Prompt

Gates git work: no commit, push or merge without the user's say-so in the current turn; a push to a branch that already landed is refused; an MR whose branch lacks the target's latest commit is refused; a pushed series must leave the project working at every commit; commit messages kept tidy (format and the issue it is about, then a Haiku review of the body).

Nexika mizan: keeps your Claude Code session in balance. A band above the prompt shows the branch and who started it, open PRs or MRs per person, the branch's CI (passed, failed with the job, running), RAM and disk with warnings, the context level (fresh, mid, full), the running agent and its task, the current task step and the session cost; /mizan opens the details. At full context it saves a...

The learning loop: every main turn is kept in a ledger across sessions (phase, effort, steps, failed tool calls, agents, advisor calls), joined with the outcomes observed at tool.call (commits kept or reverted, test exits, PR verdicts and merges, no transcript) and the ones /insights writes, and the project's goals; probe sessions are counted and kept out; a phase whose evidence, read on the...

Colours what needs you at the end of Claude's replies (must-know, decide now, decide later, defaults, your actions) and pins open decisions above the prompt

Idle-session guard for Claude Code's 1-hour prompt cache (macOS): about 50 idle minutes after the last turn of a large session it notifies you, optionally after a cache-read fork judges whether to compact, keep warm or leave it; every request and change it makes is audited. Derived from cache-tax (MIT, Karan Bansal).

Keep your API keys and tokens out of your Claude Code conversations: tokens pasted as keyfob: lines are stored before the model sees them, Claude asks for a key through the /keyfob panel, a guard stops Bash commands that would print one, and each command gets only the keys it needs from the OS keystore

Claude Code speaks through murmur (Kokoro on the Apple Neural Engine): turn summaries, needs-you alerts, background tasks, progress check-ins and rate-limit warnings, each session in its own voice (/say picks one)

Picks the model and the reasoning effort for each task with TypeSafe's Jev, a System One decision model reached either through TypeSafe's own API or the Vercel AI Gateway. Routes subagent models at agent.spawn and the main loop's effort and model at turn.step, in both directions, falling back to the engine's built-in classifier when no key is set.

/today: the day's agenda for Claude and for you. Calendar events, E3 deadlines and Reminders are read through the macos and nycu plugin scripts, shown as one line above the prompt with a countdown to the next event, and handed to Claude as a context block so it knows your day.

Picks the reasoning effort for each turn of the main conversation with a System One classifier (hosted Jev or a compatible service) and shows the level on the spinner and on the line that closes the turn. In shadow mode it shows and logs its pick and sends the session's effort; in enforce mode it sends the pick. Requires CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1.

Rules for Claude Code alerts: when one is worth sending (done, blocked, error, usage limits) and where it goes (chime, macOS notification, ntfy, Slack, Discord or a JSON webhook).

In a worktree session, tells Claude up front which commands the worktree guard refuses (heredoc and inline-script edits, multi-line scripts, cd to computed paths), so it uses Edit and plain commands instead of retrying.

Git only when you asked: commits, pushes, pull requests, merges and destructive git commands run only when your latest message (or /ship, /kickoff (also as /progress-pane:kickoff), or your answer to Claude's question) allowed them; everything else is refused with a note telling Claude to ask. No UI.

Mission control for Claude sessions, in a Notebook look: a band and /progress pane in every session (model, effort, tokens, context, agents, gate, git), and with a worklog a validated worklog tool Claude updates at every step (done only after a green gate), facts the mod sees for itself (gate runs, commits, branch, agents, context and cost), drift hints where the two disagree, a band above the...

カニード が Claude の作業中にダンジョンを自動探索する Claude Mod。回復は Lv アップとテスト成功、あとはマップで見つけたショップだけ。倒れると Lv 1 からやり直し。会話ログのパス・#N・URL もクリックで開ける。Needs function hooks (CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1) and an interactive terminal.

Keeps the prompt cache warm while you step away: a short poke before the cache expires, so the next turn reads the cache instead of writing it again. A row above the prompt turns it on and off and edits the message. Near the usage limit it books Claude a wake right after the reset.

Shows the LaTeX math in Claude's replies as real equations: images in kitty and Ghostty, text elsewhere

Keyboard picker for your saved prompts: search snippet files, fill their placeholders, and drop them into the Claude Code prompt box, plus create, edit, duplicate, move, pin, delete and restore snippets without leaving the session. Snippets are plain Markdown files, not skills, so they cost the model no context until you use one.

Counts every skill expansion (typed /name, the Skill tool, a preload) in the plugin store, across sessions; /skill-usage shows the table. Observe-only: never changes what the model reads. No files, no network. English by default, Russian by a one-line switch.

A board with two tabs: this session shows the task list that Claude and you share, and this machine shows every local session, the ones that wait for you first. /board opens it, and four model tools keep the list.

Prompt-cache meter above the Claude Code prompt: how many tokens each request read from, wrote to and sent past the cache, a live countdown to the cache's expiry, and what to do about it (keep going, compact or clear). The TTL follows Claude Code's own rules for your account (1 hour on a Claude subscription, 5 minutes on an API key, cloud provider or usage credits, plus the TTL variables and...

Checks your prompt draft's English with Haiku after a pause and shows the whole draft corrected in the band above the prompt, with the changed words colored by kind of error and a short reason for each. Mute the kinds of error you do not want to see.

Context-window squares by category above the prompt, with a prompt-cache countdown and an optional keep-warm refresh. Reads the context breakdown, the prompt draft's length and the session transcript's cache usage, all locally; a refresh (manual, or auto-refresh if enabled) sends one model request through Claude Code. Inside Herdr, it also flashes the session's entry in Herdr's agent list as...

Keeps unattended sessions going: continues after usage limits reset, retries API overload with backoff, resumes cut-off responses, and compacts idle sessions before the prompt cache expires.