Claude Code Tools

The writing plugin provides two explicit-trigger skills designed to improve prose quality. These skills do not fire automatically during ordinary writing tasks; they must be invoked by name.

• agent-style: Applies 21 literature-backed rules for formal technical prose (e.g., papers, design docs, proposals, READMEs, commit messages). Based on The Elements of Agent Style.
• remove-ai-patterns: Removes AI-writing patterns ("AI-isms") from text across any register. It uses voice profiles and a deterministic Node.js detector. Wraps avoid-ai-writing.

Usage Note: The two skills are complementary but should not be interleaved. Use agent-style for clarity and composition in formal prose, or remove-ai-patterns to de-AI text of any voice. Pick one as the final gate for a document.

The voxtype plugin provides the voxtype-install skill, which automates the installation and configuration of voxtype—a local, on-device voice dictation tool for macOS that types wherever your cursor is located.

Once installed, you can trigger the setup process by asking your agent (Claude Code or Codex) commands like help me install voxtype or set up voice dictation. The agent will guide you through the following sequence:

1. uv tool install voxtype
2. voxtype setup
3. Launching the tool and completing macOS permission steps.

aichat is a unified CLI command-group designed for managing Claude Code and Codex sessions. It provides two primary capabilities to overcome the limitations of standard session compaction:

1. Resume with lineage: Allows you to continue sessions when context limits are reached. Instead of just lossy summarization, it uses trim or rollover strategies to clear context while preserving links (lineage) to parent sessions, enabling on-demand context recovery.
2. Full-text search: Provides a fast, Rust/Tantivy-based search engine across all sessions. It includes a TUI for human users and a JSONL output mode for AI agents to find past work.

Note that aichat's search is a full-text search, whereas Claude Code's built-in search only searches session titles.

The claude-code-tools project provides a suite of CLI tools, skills, agents, hooks, and plugins designed to enhance productivity when using Claude Code and other CLI-based coding agents.

Key components and tools available in the ecosystem include:

• Core Tools: aichat, voxtype (local voice dictation), tmux-cli, agent-tunnel, lmsh, vault, env-safe, sasy-guard, statusline, and fix-session (session repair).
• Integrations: Google Docs, Google Sheets, and alternative LLM providers.
• Plugins & Extensions: Safety hooks and voice plugins.
• Guides: Claude to Codex workflows and general getting started guides.

tmux-cli includes several automation features for managing terminal sessions:

• Pane Management: Launch shells/programs in new panes and kill panes.
• Input Control: Send keystrokes with configurable delay and auto-retry.
• Output Inspection: Capture pane output for analysis.
• Synchronization: Wait for a pane to become idle before proceeding.
• Execution: Execute commands and retrieve exit codes as JSON.
• Interruption: Send Ctrl+C or Escape to any pane.
• Safety: Includes a guard to prevent killing your own active pane.

The tmux-cli skill allows Claude Code to communicate with processes running in other tmux panes. This is useful for interacting with:

• Other AI code agents (e.g., Claude Code, Codex CLI, Gemini-CLI)
• Interactive scripts waiting for input
• Long-running processes that require monitoring
• Debuggers like pdb for stepping through code

Use the tmux-cli command to send input to and capture output from these panes. Run tmux-cli --help to view all available options.

When quiet_mode is active, the following outputs are suppressed:

• Agent streaming output
• Intermediate messages and tool outputs
• Rich console spinners/status messages
• Response statistics (show_stats)
• Debug information

sasy-guard is a deterministic security guard for Claude Code. It uses policy-compiler principles to compile policies into Soufflé Datalog rules.

Unlike simple single-command checks, sasy-guard reasons over a dependency graph rebuilt from the whole session. This allows it to detect multi-step attacks, such as:

• A secret being read in one step and later leaving the machine.
• A git push occurring without a clean secret scan.
• Large, unreviewed changes attempting to ship.

tmux-cli is a command-line tool designed for terminal automation, specifically allowing CLI agents (like Claude Code) to control applications running in tmux panes and windows. It provides capabilities to launch, send input to, capture output from, and kill programs.

Auto-Detection Modes

tmux-cli automatically detects your environment and switches between two modes:

• Inside tmux (Local Mode): Manages panes within your existing, current tmux window.
• Outside tmux (Remote Mode): Creates and manages a completely separate tmux session with its own windows.

Dynamic Workflow is a system for creating, reviewing, and running durable JavaScript workflows that coordinate multiple headless Codex agents. It uses a deterministic JavaScript program to manage control flow while Codex workers handle reasoning and tool execution.

Use a workflow when JavaScript control flow materially reduces context or coordinates patterns such as:

• Fan-out/Fan-in: Discovering items and running one worker per item, then synthesizing results.
• Parallelism: Running heterogeneous agents in parallel and combining results.
• Branching/Looping: Driving logic based on structured worker output.
• Long-running tasks: Executing background runs with durable progress.
• Portability: Reusing or porting existing dynamic workflow scripts.

Do not use for small, linear tasks that a single Codex turn can handle directly.

In Codex, Workflow status and Callback status represent different lifecycle stages:

1. Workflow Status: Indicates if the work is starting, running, pausing, paused, canceling, unverifiable (non-terminal but identity uncertified), failed, completed, or canceled.
2. Callback Status: Indicates the state of the completion notification delivery to the app server.

Callback States:

• armed: Target passed preflight and awaits terminal result.
• sending: Notifier is delivering/confirming the message.
• delivered: App server confirmed receipt.
• failed: Delivery stopped without an accepted submission.
• unknown: Delivery status is ambiguous.
• invalid: Dashboard could not validate callback metadata.

Note: A delivered callback does not mean a workflow succeeded. For example, a failed workflow can have a delivered callback status.