in the repository and reads what it needs with git.

The default system prompt tells the agent it was started from tt-git, which repository and branch it’s in, to use git itself, and not to push or commit unless you ask.

Configuring it

[agent]
command = "claude"     # the agent to run
placement = "auto"     # "auto", "herdr-tab", "herdr-split" or "details"
# system_prompt = "..." # replaces the default; "" for none

[[agent.actions]]
name = "Write release notes"
description = "user-facing notes for this branch"
prompt = "Write release notes for what {branch} adds since {base}:\n{log}"

[[agent.actions]]
name = "Resolve conflicts"
when = "merging"
prompt = "Resolve the conflicts in:\n{conflicts}\nStage each file. Don't commit."

See [agent] for every setting. Setting [[agent.actions]] replaces the built-in actions, so copy the ones you want to keep.