# agentprof

> AI Agent Workspace Optimizer & Shell Latency Profiler

agentprof is a Rust CLI and TUI that measures where AI coding agents (Claude Code,
OpenCode, Cursor, Codex, Grok, Aider) lose time and context: subshell startup
latency, MCP tool-schema token load, instruction-file bloat, and skill trigger
collisions.

Design principle: every figure is measured, never guessed. Values that have not
been measured are reported as absent rather than filled in with a placeholder.

## Installation
- Homebrew: `brew tap dautovri/tap && brew install agentprof`
- Cargo: `cargo install --git https://github.com/dautovri/agentprof.git`
- Script: `curl -fsSL https://raw.githubusercontent.com/dautovri/agentprof/main/install.sh | bash`
- Repository: https://github.com/dautovri/agentprof
- Documentation: https://dautovri.github.io/agentprof/

## Commands
- `agentprof scan`: Unified audit (context + subshell + MCP + skills + secrets).
- `agentprof tui`: Interactive dashboard. Requires a real terminal.
- `agentprof context`: Token cost of instruction files (AGENTS.md, CLAUDE.md,
  .cursorrules, .windsurfrules, copilot-instructions.md, GEMINI.md, .clinerules,
  CONVENTIONS.md, CODEX.md). Honours .gitignore.
- `agentprof mcp [--probe]`: Lists configured MCP servers. `--probe` performs a
  real MCP handshake (initialize -> tools/list) against each server and counts
  the exact tokens its tool definitions add per turn. Results are cached.
  Servers never probed report no token figure rather than an estimate.
- `agentprof skills`: Audits installed skills, token weight, and trigger
  collisions. Deduplicates skills reachable through multiple roots.
- `agentprof agent`: Profiles OpenCode, Claude Code, and Grok configurations.
- `agentprof omz`: Times Oh My Zsh plugins and slow init hooks by sourcing them
  in an isolated shell. Hooks that cannot be timed are reported as not measured.
- `agentprof bench [--iterations N]`: Median subshell spawn time after warm-up,
  for the shell named by $SHELL. Reports the avoidable interactive-vs-non-interactive tax.
- `agentprof lint`: Vague rules, duplicated directives, hedged absolutes, and
  cross-file convention contradictions.
- `agentprof compress <file> [--overwrite]`: Compresses instruction files. Keeps a backup.
- `agentprof wrap "<cmd>"`: Runs an agent with AGENTPROF_FAST_PATH=1 exported and
  propagates its exit code.
- `agentprof report [--markdown] [--fail-under N]`: 0-100 health score across five
  equally weighted categories. `--fail-under` exits non-zero for CI gating.
- `agentprof history [--sessions N]`: Real token usage and cost parsed from Claude
  Code transcripts, using cache-aware pricing. `--sessions 0` scans all.
- `agentprof fix [--ignore] [--shell] [--zcompile] [--all] [--dry-run]`: Merges
  ignore patterns without discarding existing rules (backups kept). `--shell`
  installs an opt-in guard that activates only when AGENTPROF_FAST_PATH=1.
- `agentprof compile`: Splits monolithic rule files into modular JIT instructions.
- `agentprof ci [--min-score N] [--force]`: Generates a GitHub Actions workflow
  that fails the build below the score threshold. Never overwrites without --force.
- `agentprof completions <shell>`: zsh, bash, fish, or powershell completions.

## Global flags
- `--json`: Machine-readable output on stdout, with no human-facing banners.
- `--path <DIR>` or a positional path: target directory.

## Cost model
Claude Sonnet list pricing: $3/Mtok input, $15/Mtok output. Cache writes bill at
1.25x and cache reads at 0.1x the input rate.