oil-diff (oil-oil/oil-diff)Partly tested🇨🇳 CN

@I_am_oil_oil's open-source oil-diff turns a Git Diff of dozens of AI-edited files into an HTML report of 3–7 business changes that a non-coder can read

oil (@I_am_oil_oil, GitHub oil-oil) open-sourced oil-diff (MIT, 0.1.0), a skill for Claude Code / Codex and similar agents. When an agent touches dozens of files in minutes, raw Git Diff is hard to skim and a free-form HTML write-up from the model often fills with fluff. oil-diff collects the diff with a script, groups it into 3–7 business items, and emits a single-file `.oil-diff/index.html`: a Bento board, before/after diagrams, 🎯 must-read vs 📎 peripheral files, and a side code drawer. Output stays in an untracked `.oil-diff/` folder (built-in `*` gitignore) so the repo status is unchanged. We cloned commit 8f4b40c and ran the author's `python3 -m unittest` suite: 5/5 passed (collect→render→lint, Chinese paths and binaries, blocking scavenged fake screenshots, hype-word lint). We did not have an agent produce a full report on a real product repo.

Visit site oil-diff (oil-oil/oil-diff) ↗
Pricing
Free
Region
Global
Where it's from
Chinese community
Added
Nov 2, 2026
Tested on
Nov 2, 2026
Version we used
oil-oil/oil-diff commit 8f4b40c (SKILL 0.1.0)
Categories
CodingProductivity
Original author
oil @I_am_oil_oil ↗
Our test
Partly tested
Cost / time
Skill is MIT and free; only local Python and git. The agent's write-up uses your existing Claude / Codex plan—no extra API for the report itself.
Source
X @I_am_oil_oil ↗GitHub oil-oil/oil-diff ↗

The task

After an agent changes many files, explain to yourself, teammates or a non-technical lead which business areas changed, why, and which files matter—without reading the Diff in alphabetical order.

Steps

  1. Needs Python 3.10+ and system `git`, plus an agent that can install skills (Claude Code / Codex, etc.).
  2. Install with `npx skills add oil-oil/oil-diff`, or tell the agent to install https://github.com/oil-oil/oil-diff.
  3. Ask in chat with a scope, e.g. "use oil-diff on the latest changes on this branch", "uncommitted edits", or "compare main..HEAD".
  4. The skill runs `scripts/collect_diff.py` into `.oil-diff/assets/diff-data.json`, then writes a `report-spec.json` of 3–7 business items (overview, points, must-read files).
  5. Without real run screenshots it uses built-in vector diagram modes (routing / flow / schema…); `screenshots` is only allowed for live captures—lint hashes block images scavenged elsewhere in the repo.
  6. Then `render_report.py` builds `.oil-diff/index.html` and `lint_report.py` checks structure and safety bounds; open the single-file page in a browser to share.

Prompt

Use oil-diff to turn my uncommitted changes into an HTML report: group by business items, use built-in vector diagrams if there are no live screenshots, and do not touch tracked files.

What the author says

@I_am_oil_oil: when AI edits dozens of files, raw Diff is scattered and a free-form HTML write-up fills with text; oil-diff aims at "a boss who didn't write this code can get it in 30 seconds"—group into business items, add routing/control before-after diagrams, open must-read files in a side Diff drawer; built-in collect scripts, diagram rules and templates keep structure stable. Output lives in `.oil-diff/`, invisible to Git.

Our test

On a Linux box we cloned oil-oil/oil-diff commit 8f4b40c (SKILL metadata 0.1.0), AST-checked all `scripts/*.py`, and ran `python3 -m unittest tests.test_oil_diff -v`. We did not install Claude Code / Codex or generate a stakeholder report on a real product repo.

What we found when we tried it:All 5 tests passed in ~1.2 s, covering Chinese filenames and binaries in a temp repo, the full collect→render→lint pipeline, lint blocking scavenged fake screenshots and missing captureSource, structured vs plain-text diagram lint, and hype-word checks. So the claimed antifraud and structure rules have automated tests—but we did not accept an end-to-end agent-written report.

How to get in

Author post and README are in Chinese; GitHub repo is public MIT. Clone works worldwide; Claude Code / Codex still need your own access.

Who it suits

Developers whose coding agents change many files at once and need a skim-friendly report for themselves, PMs or leads—also useful as a pre-merge self-check.

Pitfalls

  • It does not hunt bugs or commit—it only explains what changed.
  • Business grouping still depends on the model: scripts lock collection and render structure, but chapter quality hinges on the agent reading SKILL.md and the references.
  • The author recommends spawning a sub-agent for the full loop so a huge `diff-data.json` does not blow the main context.
  • In our run, the author's 5 unit tests passed; `collect_diff` on a clean scope wrote 0 file records—no changes means no report body.
  • Early 0.1.0; author says macOS was verified on device, with Windows/Linux covered by static cross-platform checks.

More in this category