rigswig: what your coding agent loads

A coding agent’s harness is everything around the model: instruction files, subagents, skills, commands, hooks, MCP servers, plugins, permissions and the model setting. They don’t live in one place. Your home folder, the repo, a local file, and whatever your organization manages.

Each tool decides in its own way which of those files wins: Anthropic’s docs set out Claude Code’s order, and OpenAI’s docs set out how Codex combines its files. I couldn’t see what a session in one of my repos actually got. So I built something that reads it.

The README for rigswig says it in one line: “See everything your coding agent loads before you type”, and where each piece came from. It’s one binary, written in Rust, under the Apache-2.0 licence. The code is on GitHub.

The headline is one bar. It’s the context a session spends before your first message, split by kind, then by where each piece came from. Per rigswig’s README, it shows which scope decided each setting, what got overridden, and roughly how much context it costs.

Roughly, because no tokenizer runs. rigswig counts characters, divides by four and rounds up. The result is labelled an estimate, with a range of three to five characters per token.

It also says what it can’t see. The system prompt, the built-in tools and each MCP server’s tool list don’t come from a file it can read, so it names them as not measured.

The figure below is the project’s own invented demo setup. Not mine. Turn a layer or a kind off and the total moves as it would under rigswig’s what-if option, --without.

Set the same key in more than one place and, per Anthropic’s docs, Claude Code takes it from the highest level that sets it. The order runs managed settings, the command line, project local, shared project, then user.

Not everything works that way. Anthropic’s docs say list keys such as permissions.allow are combined across files. Environment variables, they say, aren’t a level at all.

rigswig keeps the trail for each setting: which layer set it, and which ones lost. An item overridden by one of the same name higher up gets marked that way. It isn’t counted twice.

The same repo can give two tools very different sessions. OpenAI’s docs say Codex stops adding instruction files once their combined size reaches a limit. The default is 32 KiB. They suggest raising it or splitting the files.

Anthropic’s docs say Claude Code reads AGENTS.md in only two cases: when there’s no CLAUDE.md in the working directory or above it, or when a CLAUDE.md imports it. The demo repo’s CLAUDE.md imports it.

rigswig compare puts both on one scale: one repo under two harnesses, or two repos. The README lists nine harnesses it models. Each gets a page of the rules it follows and where they came from.

rigswig audit ranks what to trim. On the list: agents and skills in your home folder that nothing in the repo refers to, plugins the repo doesn’t call for, instruction files over budget, sentences repeated across layers, paths the instructions name that don’t exist, and permission rules that have piled up.

Each finding carries its evidence, the tokens it would save and a proposed change. Fleet mode runs across many repos and says which ones need each item, so the item can move out of your home folder. Into the repos that use it.

The README for rigswig lists these as the point of the tool. No network request, unless you ask it to read a public GitHub repo. No model calls. Nothing written unless you start the page with --allow-writes.

It never prints the values in an MCP server’s environment or headers, and never reads session transcripts. Any folder you list as private, it refuses.

Its page listens only on your own machine, behind a random token. It loads nothing from anywhere else.

It’s on crates.io and on npm. Install from npm and you get the prebuilt binary for your platform. The releases page has downloads for macOS, Linux and Windows too, each with a checksum to check it against.

Every release goes here, newest first, each with the notes from its GitHub release.

1. v0.1.22026-11-06Latest
   • Five more harnesses: Devin CLI, Cursor, Gemini CLI, DeepSeek Harness and DeepSeek-CLI.
   • rigswig status now reads nine harnesses, each with a page of the rules it models.
   • Config files with comments (JSONC) are read too.
2. v0.1.12026-11-06
   • On npm: npx rigswig installs the prebuilt binary for your platform.
   • On crates.io: cargo install rigswig.
   • No change to what rigswig reads or shows.
3. v0.1.02026-11-06
   • The first release: status, audit and the local page for Claude Code, Codex, Hermes Agent and Ollama.
   • Binaries for macOS, Linux and Windows, each with a SHA256SUMS line to check.
   • The JSON output may change in a minor release until 1.0.

It’s pre-1.0. The release notes say the JSON output may change in a minor release until 1.0. They also say the binaries aren’t signed or notarized yet.

The token counts are estimates from characters, not from the model’s tokenizer. Each adapter is rigswig’s reading of a maker’s docs. Docs change. Every adapter lists what it doesn’t model.

It isn’t the only tool on this question. The README credits the others: harness-map, whatloads, claude-config-auditor, and Claude Code’s own /context command.