This page shows what you see in your first session with claudetto. It is for users who finished Install. The install steps are not repeated here.

The core plugin claudetto adds an output style, six agents, a status line, and hooks. The hooks ask, deny, and check. See Architecture for the parts.

After you apply the settings snippet with /claudetto:setup, you see these parts.

• The status line, with the parts in the next section.
• The output style claudetto:Concise. It sets how Claude writes its replies. See Prompt.
• A drift note, when the active settings differ from the snippet.

Claude receives the drift note at session start. The note lists each difference. It states that the user has not seen the difference. Some lines name /claudetto:setup as the fix. When no difference exists, the note is not added. See Settings Snippet for the keys of the snippet.

The status line has up to three rows.

• The first row shows the folder, the git state, the pull request, the active change as ◆ name 4/13 ▸5.1, and the 7-day limit.
• The second row shows the model, the context, the cost with the lines added and removed, and the 5-hour limit. The number after ⇥ in the context is the distance to the handoff trigger.
• The third row shows the spend limit and the prompt cache.

The colors follow the theme of Claude Code, and NO_COLOR turns the color off. The width comes from the COLUMNS env var. When COLUMNS is unset, the width is 120 columns, and 4 columns stay free at the right edge. The status line calls no model and no network. It runs no program except git. See Status Line.

The plugin adds six agents. Claude starts an agent with the Agent tool, when a task fits the agent.

• investigator answers a question from several files, logs, or git history. It is read-only.
• researcher answers from web pages and official docs, with quotes and sources. It is read-only.
• planner plans a change across files, and it writes the plan only under minispec/changes/<name>/.
• implementer implements one scoped slice with a known check, in its own worktree.
• debugger finds the root cause of a failure, and it makes the smallest fix.
• reviewer reviews code, a plan, or a design. It is read-only.

No agent runs the tests. The checks slice has the claudetto-run-checks CLI, which replaces a test-runner agent. See Agents.

MiniSpec keeps a plan in files in your project. The folder minispec/specs/ holds the specs, and the folder minispec/changes/ holds one folder for each open change. Run /claudetto:minispec to plan a change, /claudetto:minispec-apply to do its tasks, and /claudetto:minispec-review to review it. The minispec CLI reads and updates the same files. See MiniSpec.

Claude Code does not compact the context in the default setup. At 225,000 tokens of the 300,000 window, the handoff writes a file with the goal, the state, and the next step. For Haiku the trigger is 100,000 tokens. The file goes to minispec/changes/<name>/handoff.md when a change is open, and else to .claudetto/handoffs/. Then a clean session reads the file. Run /handoff to start it by hand. See Handoff.

The permission rules, the sandbox, and the hook module guard each call. See Guards for the full list.

The settings snippet makes Claude Code ask you first for these commands. It also asks for a force push, git reset --hard, and other commands that discard work.

• A package publish, such as npm publish or cargo publish.
• A write to a pull request or an issue, such as gh pr create or gh issue comment.
• A recursive rm, such as rm -r or rm --recursive.

A Bash command that prints a whole file is denied. The denial names the Read call with the range to use.

The permission rules deny a read of secrets, such as .env files and ~/.ssh/**. They also deny rm -rf of the root or of your home folder.

The destructive floor denies a destructive command. The floor denies git reset --hard, a force push to main or master, and a forced checkout or removal that drops uncommitted work. Claude stops, and it does not try the action another way. The denial gives you the exact command in a block. Run that command yourself with ! at the start of your prompt, if you want the action.

• Guards: what asks, what denies, and what the sandbox limits.
• Handoff: how a session ends before the bound.
• MiniSpec: the files and the tasks of a change.
• Troubleshooting: what to check when a part does not show.