Skip to content

Releases

ctxray 0.1.0

Choose a tag to compare

@github-actions github-actions released this 16 Nov 21:11

The first release. ctxray shows what your AI agents load into a model's context and tests it in CI.

  • πŸ” ctxray inspect shows what loads here. Rules, skills, commands, and MCP servers of Claude Code, Codex, Gemini CLI, OpenCode, and Cursor CLI, each with its tokens and the reason it loaded: always, for a file path you pass with --files, or by its description. Claude Code Desktop and the Cursor editor read the same files, ask for them with --harness. A warning names what one agent sees and another doesn't, like a CLAUDE.md Codex never reads.
  • πŸͺ™ Tokens counted each agent's way. Claude's count comes from o200k with factors measured on claude-sonnet-5-5, Gemini's from the Gemma vocabulary, the others' from o200k. Every local count carries β‰ˆ. With exactTokens, ctxray asks the provider for the exact number through your key.
  • βœ… Tests in seven layers. *.ctxray.ts files hold test.resolve, test.assemble, test.wire, test.activate, and test.scenario, from what the files predict (L3) to a task the agent solves end to end (L7). ctxray runs L1 Lint and L2 Policy itself. ctxray init writes a config and a first test that passes.
  • πŸ§ͺ One test, every agent. A test runs once per harness in the config, and expect.on narrows a check to some of them. discover.rules() turns one test into one per rule, by CODEOWNERS owner, folder, or activation.
  • 🎯 Does the model pick the right skill? test.activate sends prompts to a real model several times and checks the share of runs that loaded the skill, with a 95% Wilson interval. By default a skill must load in 80% of the runs, and a test makes the 16 runs it takes to prove that. The config's activate sets both for your project. Too few runs to tell gives inconclusive, not a guess. --record keeps the model's answers, and --backend replay plays them back for free on every merge request. ctxray gen-prompts writes prompts from the descriptions for you to review.
  • πŸ›°οΈ What the agent really sent. L4 and L5 launch the agent on a copy of your project behind ctxray's proxy, with an isolated home, no network but the proxy, and a mock model: free and the same every run. You see the system prompt, the built-in tools, and every MCP tool with its schema's tokens.
  • πŸ“¦ Pinned agents. ctxray harness add pins versions in ctxray.lock with integrity for every platform, and CI installs exactly those from npm or your registry. harness upgrade shows how a new version changes the context. harness pack and install --from carry them into a network without internet.
  • πŸ“‹ Your organisation's rules. Policies limit tokens per component, ban or require MCP servers, and require files. extends shares them from an npm package. ctxray baseline lets an old repository adopt ctxray and fail only on new problems. 27 lint checks catch broken frontmatter, dead rules, missing imports, and skills an agent never sees.
  • πŸ“‰ What a merge request changes. ctxray diff compares the loadout with origin/master or the --base you name.
  • πŸ‘€ Your own sessions. ctxray spectate install adds hooks to Claude Code and Cursor, and ctxray spectate shows what your latest sessions loaded. Prompts never get stored.
  • 🧩 A plugin for Claude Code. /ctxray shows what the current session loaded, /ctxray-test runs the tests, and the write-tests skill writes them. ctxray mcp serves inspect, run, and diff to any agent.
  • πŸ“Š Reports. A live view in the terminal, an HTML report in English and Russian that ctxray show-report keeps live while tests run, JUnit, Allure, JSON, and a Markdown summary that lands in the GitHub job summary.
  • 🏒 Closed networks. LLM gateways, proxies, your CA certificates, internal npm registries with .npmrc auth, read-only shared harness caches, and an offline switch. ctxray doctor checks all of it first.
  • πŸ”’ Secrets stay out. Keys, tokens, emails, phone and card numbers are hidden before anything reaches a trace, a report, a snapshot, or a cassette, with gitleaks' rules on top of ctxray's own.
  • 🚚 Install it your way. npm, Homebrew, Scoop, install.sh and install.ps1, Docker images on ghcr.io, a GitHub Action, a GitLab CI template, and release archives that work as offline bundles. Executables for Linux (glibc and musl, x64 and arm64, and x64 without AVX2), macOS, and Windows on x64 and arm64. Checksums are signed with Sigstore and the files carry build attestations and an SBOM.
  • 🌍 English and Russian. ctxray follows your system's language, or --lang.
npm install --save-dev @mairwunnx/ctxray@0.1.0
curl -fsSL https://github.com/MairwunNx/ctxray/releases/download/v0.1.0/install.sh | CTXRAY_VERSION=0.1.0 sh
brew install mairwunnx/tap/ctxray
docker pull ghcr.io/mairwunnx/ctxray:0.1.0