code-walkthrough

Turns a diff, a PR, or a part of the codebase you want explained into a page you read top down, from the big picture to the exact lines.

The page above is one /code-walkthrough run on lazygit PR #5702, "Make the side panels configurable": 30 files, +1008 -179. Select text on it and click Ask to question the agent in place; ask it to draw any diagram and it adds one to the page.

/code-walkthrough                          # working tree plus staged changes
/code-walkthrough this branch              # main...HEAD
/code-walkthrough 123                      # that PR
/code-walkthrough HEAD~3..HEAD
/code-walkthrough explain the auth flow    # code as it stands, no change needed

Why

Agents write code faster than anyone can read it now, so reading and responding is the bottleneck, not writing. A unified diff is a bad surface to read on. It hands you 30 files in alphabetical order and leaves you to build the picture yourself, bottom up.

When you explain a change to someone, you say what it does, then the steps it takes, then what it touches, and only then point at lines. code-walkthrough builds the page in that order, so by the time you reach a line of code you already know where it sits.

Top down, one question per level

1. Overview: what changed, and why. Two sentences and a few bullets, no file names. On the example: lazygit's side panels stop being a fixed five and become a list you configure.
2. The PR in one picture, then one per theme. The files grouped into three to five themes. A model draws one picture of the change and one per theme, from real names and numbers in the diff; a docs-only change may get none.
3. Call graph: how far does it reach. The call graph, under "All symbols", resolved by the compiler or a language server, at package and symbol level. That shows the blast radius: which new function is called from everywhere and which is a leaf.
4. Code: the lines, with the reason beside them. Each file gets a one line role and each hunk a note on what the code does. Inside a theme, callers come before callees.
5. Respond: send it back. Select any text and click Ask to question the agent in place. It answers in a thread, and can add or replace a block on the page, such as a sequence diagram; the edit is local and Undo removes it. Threads are read-only by default. A proposed task has Hand to my session, which passes it to your own Claude Code. Copy gh command prints the calls to post your comments as PR review comments. Nothing is posted to the PR unless you do it from the page, and the page is a local file that only gets hosted if you ask.

Changed .md files and images have a Diff | Preview switch that swaps the diff for the rendered file. If you work something out in your head on every review that could live on this page, open an issue.

Every walkthrough in one place

Every walkthrough you build is listed on one dashboard at /walkthrough, grouped by day, with stacked PRs nested. Search it, or paste a PR link, branch range or commit to start a new one. It can be installed as an app.

What a run costs

Live mode is the default. On the example PR a build took about 2 minutes and 149k tokens, and each question you ask adds a few tens of thousands more. Without live mode the skill falls back to parallel Sonnet subagents; the same PR took about 7 minutes and 490k tokens there. The graphs come from parsers and take seconds either way.

Live mode runs through a daemon. By default, after running setup, that's your own Claude Code: whatever account it uses, subscription, API key, or a gateway, and it counts against that same account. Point it at any OpenAI-compatible backend instead and it bills per token against that backend's own API key.

Live mode sends diff hunks, the files the agent reads and your questions to that backend. Only a model on your own machine (Ollama, vLLM) keeps them local. Threads can't edit files unless you trust the repo, from the Threads panel or trust.repos in the config; see skills/code-walkthrough/references/live.md.

Prerequisites

• python3, standard library only, nothing to install
• git
• gh, only to target a PR or post comments
• optional, for the graphs: tree-sitter-language-pack, go, or a language server, depending on the language. See below.
• optional, for live mode: one python3 skills/code-walkthrough/scripts/cw_mcp.py setup run after installing. /code-walkthrough does it for you when live mode is missing. Details in skills/code-walkthrough/references/setup.md.

Language support

Everything built from the diff itself works in any language: the grouping and reading order, the annotated walkthrough, per-line comments and the pictures.

The graphs have to parse the code:

Language	What you need
Python, Go, JavaScript, TypeScript	nothing
Rust, Java, Ruby, Kotlin, Swift, Scala, C, C++, C#, PHP, Lua, Elixir, Julia	pip install tree-sitter-language-pack
Interactive symbol graph	go for Go; pyright-langserver, typescript-language-server or rust-analyzer for Python, TypeScript, Rust

Where a parser is missing the graph sections are skipped and the page says so, nothing else changes.

Install

Claude Code plugin only. Other harnesses are untested.

/plugin marketplace add crisperity/code-walkthrough
/plugin install code-walkthrough@crisperity

The first run of /code-walkthrough sets up live mode. On Linux with systemd that includes a user service that keeps the dashboard at http://127.0.0.1:27183/walkthrough running from login on; python3 skills/code-walkthrough/scripts/cw_mcp.py service uninstall removes it and the daemon goes back to starting on demand.