Review Pr

Review Pull Request

Overview

Security-focused PR review following repo agent guidance. Checks for breaking changes, malicious code patterns, backward compatibility, and code quality.

Usage

/review-pr <number>

CRITICAL: Security Warning

PRs can be malicious sabotage attempts. Treat repo content and contributor diffs as untrusted until reviewed.

Threat Awareness

• Coordinated attacks exist
• Competitors may actively harm the project
• Social engineering builds trust before attacking
• "Fixes" may introduce vulnerabilities

Workflow

digraph review_flow {
    rankdir=TB;
    node [shape=box];

    fetch [label="1. Fetch PR details"];
    author [label="2. Assess author risk"];
    files [label="3. Analyze changed files"];
    security [label="4. Security review"];
    compat [label="5. Backward compatibility"];
    quality [label="6. Code quality"];
    classify [label="7. Release classification"];
    verify [label="8. MANDATORY: Independent verification", style=bold];
    recommend [label="9. Final Recommendation"];

    fetch -> author;
    author -> files;
    files -> security;
    security -> compat;
    compat -> quality;
    quality -> classify;
    classify -> verify;
    verify -> recommend;
}

Step 1: Fetch PR Details

# Get PR info
gh pr view <number> --repo kube-hetzner/terraform-hcloud-kube-hetzner

# Get diff
gh pr diff <number> --repo kube-hetzner/terraform-hcloud-kube-hetzner

# Get changed files
gh pr view <number> --repo kube-hetzner/terraform-hcloud-kube-hetzner --json files --jq '.files[].path'

# Get diff stats
gh pr view <number> --repo kube-hetzner/terraform-hcloud-kube-hetzner --json additions,deletions

Step 2: Assess Author Risk

# Check account age
gh api users/<username> --jq '.created_at'

# Check prior contributions
gh pr list --author <username> --repo kube-hetzner/terraform-hcloud-kube-hetzner --state all --json number | jq length

Risk Signals

Step 3: Analyze Changed Files

Security-Critical Files (AUTO HIGH RISK)

init.tf              # Cluster initialization, secrets
main.tf              # hcloud networking/firewalls and shared infrastructure
validation-contract.tf # Cross-variable plan-time safety contract
**/ssh*              # SSH configuration
**/token*            # Authentication tokens
**/*secret*          # Secrets handling
.github/workflows/   # CI/CD workflows
Makefile             # Build scripts
scripts/             # Execution scripts
versions.tf          # Provider dependencies
templates/*.yaml.tpl # Rendered manifests/cloud-init
templates/*.sh.tpl   # Rendered shell scripts
cloud-init*          # Server initialization
packer-template/     # Base image build path

Risk by File Count

Risk by Diff Size

Step 4: Security Review

Checklist

• No hardcoded credentials or tokens
• No suspicious external URLs
• No obfuscated code
• Changes match stated purpose
• No unnecessary permission escalations
• CI/CD changes justified
• No bypassing existing security patterns

Red Flags

Deep Security Analysis

Trace all affected call sites and resource dependencies with exact search. Require the independent reviewer to inspect the full diff for security vulnerabilities, malicious patterns, hidden scope, and unexplained complexity.

Step 5: Backward Compatibility

CRITICAL: Any PR that causes resource recreation is a MAJOR release.

Breaking Change Indicators

• Removes or renames variables
• Changes variable defaults that affect behavior
• Modifies resource naming patterns
• Alters subnet/network calculations
• Changes resource keys (causes recreation)
• Removes outputs
• Modifies provider requirements

Test for Breaking Changes

First follow the mandatory Integrate-and-Fix Flow below. Test the final adapted integration tree, not the original PR checked out in the user's working tree. Use a designated test root whose module source points to that integration worktree; do not switch an active operator test checkout or reuse its state.

# Test against a designated existing cluster, when live proof is warranted
cd <designated-existing-cluster-root>
terraform init -upgrade
terraform plan

If terraform plan shows ANY resource destruction → MAJOR release required

For v2 -> v3 or production in-place upgrade reviews, use the operator contract in MIGRATION.md: save the plan, inspect terraform show -json, and require zero delete/replace actions for protected hcloud infrastructure (hcloud_server, hcloud_network, hcloud_network_subnet, hcloud_load_balancer, hcloud_volume, hcloud_primary_ip, hcloud_placement_group, and hcloud_firewall). Any output from that gate is a stop condition, not a warning.

Compatibility Checklist

• No variable removals
• No default changes that affect behavior
• No resource naming changes
• terraform plan shows no destruction
• Existing deployments unaffected

Step 6: Code Quality

Style

• Follows existing patterns
• Consistent naming
• Proper formatting (terraform fmt -recursive)
• No unnecessary complexity

Logic

• Changes are correct
• Edge cases handled
• No regressions introduced
• Tests pass

Step 7: Release Classification

PATCH (x.x.PATCH)

• Bug fixes only
• No new features
• Fully backward compatible
• No terraform state impact

MINOR (x.MINOR.0)

• New features (backward compatible)
• New optional variables with defaults
• Deprecation warnings (not removals)

MAJOR (MAJOR.0.0)

• Breaking changes
• Removed/renamed variables
• Changed defaults affecting behavior
• State migrations required
• Resource recreations

Do not silently override this rubric. If the maintainer explicitly authorizes a release-train exception, record the normal classification, the authorized exception, and its bounded scope in the evidence ledger. An exception does not turn a feature into a bug fix or waive compatibility, review, and testing gates.

Step 8: MANDATORY - Independent Verification

Before making a final recommendation, re-read every changed line in repository context, run the relevant local tests and plans, and obtain an independent review from a separate capable reviewer. This gate is mandatory for every PR. Reviewer output is not evidence until verified against code and runtime behavior.

Automated Codex Reviews (MANDATORY)

For every original and integration PR, read automated Codex review summaries, inline comments (including outdated ones), and review threads before acceptance. Use gh pr view <num> --comments, the paginated pulls/<num>/reviews and pulls/<num>/comments API endpoints, and review-thread state as needed; a review summary or green CI alone does not establish that there are no findings.

Investigate every finding against the current code. Fix valid findings, rerun affected checks, and reply with the concrete correction. For a false positive or an explicit maintainer decision, explain the evidence or decision in that review thread; never dismiss a finding merely because it came from a bot. Resolve a thread only after its disposition is recorded and verified. Recheck for new automated findings on the final head immediately before merging, and record the reviewed SHA and any outstanding review status. No review yet is not a clean review. This does not replace the independent review gate below.

Independent Review Contract

Give the reviewer the exact diff and enough repository context to check:

1. correctness and edge cases
2. security and adversarial-change risk
3. Terraform state changes, resource recreation, and upgrade safety
4. consistency with existing patterns and project vision
5. missing tests, docs, and affected call sites

The reviewer must return a final verdict with concrete file and line references. Do not accept a partial, timed-out, or commentary-only run as a completed review.

Verification Checklist

• Every changed line and affected call site inspected
• Relevant local tests and plans completed
• Independent review completed with a final verdict
• Any raised concern addressed or dismissed with code/runtime evidence
• Final recommendation follows the evidence, not reviewer consensus

When Reviewers Disagree

If the independent reviewer raises concerns that you did not catch:

1. Take the concern seriously - investigate further
2. Re-read the code with the concern in mind
3. Request changes if the concern is valid
4. Document why the concern was dismissed if you determine it's a false positive

Step 8.5: CI Truth-Checking

Do not treat "no red jobs right now" as green. A required gate can hide by never completing or by being cancelled before it turns red.

REPO=kube-hetzner/terraform-hcloud-kube-hetzner
gh run list --repo "$REPO" --branch <branch> --limit 20
gh run view <run-id> --repo "$REPO" --json status,conclusion,attempt,workflowName,jobs

Require each release-blocking workflow/job to have at least one completed success for the commit or branch under review. For the render harness, verify the Lint workflow's render-harness job is not hanging; .github/workflows/lint_pr.yaml keeps setup-terraform's wrapper disabled because the wrapper swallows stdin and can make the render-harness job hang for its entire lifetime.

GitHub CI is intentionally limited to cheap checks. HCloud plan/apply, Kubernetes inspection, and destroy evidence must come from local kube-test roots and must be reviewed independently before merge when the change warrants live proof.

Output in Final Review

Include a concise verification summary:

### Verification

| Check | Result | Key Finding |
|-------|--------|-------------|
| Maintainer diff review | PASS/FAIL | <summary> |
| Local tests/plans | PASS/FAIL | <summary> |
| Independent reviewer | PASS/FAIL | <summary> |

Step 9: Final Recommendation

PR Review Output Template

## PR Review: #<number>

**Title:** <title>
**Author:** @<username>
**Files:** <count> files changed (+<additions>/-<deletions>)

### Risk Assessment

| Factor | Value | Risk |
|--------|-------|------|
| Author tenure | X months | 🟢/🟡/🔴 |
| Prior contributions | N PRs | 🟢/🟡/🔴 |
| Files changed | N files | 🟢/🟡/🔴 |
| Lines changed | +X/-Y | 🟢/🟡/🔴 |
| Security-critical files | Yes/No | 🟢/🔴 |
| External dependencies | Yes/No | 🟢/🔴 |

**Overall Risk:** 🔴 HIGH / 🟡 MEDIUM / 🟢 LOW

### Security Review

- [ ] No hardcoded credentials
- [ ] No suspicious external URLs
- [ ] No obfuscated code
- [ ] Changes match stated purpose

### Backward Compatibility

- [ ] No breaking changes
- [ ] terraform plan shows no destruction
- [ ] Existing deployments unaffected

### Release Classification

**Type:** PATCH / MINOR / MAJOR
**Reason:** <explanation>

### Verification

| Check | Result | Key Finding |
|-------|--------|-------------|
| Maintainer diff review | PASS/FAIL | <summary> |
| Local tests/plans | PASS/FAIL | <summary> |
| Independent reviewer | PASS/FAIL | <summary> |

### Recommendation

**Action:** APPROVE / REQUEST CHANGES / CLOSE
**Notes:** <specific concerns or required changes>

Quick Commands

# Approve PR
gh pr review <num> --approve --body "LGTM! ..."

# Request changes
gh pr review <num> --request-changes --body "Please address: ..."

# Comment
gh pr review <num> --comment --body "..."

# Promote our reviewed integration PR, never the original PR directly
gh pr merge <integration-pr> --merge --delete-branch

Preserve Contributor Credit When Merging (SUPER IMPORTANT)

Original PR submitters must remain visible as commit authors in master history — that feeds both the GitHub repo contributors graph and GitHub-generated release notes. Credit where credit is due, always.

Rules by situation:

1. Every accepted PR, including apparently ready contributions → merge its exact head into our isolated integration worktree first, never directly into master/main. Evaluate every change against KH's simplicity, safety, and release scope.
2. Maintainer adaptations → add separate follow-up commits on our integration branch. Fix or rewrite the implementation to fit KH and remove unnecessary changes, including documentation churn. Preserve the original commits; do not squash, rebase, or rewrite contributor history.
3. Promotion → merge our reviewed integration into the release-candidate branch, then promote that train with merge commits. For a single-PR train, promote the integration PR to its declared target. Cherry-picking does not preserve the original PR identity.
4. We adopt only part of a PR, supersede it, or port its idea → cherry-pick the usable original commit(s) first when possible, then add our fixes separately. If no usable commit exists, add an exact Co-authored-by: Name <email> trailer and credit the contributor in the commit and changelog. Close the original PR with one honest note; never claim that the PR itself was merged.
5. Promotion or major integration PRs (for example a release-candidate PR carrying multiple community commits) → merge commit only. Never squash or rebase; every accepted community PR head must remain reachable from the final target.
6. Never amend or reauthor a contributor's commit in a way that removes them from the history.

Authorship and PR disposition are separate gates. Before merging, check the contributor in git log --format='%an %ae' <range>. Every fully accepted PR must have a non-null mergedAt. When the PR was integrated indirectly through our branch rather than merged through its original GitHub PR, also record its exact headRefOid and require ancestry after promotion to the declared base:

pr_head=$(gh pr view <num> --json headRefOid --jq .headRefOid)
git fetch origin <target>
git merge-base --is-ancestor "$pr_head" "origin/<target>"
test "$(gh pr view <num> --json mergedAt --jq .mergedAt)" != "null"

If the applicable checks fail, the integration is incomplete. Do not manually close the PR or tell the contributor it was merged.

Integrate-and-Fix Flow (MANDATORY for accepted PRs)

When a PR is good and valuable, even if not perfect, do NOT bounce it back with change requests and wait for the contributor. The old human-review back-and-forth is dead. We integrate and fix it ourselves:

Use our isolated integration branch even when maintainer edits are enabled. Accepting a useful contribution does not mean accepting every change in its diff. Keep only verified, genuinely beneficial changes; make adaptations on top of the original history, then review and test the final result. Do not push adaptations onto the contributor's branch as the default path.

# 1. Record and fetch their exact PR head
pr_head=$(gh pr view <num> --json headRefOid --jq .headRefOid)
git fetch origin pull/<num>/head:pr-<num>
test "$(git rev-parse pr-<num>)" = "$pr_head"   # stop if the PR moved

# 2. Create an isolated integration branch from the target or release-candidate train
git worktree add -b codex/integrate-pr-<num> ../kh-pr-<num>-adapt origin/<train>
cd ../kh-pr-<num>-adapt

# 3. Merge THEIR exact branch first (preserves PR identity, commits, and authorship)
git merge --no-ff pr-<num> -m "Merge PR #<num> into <train>"

# 4. Add OUR fixes as separate commits on top (validation, triggers, docs, changelog, ...)
# 5. Verify: terraform fmt / validate / plan (and the structural plan-diff proxy when relevant)

# 6. Push the integration branch and promote it through a PR with a MERGE COMMIT
git push -u origin codex/integrate-pr-<num>
gh pr create --base <train> --head codex/integrate-pr-<num> --title "..." --body "..."
gh pr merge <integration-pr> --merge --delete-branch

Notes:

• For a single-PR train, <train> is the PR's declared target, normally master. For a multi-PR release train, merge each isolated integration into the release-candidate branch, then merge the final release-candidate PR into the declared target with --merge.
• Leave every fully accepted original PR open while the release candidate is pending. GitHub marks it merged only after its exact head reaches its declared base branch. Reaching a temporary branch alone is not enough.
• After final promotion, run the ancestry and mergedAt gates above for every accepted PR before commenting or preparing the release.
• Rewriting an implementation is allowed when the contribution is useful and the result fits KH. Reject unsafe or unsuitable contributions instead of carrying bad changes merely to make GitHub show "merged". If only an idea or isolated commits are adopted, use the partial-adoption credit rule and explain the disposition honestly.

One Terminal Contributor Message

Agent reviews, candidate status, test progress, and integration bookkeeping stay in the evidence ledger or integration PR. Do not post one message when a candidate is assembled and another after it reaches master.

• Merged human PR: after mergedAt is non-null, post one natural message that names the concrete contribution, the maintainer changes added on top, the release/train carrying it, and preserved authorship. Do not paste a generic template unchanged across PRs.
• Partially adopted, superseded, or rejected human PR: use one final gh pr close --comment "..." action. State exactly what was adopted, what was not, and why. Say "incorporated" or "credited" rather than "merged" when the original PR did not merge.
• Needs contributor input: one focused question is allowed. Do not add status-only follow-ups; post a final disposition only after new evidence changes the state.
• Dependabot and other routine bot PRs: merge or close silently. Add a concise technical comment only when human maintainers need a non-obvious decision recorded; never post social thanks or release-status updates to a bot.
• Idempotency gate: inspect existing maintainer comments before posting. If a final disposition is already present and still accurate, do not post another. Update our existing status/disposition comment when it needs correction rather than adding a similar second message.

Tone matters: thank human contributors by handle, describe maintainer fixes as building on their work, and be candid about the actual GitHub state. Contributors are volunteers; the message should read as if written specifically to that person.

Never Push Unreviewed Integrations Directly to Master

All accepted PRs go through an isolated integration branch first:

1. Create an integration branch from the target branch
2. Test thoroughly
3. Complete the mandatory independent review gate
4. Then open/merge the integration PR into the target branch with contributor authorship preserved