Upgrade Preflight

• The automated CI integration test boots one network. The two-network path runs only in the manual Real cross-version diff workflow.
• The example contracts are tiny; their cost figures are RPC simulation estimates and should not be extrapolated.
• The roughly 6% to 9% instruction drop from 27 to 28 has not been traced to a cause.
• The GitHub Action wrapper is not yet verified end to end.
• Requires Docker. Networks use stellar/quickstart:latest with --protocol-version.

A CLI + GitHub Action that tells Soroban developers whether their smart contracts behave the same, cost the same, or break after a Stellar protocol upgrade — by actually running both protocol versions on real local networks and diffing the results. Node operators have upgrade tooling built around them; contract developers get a short Testnet window before each Mainnet vote and no easy way to answer "will my contract break or get more expensive on the new protocol?" This does that check, deterministically, with no AI/LLM involved anywhere.

See a real result without Docker: the Upgrade Preflight playground shows the real recorded 27 to 28 and 28 to 29 reports and re-runs this tool's diff engine on them in your browser.

How it works

1. You describe a set of contract-call scenarios in preflight.config.yml.
2. upgrade-preflight run --from 27 --to 28 boots two real, local stellar/quickstart networks — one per protocol version — deploys your contracts, and runs every scenario against both.
3. It diffs the results and reports one of five verdicts per scenario (and overall): SAME, COSTS_CHANGED, BEHAVIOR_CHANGED, BROKE, or ERROR (a tool/network failure, never a statement about your contract).

See docs/ARCHITECTURE.md for exactly what each verdict means and two non-obvious rules (an exactly-at-threshold cost change is SAME; two versions failing identically is SAME, not BROKE).

Quickstart

Requires Docker.

git clone https://github.com/stellarbrief/upgrade-preflight.git
cd upgrade-preflight
npm install
npm run build
node dist/cli/index.js run --from 27 --to 28

This runs the bundled preflight.config.yml and its four example contracts (see examples/) against protocol 27 and 28.

What has and hasn't been verified

• The diff engine (verdict rules, thresholds, bigint-safe comparison) is unit-tested.
• The automated CI integration test boots one stellar/quickstart network at protocol 27 (on a pinned image tag, see below), deploys the example contracts, runs every scenario, and checks resource data comes back. It does not compare two protocol versions.
• The full two-network comparison (run --from X --to Y) is exercised by the manual Real cross-version diff workflow (.github/workflows/real-diff.yml), which uploads the Markdown and JSON reports as artifacts. Before each report the tool checks that the RPC's getNetwork().protocolVersion (documented as the protocol of the latest ledger) matches the requested protocol, and fails if it doesn't.

A real captured result: protocol 27 → 28

From the workflow's runs on the bundled preflight.config.yml and its four example contracts. Two independent runs produced identical numbers. The full report and where it came from are in docs/samples/.

Return values, emitted events, disk reads, writes and footprint entries were identical across both protocols; the minimum resource fee moved by 0.1% to 0.2%.

What this does and doesn't show:

• These are tiny example contracts, and the figures are simulateTransaction estimates from the RPC, not on-ledger metered costs. Don't extrapolate them to real contracts.
• The drop is roughly constant across very different scenarios, which suggests a change in fixed per-call overhead rather than in contract logic. That is an interpretation; this project has not traced it to a specific protocol change.
• The automated CI integration test still boots a single network. The two-network path runs only in the manual workflow above.

A second real result: protocol 28 → 29

Taken the day after protocol 29's stellar-core release, on the then-current stellar/quickstart:latest. Two runs were identical. Details are in docs/samples/.

All verdicts are SAME because the changes are under the default 5% threshold; behavior, events and every other resource are identical. The protocol 28 figures here match the protocol 28 figures in the 27 → 28 table exactly, even though they came from different image builds.

The quickstart image decides what you can test

The core version inside a stellar/quickstart image determines which protocols it can run, and latest moves. On 2026-10-06 latest was republished; after that, protocol 27 stopped running on it (the first contract upload failed with a host error), while 28 and 29 ran fine. The 27 → 28 result above was taken on the image that latest pointed to before that (v672-b1475.1-latest). The cause inside the image has not been investigated.

• Pass --image <ref> (CLI) or the image input (Action and workflow) to choose the image. The report records it.
• The CI integration test pins a tag so it doesn't break when latest moves.
• For reproducible results, pin a tag from Docker Hub's list; for a new protocol, use a newer one.

How verdicts work, and their limits

Every report states exactly how many scenarios ran and includes an explicit "untested code paths are not covered" disclaimer — upgrade-preflight only tells you about the specific contract calls you gave it scenarios for. See docs/ARCHITECTURE.md.

Writing your own scenarios

See docs/WRITING_SCENARIOS.md.

Using it in CI

See docs/CI_USAGE.md for the GitHub Action and exit codes.

Roadmap

See ISSUES_BACKLOG.md for scoped, ready-to-pick-up issues, and docs/ARCHITECTURE.md for the fuller design writeup.

Contributing

See CONTRIBUTING.md for setup, the PR flow, and how issues are rated.

License

MIT — see LICENSE.