GitHub pull request → LLM summary → PR comment: one importable n8n workflow, tested end to end

A pull request arrives with the title "fix stuff" and 40 changed files. This workflow reads the diff and posts one comment: what changed, why it matters, the risks it can see, and where the reviewer should look first — updated on every new commit, never duplicated.

One file you import into n8n. No sub-workflows, no database, no GitHub App, no LangChain nodes — core nodes only (Webhook, Crypto, Set, Code, If, HTTP Request, Respond to Webhook).

What makes it safe to point at a real repository:

• Deliveries are verified. HMAC-SHA256 over the raw body, compared in constant time, before anything is fetched or any token is spent. Bad signature → 401, nothing else happens.
• A repository allowlist, so a leaked webhook URL cannot make you summarise someone else's code. Drafts, closed pull requests and bot senders are skipped.
• Secrets in a diff never reach the model — AWS keys, tokens, JWTs, private keys, connection strings and api_key = … assignments are redacted first, and the comment says how many.
• Lockfiles, dist/, vendor/, images and snapshots are never sent, and the diff is capped at a byte budget you set. The comment discloses what was left out.
• One comment per pull request. It is matched by a hidden marker and updated in place; human comments are never touched.
• Nothing is invented. If the model fails or answers something unparseable, the comment says so (or posts nothing, your choice).
• Model output is sanitised before it becomes a GitHub comment: it cannot forge the marker or @mention your whole organisation.

Every one of those claims has an end-to-end test in this repository (see below).

By Fractal Techware. MIT licensed.

What's included

File	What it is
workflows/github-pr-summary.json	The workflow: webhook → HMAC verify → diff → LLM → upserted PR comment
tests/run_tests.py	12 end-to-end tests that run the real workflow inside a pinned n8n container
tests/mock/mock.py	Stdlib mock of the GitHub REST subset (with real Link pagination) and the LLM
tests/lib/stack.py	The Docker harness: imports the shipped JSON, publishes it, signs real deliveries
tests/lib/payloads.py	pull_request webhook payloads and HMAC delivery signing
tests/fixtures/pr-files.json	A realistic changed-file list: source, tests, docs, a lockfile, a binary, a file with no patch
examples/docker-compose.yml	A pinned n8n to try it in
examples/send-test-delivery.sh	Sign and send a delivery yourself — no repository needed
docs/setup.md	Credentials, token permissions, every settings field, costs, troubleshooting

1 workflow · 14 nodes · 12 end-to-end tests · pinned to n8n 2.39.7.

60-second quick start

git clone https://github.com/Fractal-Techware/n8n-github-pr-summary.git
cd n8n-github-pr-summary
./run-tests.sh      # proves it works: pinned n8n + mocked GitHub and LLM, no API keys needed

stack ready in 16s (http://127.0.0.1:5698)
PASS  test_drafts_bots_and_foreign_repos_are_skipped  (3.3s)
PASS  test_human_comments_are_never_touched  (0.6s)
...
12/12 tests passed in 53s

Then, in your own n8n:

1. Import workflows/github-pr-summary.json (Workflows → Import from File).
2. Create three credentials: GitHub webhook secret (Crypto — openssl rand -hex 24), GitHub API token (Header Auth, Authorization: Bearer <token>) and PR summary LLM API key (Header Auth, Authorization: Bearer <llm key>).
3. Edit the Settings node: github.repo_allowlist (it is empty of your repos by default, so nothing is processed until you add them), llm.url and llm.model.
4. Activate, then add a repository webhook pointing at https://<your-n8n>/webhook/ftw-gh-pr-summary, content type application/json, the same secret, and the Pull requests event only.

The GitHub token needs just two repository permissions: Pull requests: read and write and Contents: read-only. Full walkthrough and troubleshooting: docs/setup.md.

What the comment looks like

<!-- ftw-pr-summary -->
### AI pull request summary

**What changed**
- The payment client retries idempotent charges up to three times on 5xx responses
- Every attempt now carries the same Idempotency-Key header
- New unit tests cover the retry and the header

**Why it matters**
Checkout stops failing on transient 503s from the payment provider, without risking a double charge.

**Risk notes**
- Retrying a non-idempotent POST would double charge if the key is ever omitted

**Suggested review focus**
- src/payments/client.py: the retry predicate and the backoff bounds

<sub>commit `d34db33` · 6 files (+1333/-174) · 2 ignored (package-lock.json, payment-flow.png) ·
1 file with no diff (binary or too large) · gpt-5-mini · AI-generated, verify before relying on it.</sub>

(Everything in the <sub> line is computed from the diff, not from the model.)

How it is tested

./run-tests.sh starts one n8n 2.39.7 container and one Python container that pretends to be the GitHub REST API and your LLM endpoint. The workflow JSON is imported exactly as it ships — only the Settings node is overlaid so the two URLs point at the mock — then published, and the tests deliver real, HMAC-signed GitHub webhooks to it and assert on what n8n sent and stored:

Test	Proves
test_pull_request_is_summarised_and_commented	Diff and patches reach the model with the model from Settings; the comment carries the marker, sections, counts and the AI disclosure
test_second_delivery_updates_the_same_comment	A new commit PATCHes the existing comment — one comment, not two
test_human_comments_are_never_touched	Matching is on the hidden marker only
test_invalid_signature_is_rejected_before_any_work	Bad HMAC → 401, no GitHub call, no model call
test_ping_and_unhandled_actions_are_ignored	ping and unhandled actions answer 200 and stop
test_drafts_bots_and_foreign_repos_are_skipped	Drafts, bot senders and repositories outside the allowlist cost nothing
test_ignored_files_are_not_sent_to_the_model	The lockfile and the image stay out of the prompt, and the comment discloses it
test_secrets_in_a_diff_are_redacted_before_the_prompt	An AWS key and an sk-… key added in the diff never reach the model
test_paginated_file_lists_are_followed	A 150-file pull request is fetched across pages and fully counted
test_llm_failure_posts_an_honest_note	LLM 500 → a comment that says so, nothing guessed
test_malformed_and_fenced_answers	Prose is refused; JSON inside a ```json fence is accepted
test_model_output_cannot_forge_the_marker_or_mass_mention	Model text cannot inject a second marker or @everyone

No test reaches the internet and no API key is involved; CI runs the same script (.github/workflows/test.yml).

Want all the workflows?

This repository is the free edition of the n8n GitHub AI Workflows pack — same build, same test standard, the rest of the repository chores:

	Free (this repo)	Starter $19	Pro $49	Studio $99
Importable workflows	1	9	15	18
AI pull request summaries	yes	yes	yes	yes
LLM providers	any OpenAI-compatible endpoint	+ Anthropic, Ollama, OpenAI Responses API (router)	same	same
Auth	personal access token	+ GitHub App, shared webhook verification with replay protection	same	same
Release notes into the draft release, CI failure explanations	–	yes	yes	yes
PR risk labels, issue triage, dependency digest, stale PR nudge, daily digest, error notifier	–	–	yes	yes
Per-repo .github/ftw-ai.yml config, Conventional Commit title status check, weekly security digest, prompt eval kit (16 golden cases), GitHub App + GHES guides	–	–	–	yes
License	MIT	own organization	own organization	client / agency use

See the full pack on Gumroad →

More free, tested building blocks: github.com/Fractal-Techware.

Contributing

Issues and pull requests are welcome — see CONTRIBUTING.md. Every behaviour change needs a test in tests/run_tests.py.

License

MIT © 2026 Fractal Techware SRL