Reviewing Nominal API Bumps

Reviewing nominal API version bumps

Compare two published versions of the generated Nominal bindings and report what changed and what it costs nominal-client to migrate. Works for both transports:

• conjure → nominal-api (the Conjure HTTP bindings; one big nominal_api/_impl.py)
• proto → nominal-api-protos (gRPC/protobuf; nominal/protos/**/*_pb2.pyi + *_pb2_grpc.py)

The output classifies every delta as ⚠️ Breaking, ➕ Additive, or ℹ️ Behavioral, and — critically — scans the client so a breaking change with 0 references is flagged as a safe bump, while an additive with 0 references is flagged as a new capability that isn't wired up yet.

The default deliverable is a self-contained HTML report served on localhost (http://localhost:<port>/…); the same content is also written as markdown.

When to use

• Bumping the nominal-api / nominal-api-protos pin in pyproject.toml.
• Answering "what changed between 0.1282.0 and 0.1286.0?" or "is this bump safe to take?"
• Deciding what migration work a dependency upgrade implies before doing it.

Not for: editing the client to perform the migration (this skill only reports), or non-Nominal deps.

Quick reference

SKILL=.agents/skills/reviewing-nominal-api-bumps

# One-shot: pinned-in-pyproject -> latest on PyPI, scanning this repo
uv run --no-project python $SKILL/bump_scan.py conjure --repo .
uv run --no-project python $SKILL/bump_scan.py proto   --repo .
uv run --no-project python $SKILL/bump_scan.py both    --repo . --out ./reports

# Explicit versions
uv run --no-project python $SKILL/bump_scan.py conjure --repo . --old 0.1282.0 --new 0.1286.0

# Serve the built HTML on localhost (run in the background; prints http://localhost:<port>/…)
uv run --no-project python $SKILL/serve_report.py ./reports --file bump-report-0.1282.0-to-0.1286.0.html

bump_scan.py resolves versions (old = pin in pyproject.toml, new = latest on PyPI), installs both versions, extracts each surface, diffs them, and scans nominal/ + tests/ for references. It writes to --out (default .):

• bump-report-<old>-to-<new>.html — a self-contained styled report (combined across transports for both); the default deliverable, served on localhost by serve_report.py.
• migration-report-<transport>-<old>-to-<new>.md — the same content as markdown (greppable; the inline summary), and printed to stdout.

Then serve it: python serve_report.py <out> --file <name> binds 127.0.0.1 on a free port and prints http://localhost:<port>/…. bump_scan.py prints the exact serve command to run.

How to run: use uv run --no-project python …. --no-project makes uv ignore the project, so your .venv is never read, synced, or modified — whereas plain uv run would sync it to the lockfile first. Bindings are fetched with uv pip install --target into a temp dir with the interpreter pinned and run outside the project, so .venv is untouched regardless. Bare python … works too (stdlib-only), and uv run --no-sync python … reuses the project interpreter without a sync. Only uv need be on PATH.

Workflow

1. Run the scan. Use bump_scan.py <transport> --repo <client-root> --out <dir>. Prefer a scratch/temp --out dir so reports don't dirty the repo. For a full bump, run both (one combined HTML page).
2. Serve it. Launch serve_report.py <out> --file <name> in the background (it runs until killed) and give the user the printed http://localhost:<port>/… URL — this is the default way to view the report. bump_scan.py prints the exact command. (The HTML is self-contained, so file:// works too.)
3. Read the verdict line. ✅ No breaking change references the client ⇒ the pin bump is safe on its own. ⚠️ N breaking change(s) touch the client ⇒ open the Breaking table and address each.
4. Work the Breaking table. For each row with non-zero Client refs, open the listed file: paths and update call sites for the exact change in the Change column (e.g. ~ inputs: List→Dict[str,…], − fill_strategy + alignment_configuration (req)).
5. Skim Additive for opportunities. Additive rows with non-zero refs mean the client already touches that type and could adopt the new optional field/endpoint. Rows with 0 refs are net-new capabilities — adopt only if there's a reason.
6. Behavioral rows are enum members added to existing enums — revisit any exhaustive match/if over that enum in the client.
7. Relay the localhost URL + the verdict + the Breaking table inline.

Interpreting the report

• Client refs are textual matches of the element's leaf name in nominal/ + tests/. They are candidates, not proof: common names (File, Error, State) over-match unrelated code — verify the import context before trusting a count. A field-level change (e.g. a removed field) is scanned by the enclosing type's name, so also grep the field name when triaging.
• Rows whose key contains Internal are internal/server-side services, sorted last; they rarely appear in the public client.
• Visitor classes are skipped — a change to a union shows up on the union row itself.
• Proto stub-generator boundaries. A proto diff that straddles a mypy-protobuf generator change (e.g. nominal-api-protos 0.1286 → 0.1287, which moved to keyword-only __init__ and a different annotation vocabulary) shows widespread Change churn from type-string differences alone (_Optional[str] vs _builtins.str) even where the proto field is unchanged. Added/removed elements, enum members, and field names stay accurate; treat the bulk of type-only deltas at such a boundary as noise. Diffs within one generator era are clean.

The tools (run individually if needed)

bump_scan.py calls these in-process; run them standalone only for a custom flow.

• extract_surface.py {conjure|proto} <unpacked-wheel-root> → the structured surface as JSON Lines (one element per line, sorted by key). Line-diffable and greppable; types kept fully qualified so List→Dict is visible. Importable as extract(transport, root) -> list[Element].
• diff_surface.py <old.jsonl> <new.jsonl> --label "..." [--repo <root>] → the markdown report. Classifies field-level deltas (removed field / new required field / type change / optional↔required) and renders the three severity tables. Importable as diff_and_scan(old, new, repo), render_markdown(...), build_report(...).
• render_html.py → render_page(reports) builds the combined self-contained HTML page from the diffed data (styling in report.css). serve_report.py <dir> serves it on localhost.

Common mistakes

• Trusting a ref count for a generic name. Error/File/State over-match. Open the file.
• Treating an additive optional field as breaking. New Optional[...] fields and new endpoints never break existing callers — they live in the Additive table for a reason.
• Default new needs network. It's resolved from the PyPI JSON API; if offline or it fails, pass --new explicitly. uv must be on PATH to fetch the bindings.
• Running proto against the conjure package (or vice-versa). The transport selects both the PyPI package and the extractor; don't mix them.