Writing PTE PR descriptions

Process

• Before gh pr create, run the static checks CI runs, from the worktree being pushed: pnpm check:format, pnpm check:lint, pnpm check:types, pnpm check:knip (repo root), plus the affected package's test:unit. These are cheap; run them all, every time. A green test run is not a green PR: lint enforces the house assertion style (see the writing-tests skill), and test:unit and check:types typecheck test files where stale API shapes surface.
• Browser tests: complement AGENTS.md's per-package baseline. A contained change may run just its own file (vitest run --project "browser (chromium)" tests/<file>.test.tsx); a change with general blast radius (engine, selection, input pipeline) runs the package's full test:browser:chromium.
• Always open as draft (gh pr create --draft); a maintainer marks it ready and merges. Never gh pr merge.
• Title = the main commit subject (conventional format).
• Use --body-file, never a heredoc with backticks inside $( ): backticks in the body survive intact.
• The repo has no PULL_REQUEST_TEMPLATE.md, so the body is prose, not a form to fill in.
• The description stands on its own: no internal context, ever. No issue-tracker ticket IDs, no customer or partner names, no support threads, no Slack discussions, no "a user reported" or "requested by". Describe the observable failure or need itself ("the floating comment button intermittently fails to appear on the first selection"), never who reported or requested it. Related work is referenced by PR number or by naming the mechanism ("tracked separately as a follow-up"); the linked issue links to the PR, never the reverse. Ticket and support context lives wherever the work is tracked and in support channels, not in the PR.

Shape

• User-facing PRs open with the changeset's prose, verbatim. The changeset is the calibrated consumer sentence (see the changesets skill), and it is what the reviewer actually reads; a body that buries it under paragraphs gets skipped whole. After it, the body adds only what changesets ban: a concrete example of the failure, one sentence naming the fix mechanism, any behavioral change or honest caveat. When the changeset alone carries the change, it is the whole body. This mirrors the changeset's prose; body sentences about the changeset ("a changeset rides along") remain banned.
• Fix/refactor PRs: heading-free narrative prose. After the changeset opening: diagnosis in code terms → fix → what was verified, compressed into the paragraph the budget below allows. Larger feature PRs may use ## What / ## Design notes.
• Lead with motivation (internal-only PRs, which have no changeset to open with): the pain that makes the change necessary, before the mechanism. The description stands on its own; explain the problem in terms of the code, not the PR graph, and reference another PR only when the reader cannot follow the description without it.
• Prose over bullets. Each design decision gets a paragraph carrying its own reasoning; bullets flatten the why out. A short bullet list is acceptable only for genuinely parallel items (e.g. "things I was careful about").
• No ## Tests section. A test inventory is listy and disconnected from the reasoning. Tests appear in the narrative next to the decision or claim they pin, "pinned" meaning a test that fails without the change ("pinned by a test expressing...").
• Justify each non-obvious decision inline, paired with the test that pins it.
• Fenced code examples when introducing API.
• Stacked or extracted PRs state why they are separate and what depends on them, referencing the other PR only when the reader needs it (#2777: "Extracted from #2772 because it changes behavior for all existing editor.on events ... #2772 stacks on this").
• No generated-by footers, no emoji banners.

Honesty

• Volunteer behavioral changes: one nobody would catch in review gets its own sentence, even a narrow one.
• Say what is not covered: an untested path, a case deliberately left out of scope.
• Perf numbers carry before/after and methodology (environment, run count, what was measured), never a bare percentage.
• No boilerplate self-assessment. "Type-only change", "no changeset needed" restate the diff; state something non-obvious about blast radius instead, or say nothing.

Length: the budget

• The body scales with the diff's blast radius, not the investigation's length. Default for a fix PR: the changeset's prose plus at most one short paragraph (~100 words): the concrete example, the mechanism, one honest caveat. The 3-4 paragraph, ~200-word shape (failure, design, blast radius with its pin, caveat) is earned only by PRs with a genuinely large blast radius; it is not the template.
• A trivial diff earns two sentences: a verdict, and a rider naming anything else the diff also does. Zero citations, zero mechanics.
• Everything cut must already live somewhere: mechanism in the commit bodies, diagnosis narrative and decisions on the linked issue, specifics in the tests. The body is a map, not the territory.
• When scope grows mid-PR, rewrite the body top to bottom: never append, and never patch just the sentences you remember; re-read the whole body and rewrite. Appended addenda are how walls form, and patching only the remembered sentences is appending's twin: both skip the whole-document read. After any commit lands on, leaves, or is reworded on a branch with an open PR, re-read the entire body against the tip before pushing; stale claims survive wherever you patch from memory.
• One-screen test: after one screen the reviewer knows what changed, why, and what to scrutinize. If not, cut until they do.

Exemplar: fix PR, heading-free narrative (#2774, merged)

This is the earned-length exception (a factorial blow-up in a public conversion path), not the default fix-PR shape; the default is the changeset-led body above.

fix(sanity-bridge): avoid combinatorial blow-up converting mutually-embedding types

While upgrading our studio from v4 to v5 we hit a complete freeze: opening any document with a portable text field would pin the main thread until the tab crashed. No errors, nothing in the console. We eventually traced it to sanitySchemaToPortableTextSchema.

The walk added in #2630 tracks ancestors per branch, so it only stops when a type repeats within its own chain. Our schema has ~30 block object types, and about a dozen of them embed the shared blockContent array again (accordions, quotes, cards and so on). With that shape the walk ends up visiting every possible path through the type graph, which grows factorially with the number of block objects. For us that was ~2.1s per conversion — and since PortableTextInput converts on render, the studio just locked up. A slightly more connected schema never finishes at all.

The fix relies on the fact that Schema.compile produces one canonical instance per named type [...] I added a small memo (scoped to one conversion call) keyed by the compiled type instance [...] which makes the walk linear in the size of the compiled schema.

A few things I was careful about:

• The memo key is the instance, not the name. Same-named inline declarations with different fields compile to distinct instances [...] the cases in same-name-objects.test.ts are unaffected.
• Cycle detection via ancestor names is unchanged, and only completed expansions are memoized [...]
• All 15 existing tests pass without modification. [...] output is byte-identical.

Numbers: the new regression test (12 block objects sharing a named array type) runs in ~5ms with the fix. Without it, vitest times out [...] Against our production schema the conversion goes from 2,122ms to 1ms with identical output.

One semantic note worth flagging: a memoized expansion is computed under the ancestor chain of wherever it was first reached. [...] No existing test pins a case where the difference is observable.

The merged PR body ends with a generated-by footer; that's the one part not to copy — the anti-patterns below ban it.

Why it's canonical: opens with the lived symptom, diagnoses with the actual growth mechanism, states the fix as an insight about the platform (Schema.compile canonical instances), the careful-abouts each pair a decision with the test guarding it, numbers have before/after + methodology, and it volunteers a behavioral change nobody would have caught in review.

Exemplar: feature PR with headings (#2772, merged)

Structure to copy (see https://github.com/portabletext/editor/pull/2772 for the full text):

• ## What opens with the concrete driver (Studio's render-pipeline work needs listIndex), generalizes to the class of consumers (derived state), then shows the API with a fenced example and the exact test pinning the motivating use case end-to-end.
• API-surface discipline gets its own paragraph: the closed five-variant union, the type-level tripwire that fails compilation when the engine vocabulary grows, the runtime allowlist, "exposing a future operation is a decision made twice, never an accident".
• ## Design notes gives each contested decision a paragraph with its reasoning and its pinning test: why the stream is ungated (with the test pinning the contrast), delivery order under normalization (both real orders, which test pins which), error isolation, wildcard listeners.
• Performance measured and stated with methodology: "chromium, medians of 3 runs, inserting 1000 blocks: 260ms vs 259ms [...] Parity within noise."

Anti-patterns

• ## Tests / ## Changes inventory sections.
• Bullets where each item secretly needed a paragraph of why.
• "This PR fixes a bug in X" openings, open with the observable pain instead.
• Explaining the problem only via the PR graph ("as discussed in #1234").
• Claiming coverage the tests didn't actually run; say what was and wasn't exercised.
• Self-assessments that restate the diff; generated-by footers.
• The words "delta" (write "change") and "rides along"/"ride along", and labels such as "One additional change:", anywhere in artifact prose. Name a secondary change in a plain sentence instead.
• The investigation narrative leaking into the body: how the bug was found belongs wherever the work is tracked; the body describes the change.
• Appending paragraphs as the PR evolves instead of rewriting the body.