All mods

After each turn, suggests three next prompts above the prompt box. Type 1, 2 or 3 (or click) to send one. Uses a small Haiku call per turn.

When you approve a plan, saves it as .claude/tickets/ /plan.md in the project and points Claude at that copy.

When a prompt names an earlier session's ID, attaches a digest of that session (prompts, files, commits, last reply) for Claude.

A band above the prompt with the repo's branch, changed files, unpushed commits and PR, plus Commit & push, Open PR and Diff buttons.

A themed bar above the prompt: live context meter, plan limits, session cost, a token-saving tip, and one-key Compact and Theme buttons, plus a /glow theme picker. Uses Claude Code's early-access function hooks (2.1.286 or newer).

Sidequest for Claude Code: play Snake in a side pane while Claude does the main quest; works in any terminal

Shows above the prompt how long the cache stays warm, what each message costs on your Claude plan, and one habit that saves usage

A pixel theater above the prompt: little orange Claude acts out each tool call at a desk while a monitor shows the command and Haiku narrates it.

入力欄の上に、真上から見た浅瀬の波をドット絵で動かす Claude Code のモッド。朝・夕焼け・夜が巡り、波の音 (Windows / macOS / Linux) と、ときどき Claude くんが遊びに来る。(非公式)

Grammar fixes for prompts you type or dictate: optional auto-fix on send (LanguageTool rules or a light Haiku fix), Haiku drafts that know your project's terms, and word tools (your language to English, synonyms).

A live map of your Claude Code session: a git-style flow chart, a live feed of every tool call, a drag-and-drop task board, async grill questions, agent chat, stats and recap, in a pane you can steer from.

A pixel dinosaur park under the spinner: a coral T-rex acts out the agent's work, and every failing test run brings the meteor closer. No model calls.

Pin a session to its git worktree: every Bash call runs from the anchor, and the primary checkout is guarded against stray edits and git writes.

No claims without evidence: flags "tests pass", "CI green" and "verified" claims that nothing in the turn backs up.

A second player reviews before the PR: gh pr create waits for a second-model review of the branch diff, and a failing review with a high finding blocks it.

Container stacks per worktree: a pixel-art pane of every compose stack with its memory and worktree, orphans whose worktree is gone, and a guard that asks before booting another stack when memory headroom is low.

Make silent shell failures loud: a glob that matched nothing, a command missing from the agent's shell, or a pipe that hid a failing exit status gets a note the model reads and a peach pixel-art toast.

Fix the arguments, not the model: repairs malformed MCP tool arguments ("true" for true, numbers as strings, unknown keys) against the tool's schema before the call, tells the model what it fixed, and flags servers that are down. A pixel-art patch log.

A raid frame for every live session on the machine: who is working, who is waiting on you and for how long, per-PR locks so two sessions never act on one PR, and /broadcast to the rest of the party.

The fix must make a test fail first: reverts your source changes, runs the changed tests and requires a failure, restores (verified by hash) and requires a pass. Optional gate on git push and gh pr create.

A save point before compaction: the task's goal, step, branch, links, decisions, rules and next step are saved before /compact and handed back to the model after it, with a pixel-art SAVE POINT band when it is safe to compact.

Memory at the moment it matters: when a tool call touches what one of your memory files is about, radar attaches that memory's description and rules for the model to read. A lavender pixel-art radar pings each match.

Auto-continue after a turn dies on a network or API error: a pixel-art CONTINUE? countdown with backoff and 3 lives per hour.

Watch PRs outside the model: polls CI, reviews and threads with gh, wakes the session only when something actionable happens, and denies sleep-poll loops while it watches.