DEV Community

XenoCyber0
XenoCyber0

Posted on

I Built a GitHub Action That Writes Your PR Descriptions

#ai #githubactions #opensource #typescript

I Built a GitHub Action That Writes PR Descriptions From Real Diffs (Works With Ollama)

Yep, this does what the title says. StandupBot reads the actual PR diff and commit messages, runs them through any OpenAI-compatible LLM endpoint, and writes a structured Summary / Changes / Testing description — so you never have to write "fixed stuff" again.

I built this for my own team because I was tired of staring at empty PR boxes after every fix. Then I put it on GitHub Marketplace so anyone can use it. MIT licensed, no strings.

GitHub logo XenoCyber0 / StandUpBot

StandupBot

License: MIT GitHub Marketplace Open Source Love

A GitHub Action that writes your PR descriptions for you. It runs the real diff and commits through an LLM you choose, and fills in a structured Summary / Changes / Testing description plus up to three labels — so you stop hand-writing PR bodies and standup updates.

Bring your own model: it talks to any OpenAI-compatible endpoint (OpenAI, OpenRouter, Ollama, LM Studio, …). No provider, URL, or model is hardcoded.

Sample output

Generated end-to-end by StandupBot for a real PR (nodejs/node#64573) from its actual diff, using a real LLM endpoint. This is the description it produced verbatim:

Summary

Add lchownSync to the VFS implementation so that symbolic link ownership can be changed without following the link, matching the behavior of fs.lchownSync.

Changes

  • doc/api/vfs.md: added lchownSync(path, uid, gid) to the list of VFS API signatures.
  • lib/internal/vfs/file_system.js: added synchronous lchownSync method to VirtualFileSystem and updated the async…
View on GitHub

The Problem

You open a PR. Forty files changed. Two hours of focused work. The description field is empty.

So you write: fixed stuff. Or updated code. Or changes requested by reviewer.

Three months later, someone (you) is running git log -p trying to figure out why the config format changed in April. The PR description was supposed to save that investigation. Nobody wrote it.

We tried before:

Approach Why it failed
PR templates Everyone types "see title" into the template
Checklist bots Nagging doesn't scale; people ignore them
Reviewer enforcement Becomes the team's most hated job

The gap isn't discipline. It's that writing PR descriptions is unrewarded work. You get the same merge button whether you write a detailed description or "fixed stuff."

So I built StandupBot.

What It Produces (Real Example, No Cherry-Picking)

This is verbatim output from nodejs/node#64573, generated from the actual diff:

Summary

Add lchownSync to the VFS implementation so that symbolic link ownership can be changed without following the link, matching the behavior of fs.lchownSync.

Changes

  • doc/api/vfs.md: added lchownSync(path, uid, gid) to VFS API signatures.
  • lib/internal/vfs/file_system.js: added synchronous lchownSync method and updated the async wrapper.
  • lib/internal/vfs/provider.js: added default lchownSync method with JSDoc comment.
  • lib/internal/vfs/providers/memory.js: implemented lchownSync that updates uid/gid of the link entry.
  • lib/internal/vfs/setup.js: changed handler to invoke vfs.lchownSync instead of vfs.chownSync.
  • test/parallel/test-vfs-lchown-symlink.js: new test verifying sync, callback, and promise variants.

Testing

The new test exercises fs.lchownSync, fs.lchown (callback), and fsp.lchown (promise) on symlinks inside a VFS mount, asserting correct uid/gid changes.

Nobody edited that. It went straight from the model into the PR body. The action also returns a normalized title and labels clamped to bug | feature | chore | docs | refactor — so your label taxonomy stays clean.

30-Second Setup

Add .github/workflows/standupbot.yml:

name: StandupBot
on:
  pull_request:
    types: [opened, synchronize]

permissions:
  pull-requests: write
  contents: read

jobs:
  describe:
    runs-on: ubuntu-latest
    steps:
      - uses: XenoCyber0/StandUpBot@v1
        with:
          llm-base-url: ${{ secrets.LLM_BASE_URL }}
          llm-api-key: ${{ secrets.LLM_API_KEY }}
          model: ${{ vars.LLM_MODEL }}
Enter fullscreen mode Exit fullscreen mode

Set LLM_API_KEY as a secret, LLM_MODEL as a variable. That's the whole thing.

BYO Endpoint: No Vendor Lock-In

The action talks to any OpenAI-compatible chat API. Nothing is hardcoded. All of these work:

OpenAI        → https://api.openai.com/v1
OpenRouter    → https://openrouter.ai/api/v1
Ollama        → http://localhost:11434/v1
LM Studio     → http://localhost:1234/v1
LocalAI       → http://localai:8080/v1
You decide.
Enter fullscreen mode Exit fullscreen mode

You pick the model. You control the data path. Switching providers later is one secret update.

The Self-Hosted Setup (My Favorite Part)

If you want your code to never leave your network:

GitHub repo ──webhook──▶ self-hosted Actions runner ──HTTP──▶ Ollama
                             (on your LAN)              (your LLM box)
Enter fullscreen mode Exit fullscreen mode
  1. Run a self-hosted Actions runner on the same LAN as your LLM box
  2. Point llm-base-url at your Ollama/LM Studio instance
  3. The llm-api-key can be any non-empty string — Ollama ignores it

Your diff never crosses the internet. The only outbound call is writing the PR body back to GitHub, which... GitHub already owns.

Model notes: In my testing, Qwen2.5-Coder 7B and similar 7–9B instruction-tuned models work well. Below ~3B params, file-name hallucinations start appearing.

It Never Overwrites You

The rule is simple:

  • PR body empty → StandupBot writes it
  • PR body has StandupBot's marker → updates its own output
  • PR body has anything else → does absolutely nothing

Your hand-written descriptions are safe. If you ever overwrite what it generated, it takes the hint and stays out.

How It Works Internally

Three parts worth stealing:

1. Diff budget. Fetches the PR diff, applies gitignore-style exclusions from .standupbot.yml before anything hits a prompt:

exclude:
  - package-lock.json
  - '**/*.lock'
  - '**/dist/**'
Enter fullscreen mode Exit fullscreen mode

The diff is hard-capped at ~24KB — a monster PR can't blow up your LLM context window.

2. Map-reduce for big diffs. Under the budget, one call. Over it? Each file gets summarized individually (max 8 LLM calls), then merged into the final description.

3. Structured output enforcement. The prompt requires a rigid schema — title, summary, changes[], testing, labels[] — and the parser clamps labels to the allowed set. If the model invents urgent-pls, it gets dropped, not shipped.

Honest Limitations

  • Testing section can be inferred, not evidenced. Sometimes the model writes "built and renders correctly" for a docs-only change because the pattern is tempting. Treat it as a draft. Known issue, prompt tightening planned.
  • Huge PRs lose their tail. Beyond the character cap, the map-reduce path works on summaries — a detail in file #47 may not survive.
  • GitHub reads are rate-limited (~60/hr) without a token, but the default workflow token covers this.

What's Next

  • Prompt tightening for the testing-overclaim problem
  • More testing on genuinely enormous PRs
  • Whatever GitHub issues say — the roadmap is user-adaptive because there are currently about four users

Try It

Marketplace: github.com/marketplace/actions/standupbot-pr-describer
Source: github.com/XenoCyber0/StandUpBot (MIT)
Issues/feedback: github.com/XenoCyber0/StandUpBot/issues

Written by the person who finally read "fixed stuff" one too many times and decided to do something about it.

Top comments (0)