agent-box

Shadow and version-control another agent's conversation. Use when asked to track an agent, create agent biography, shadow an agent, document agent work, or make agent output more accessible.

Configure Agent HQ Electron crash mitigations + diagnostics via repo .env (added after 2026-01-22 renderer crash).

Start the full agent stack (tmux sessions for Claude Code agents, SQLite ingestors, announcers, watchdog, and mitmproxy) after a fresh Mac reboot. This skill should be used when the user asks to "start everything", "bring up the agent stack", "start after reboot", "launch tmux sessions", or needs to initialize the development environment with all monitoring and agent processes.

Bootstrap the complete agentic ecosystem (Claude Code, Codex, vscode-shims, agent-box) on a remote machine. Use when user asks to install ecosystem, setup remote agent, bootstrap agent box, deploy to remote, or install claude/codex on another machine.

This skill should be used when containerizing a multi-service agentic ecosystem (UI + Agent Management Server, optionally launchers/ingestors) with Docker Compose while preserving localhost-only security, correct in-container networking, centralized log volume mounts, and working Codex/Claude CLIs.

Safe incremental deployment and updates for the agentic ecosystem after initial setup. Focuses on config preservation, safe service restarts, verification, and optional testing. Use after agentic-ecosystem-remote-deployment has completed at least once.

Send prompts directly to Google Antigravity (Gemini + Claude) via Python using OAuth refresh tokens, without OpenCode; includes correct headers, model ID normalization, endpoint selection, SSE streaming, and 429 backoff.

Bootstrap a new daily project folder for stock transaction screenshot processing. Use when user asks to "bootstrap new day", "create new day folder", "set up YYYY_MM_DD folder", or mentions starting work on a new trading day. Creates directory structure, copies documentation and prompts, optionally copies stock mappings, and opens known_issues.txt for user editing.

Use Claude Code CLI in stream-json mode for real-time bidirectional JSON communication. Enables agents to pipe JSON messages to Claude and receive streaming JSONL output with full session observability including tool calls, deltas, and results.

Convert skills between Claude Code and Codex CLI formats. Use when porting skills from ~/.claude/skills to ~/.codex/skills (or vice versa), converting project-level skills between .claude/skills and .codex/skills directories, or troubleshooting why a copied skill isn't loading. Covers all skill levels (user, project/repo, system, admin) and gotchas like symlinks, hidden dirs, and field length limits.

Document how to capture the directional Claude extension/shim traffic so the renamed JSONL streams (`claude_*` and `claude_shim_*`) are always written and can be compared against Codex logs.

Benchmark Claude Code startup/response time across different CLAUDE_CONFIG_DIR locations. Use when user asks to test Claude speed, benchmark config dirs, compare startup times, or measure latency.

Programmatically fetch Claude Code usage limits and remaining quota via API. Use when user asks about usage, rate limits, quota, billing blocks, or remaining capacity. Also handles OAuth token refresh for API access.

Clone a repo to ~/temp with unique -cloneN, bump all *_PORT values in clone .env by +100*N, then work only in the clone and never touch the original unless explicitly asked.

Deploy Codex CLI authentication to remote machines. Use when user asks about codex auth, codex authentication, codex oauth, codex remote, codex login remote, copy codex credentials, or deploy codex to server.

Display terminal color gradients and palettes for choosing colors. This skill should be used when the user asks to see color options, color gradients, color palettes, ANSI colors, RGB colors, terminal colors, or needs help picking a color for terminal/shell configuration (e.g., Powerlevel10k, zsh themes, prompt colors).

Ultra-fast token+cost computation for Codex + Claude SSE logs (Rust) plus Python anchor/fast comparisons; includes build+run+benchmark guide.

Configuration philosophy requiring centralized, transparent .env files with no hidden variables, silent fallbacks, or defaults. Use when dealing with config, configuration, env variables, background variables, fallbacks, or dev philosophy. All configuration must be explicit in .env with .env.example as ground truth.

Create a new Viber Stock OCR project folder for a new trading day. This skill should be used when the user wants to set up a fresh project folder for processing stock transaction screenshots from a new date. Triggers on requests like "create new day project", "set up folder for 2025-12-25", or "initialize new OCR project from old day".

Split multi-day A5 Excel files into single-day files for the db23 trading data pipeline. Use when processing Excel files in /home/ubuntu/Downloads/multi-day that contain multiple trading days, when the user mentions "split excel", "multi-day excel", "A5 format", or needs to prepare Excel files for the 6-step pipeline. This skill covers validation of sheet names, column structure, and data format requirements.

Run the 6-step db23 trading data ingestion pipeline. Use when processing single-day Excel files through the pipeline, when user mentions "run pipeline", "6-step", "push day", "label day", or needs to ingest a new trading day into df_trans. Covers all steps from parsing Excel to remote sync.

Prompt template for documenting debugging sessions after successfully fixing bugs. Use this to capture debugging methodology, hypothesis evolution, red herrings, breakthrough moments, and lessons learned for future agents. Trigger keywords - debug knowledge base, document debugging session, debugging methodology, post-mortem analysis, incident report debugging, debug report template, capture debugging process.

Technical reference for dumping Codex and Claude agent conversations to searchable text files. Command reference, flags, and output format. Used as prerequisite by search-agent-conversation skill.

Patch Claude Code cli.js to enable fine-grained tool streaming (input_json_delta). Use when tool call arguments arrive as a single burst at the end instead of streaming incrementally, when sse_lines.jsonl shows few/no input_json_delta events, or when asked to "enable tool streaming", "fix partial json", or "patch fine-grained streaming".

Estimate OpenAI weekly usage limits from ground-truth token cost + reproducible pricing snapshots (billing week resets 14:47 local).

Create comprehensive incident reports with knowledge graphs. Use when debugging production issues where you need to trace root cause through multiple code entities. Documents debug process, entity relationships, reasoning→pattern→codebase chain, and prevention strategies.

Generate interleaved multi-agent conversation reports organized by theme. Use when user asks for "interleaved conversation", "conversation by theme", "multi agent conversation timeline", or wants to see chronological prompts across agents grouped by user intention.

This skill should be used when investigating why the Codex VS Code extension fails to spawn or connect to a local Codex agent, especially after updates or CLI changes.

Create iTerm2 multi-pane grids for running parallel agents (gemini, claude, codex, etc.) via Python/AppleScript automation. This skill should be used when the user asks to create iTerm2 pane layouts, run multiple CLI agents in parallel, automate terminal workflows, or send commands to multiple panes. Covers critical gotchas like properly triggering Enter for interactive programs.

This skill should be used when MCP stubs or broken control_request flows need to be replaced with ground-truth implementations aligned to extension.js and cli.js, or when MCP-related UI behavior (missing tool outputs, spinner stuck, empty MCP list) needs debugging.

Configure OpenCode + opencode-antigravity-auth plugin to use Google OAuth (Antigravity) for free Gemini/Claude usage; includes auth-store gotchas, verified workaround, and troubleshooting with logs.

This skill should be used when ensuring Codex emits centralized observability logs (sse_lines.jsonl) while running in app-server mode, especially when used by the VS Code extension.

Download, prettify, and archive @anthropic-ai/claude-code npm packages. Use when user asks to "archive claude code", "download claude code", "get latest claude code", "npm pack claude code", or needs to set up a new Claude Code version in ~/swe/archive with prettified cli.js and git history.

Run the 'push to df_trans' workflow on the remote M3 Ultra machine. This skill should be used when the user asks to push new trading days to df_trans, update transaction data, or run known_good_version.py on the remote server. Handles SSH connection, PYTHONHASHSEED requirements, and sequential day processing.

Verify RAM consumption of processes by measuring system memory before/after killing them. Uses htop in a tmux session for accurate measurement.

Debug and start Agent HQ services (AMS, UI) on remote machines. Use when AMS not responding, port 8041 issues, environment variable errors, vite config failures, or SSE connection lost errors.

Adds project root directories to sys.path for proper imports, with try/except to handle REPL environments (PyCharm, ipython, Jupyter) where __file__ is undefined. Use when writing Python scripts that need to import from parent/grandparent directories and should also work in interactive environments.

Create or tune a Rust SQLite ingestor that ingests huge JSONL logs fast in historical mode and stays under a strict RSS budget in live mode using byte-scan parsing, mode-specific SQLite PRAGMAs, exec-based live restart, and a minimal index set.

Kill a TCP port safely by targeting only the LISTEN process (the server) and never killing connected clients. This skill should be used when asked to "kill a port", "free port", "stop what's running on port X", or when writing code/scripts to do so.

Search past agent conversations, chats, prompt history, and agent data (Codex/Claude). Use when user asks to "search agent chat", "search agent conversation", "find conversation", "agent prompt history", "what did agent X do", "query agent data", or needs to look up past agent activity. PREREQUISITE - Always read dump-agent-data-for-rg skill first.