Guards

Attaches a verification reminder to every delivery from another agent: a subagent's hand-back, a background task's notification, another session's message

The review gate as a tool.check function hook. Every shell call is decided in process, so the non-commit calls -- almost all of them -- cost no subprocess at all, and a blocked commit can be put to you as a question instead of refused outright. The decision itself is still bin/review-gate, run only for a command that mentions a commit, so both routes read the same receipts and cannot disagree.

一時ディレクトリ (tmp/scratchpad) からの Artifact publish を拒否し ~/Documents/claude-artifacts への保存へ誘導する

dotfiles のガード群 (git 操作のスキル強制・ブランチロック・コメント規約・文脈注入)。機能ごとに hooks/ 配下の 1 ファイル

ship フローの状態機械をペインで可視化する。/ship-hud でトグル、進行中フローがあれば自動で開く。モデル向けに status ツールも提供

Switches the Claude in Chrome extension between named browser profiles with select_browser, which needs no approval click, and picks the default before the first browser call of a session.

After a deploy command succeeds, curls the live URL with cache-busting and puts the verdict in the model's context, so a deploy cannot be claimed without evidence.

After every Edit, Write, NotebookEdit or sed -i, a docked pane shows that file's uncommitted hunk with [keep] and [revert] buttons; a press dismisses or runs git checkout with no model turn.

Denies a Bash merge (gh pr merge, git merge on main, push onto main) unless the latest human message says the word merge; ship, push and deploy do not count.

Routes heavy Bash commands (test suites, builds, Docker) to a second machine over ssh, with the PATH export and working directory already right, and the commit synced first.

Keeps a per-turn tally of edits, runs, curls and destructive commands and shows it on the turn footer; a claim with no run flags the spinner, a destructive Bash unfolds its tool group, and an edit-without-run turn suggests running the tests.

Counts the distinct files a turn edits and stops at a threshold to make the goal get restated, so a small ask cannot quietly become a refactor.

A pixel-art Ragtag athlete above the prompt who does a rep every turn: the session is an AMRAP of thrusters, burpees and pull-ups.

Bouncer for Claude Code: enforce condition-based markdown guardrail rules on tool calls — deny or annotate any call whose content matches, instead of file-path-only rules.

Jev-scored context pruning for Claude Code through OpenRouter: at compaction every tool call and result is scored by TypeSafe's Jev decision model, stale ones are dropped or truncated, everything kept stays verbatim. An optional gate truncates oversized tool outputs as they arrive.

Every tool call your agent makes, checked before it runs. Verifies each pending tool call against the session's plan using TypeSafe AI's Jev model, then allows it, asks you, or blocks it.

A pane beside the transcript with the GitHub pull requests and issues related to the current session: drag a title or description to leave a comment on it, add as many as you like, then Submit sends them all as one prompt. Watch each pull request's checks, review decision and mergeability update on a timer.

An unofficial fan-made pixel cat that lives in the band above the prompt and reacts to what Claude is doing: thinking, editing, craning after a search, yowling in a shell, squeezing its eyes shut at a tool error, and folding its ears down to sleep between turns.

A prompt typed while Claude is working is held in a stack above the prompt instead of landing in the running turn, and sent once the turn ends. Reorder the stack, edit a row in place, send one first, remove one, flush the lot. Needs function hooks (early access) and an interactive terminal.

Catch wild creatures above the Claude Code prompt: where your prompt leads (water, grass, fire, bugs, plans, builds) decides what appears. Pick a partner that draws under the logo, beside the spinner and above the prompt, and levels up and evolves as Claude works. Needs function hooks (early access) and an interactive terminal.

Dim 'time ago · short summary' labels under the last 20 messages in Claude Code

PII and secrets are swapped for placeholders (PII_PERSON_1) by regex and a local Gemma 4 before anything reaches Claude; placeholders are restored only where the work stays on this machine (local tool calls) and on screen.

Meters above the prompt: the context window, the session's spend and what is left of your limits, a small mascot that walks them while Claude works, and a line under them with an AI tip, a joke or a piece of good news.

A live Codex session in a side panel: chat with it, pick model, effort and access; Claude and Codex message each other (the codex tool, @claude:), and Codex sees Claude's prompts, tool calls, replies and transcript.