writing-changesets

name	writing-changesets
description	Create changeset files for changes to ePDS that affect end users, client app developers, or operators. Use when adding features, changing configuration, altering OAuth flows, or any change a downstream consumer needs to adapt to.

Writing Changesets

Create a changeset file to document changes for release that affect end users, client app developers, or operators.

When to Use

Add a changeset when making changes that affect anyone consuming or running ePDS:

• Adding, removing, or renaming environment variables
• Changing default behaviour that operators would notice
• Changing or extending the OAuth authorization flow (new query params, new client metadata fields, new response fields)
• Changing email templates or OTP format in a way downstream consumers can observe
• Fail-fast validation that could cause previously-working configurations to error out on startup
• Changing default ports, URL schemes, or healthcheck behaviour
• New end-user-visible signup/login UI steps

Skip changesets for:

• Internal refactors with no observable behaviour change
• Tests-only changes
• CI / infra / tooling changes that do not affect anyone running or consuming ePDS
• Docs-only changes (unless they document a new operator-facing workflow)
• Changes to the internal trust boundary between auth-service and pds-core (e.g. the HMAC callback signature, internal-only routes under /_internal/). These are contributor-facing, not consumer-facing.

If in doubt, add one. An "empty" changeset (intentionally no release) can be created with pnpm changeset add --empty to document the decision.

Audiences

ePDS changesets should call out which of the following audiences a change affects, listed in this order — from the person closest to the running software outward to the person maintaining it:

• End users — people signing up or logging in via an ePDS-powered app. Most changes are transparent to them, but changes to the signup/login UI flow, to what handles they can pick, or to OTP format are worth mentioning so that operators can write their own release announcements.
• Client app developers — people building apps that authenticate against an ePDS instance via OAuth. They care about OAuth client metadata fields, query parameters on /oauth/authorize, callback handling, and anything visible at the HTTP boundary.
• Operators — people deploying and running an ePDS instance. They care about environment variables, config files, reverse proxy / Caddyfile setup, healthchecks, migrations, and breaking validation.

List only the audiences that actually need to adapt or will notice the change. Omit audiences that are not affected.

Do not list "library consumers" as an audience. The pds-core, auth-service, and shared packages are all private and never published to any registry; the only "consumer" of the shared exports is ePDS itself, and that's contributor territory.

Format

ePDS treats itself as a single release unit even though the source is organised into packages/*. Changesets is configured to track only the root ePDS package, so every changeset has a single frontmatter entry and the resulting CHANGELOG is a single file at the repo root.

Create a file in .changeset/ with a descriptive kebab-case name (e.g. configurable-otp-length.md, handle-selection-flow.md). Do not use the random name Changesets generates — give it a name that reflects the change.

---
'ePDS': minor
---

One-sentence summary in the voice of the final changelog entry.

**Affects:** End users, Client app developers, Operators

Longer explanation with enough concrete detail that each listed audience can adapt their environment or code without reading the PR or source. Use concrete names, values, file paths, and example snippets where applicable. Write the per-audience sections in the same order as the `**Affects:**` line: end users first, then client app developers, then operators.

Frontmatter

The frontmatter is always just "ePDS": <bump>. Do not list @certified-app/pds-core, @certified-app/auth-service, or any other package name — those packages are deliberately not tracked by Changesets.

Bump types:

• patch — bug fixes, non-breaking internal-behaviour changes that downstream consumers can observe.
• minor — new features, new config options, additive API changes. Also used for breaking changes while ePDS is still at a 0.x.y version, following the semver-for-0.x convention.
• major — breaking changes, only after ePDS reaches 1.0.0. Avoid for now.

If a single changeset mixes a feature and a bug fix, pick minor (the larger bump wins anyway when aggregated).

Body structure

Every non-trivial changeset must contain, in this order:

1. Summary sentence — one physical line, in the voice of the final changelog entry (not a commit message).

   The summary line is read by everyone listed in **Affects:**, not just the most technical audience. It is the text that appears in the "Who should read this release" block at the top of the release section, and the "Who should read" block groups bullets by audience — so End users, Client app developers, and Operators all see the same summary line under their own heading. Write the summary in language the least technical listed audience can understand.

   Concretely: if the changeset affects End users (even just as one of several audiences), the summary must be written in end-user language. Avoid:

   • Acronyms and protocol terms (OTP, DID, PAR, OAuth, HMAC, DPoP) — spell them out or rephrase.
   • Parameter and field names (login_hint, client_metadata, epds_handle_mode) — the user doesn't know these exist.
   • Implementation concepts (resolver chain, precedence, schema validation) — describe the observable effect instead.
   • Monorepo / infra language (auth service, pds-core, internal endpoint) — say what the user sees happening.

   Technical naming belongs in the per-audience sections further down, where a client-app developer or operator reading their section expects to see exact field names and env vars. The summary line is not the place for it.

   Examples of the same change written badly and then well:

   • ❌ "Configurable OTP code length and charset."

   • ✅ "Longer sign-in codes, optionally mixing letters and numbers."

   • ❌ "Handle selection step in the new-user signup flow."

   • ✅ "Choose your own handle when signing up, instead of being given a random one."

   • ❌ "Accept AT Protocol handles and DIDs as OAuth login_hint values."

   • ✅ "Sign in faster from third-party apps that already know who you are."

   If the change is purely operator-facing (no End users in the **Affects:** line), operator language is fine in the summary — the only audience reading it is operators.

Do not use ## or ### markdown headings anywhere in the changeset body. @changesets/changelog-github renders each changeset as a single bullet in the release section: the first line becomes the bullet, every subsequent line is indented by two spaces to stay inside the list item. An indented ## will either break out of the list (GitHub) or render as literal text (other renderers), in both cases mangling the changelog. Use bold inline labels (**End users:**, **Operators:**) instead. Fenced code blocks and nested bullet lists are fine because they survive the 2-space indent cleanly.

The goal: a reader in one of the listed audiences should be able to adapt their environment or code without reading the PR or the source. Changesets are not release-note marketing copy — they are migration instructions.

Keep it as short as the change deserves

"Concrete" is not "exhaustive". Include exactly what the reader needs to adapt, and stop. Most changesets are two to four sentences per audience; a small change is often a single sentence. Length must track the size of what the reader has to do, not the effort that went into the change. A one-line behaviour tweak gets a one-line changeset even if the PR was large.

Brevity and bullets are not in tension: cutting is about how many points survive, bullets are about how you lay them out so the survivors stay scannable. First cut to the points the reader must act on; then, if 3+ remain, lay them out as bullets rather than packing them into one paragraph — a short multi-point section is still a wall of text when run together. Keep one- or two-point sections as inline prose. If you find yourself writing a fourth or fifth bullet, that is a signal to cut, not to keep listing.

Before you finish, delete anything the reader does not act on:

• A verbatim example (a full log line, a whole JSON body) when naming the fields already tells them what to grep or send. Show a shape only when the exact serialization is the thing they must match.
• Rationale and "why" prose — "since this can signal X", "so that Y". State the observable behaviour; the reasoning belongs in the commit or PR, not the release note.
• Exhaustive enumerations — every endpoint, every reason string, every excluded case — when a representative name plus "and related …" conveys the same adaptation. List the full set only when the reader must match against each member.
• Internal filtering the reader never sees — "3xx are filtered before this runs, 5xx are logged elsewhere". If it doesn't change what they do, it doesn't belong.
• Caveats about edge cases the reader cannot act on. One clause at most; usually cut.

Rule of thumb: if removing a sentence would not change what the reader does, remove it. A tight four-line changeset that names the fields and the observable effect beats a fifteen-line one that also explains the mechanism, quotes a sample, and enumerates every case.

What "concrete detail" means

Good adaptation detail includes:

• Exact env var names and their accepted values, defaults, ranges, and validation errors.
• Exact field names in OAuth client metadata, query params, HTTP request/response bodies.
• Exact function/export names being deprecated or renamed (only when the change actually crosses a public boundary).
• Copy-pasteable examples for client app developers — e.g. an example client metadata JSON with the new field highlighted.
• The error message that operators will see if they hit a new fail-fast path, quoted verbatim so they can grep for it.
• The old behaviour vs the new behaviour, explicitly, whenever defaults change.

Bad (avoid):

• "Various improvements to the OTP flow." — tells the reader nothing actionable.
• "Now configurable via env var." — which env var? What values?
• Paraphrased field names like "the handle mode setting" when the actual field is epds_handle_mode.

It is fine to reference documentation (e.g. .env.example) for long details, but the changeset itself must still name the thing and describe the change. Do not offload the entire explanation.

Depth check: adaptation detail, not architecture detail

Per-audience sections answer "what do I need to do differently" for that audience. They are not PR descriptions, design docs, or architecture overviews. Even when writing for a technical audience (Client app developers, Operators), keep the focus on the adaptation surface and resist the urge to explain how the change is implemented.

The bar to clear before including any technical detail in a per-audience section: does the reader's adaptation depend on knowing this? If they could change their config / code correctly without it, leave it out.

Avoid in per-audience sections:

• Internal package, module, or middleware names (pds-core mounts a pre-route guard…, the welcome-page-guard…, auth-service's session-detection middleware…) — readers consume ePDS as a single product, not a monorepo.
• References to upstream-source-code paths, function names, or internal upstream behaviour (upstream's request-manager.js force-overrides…, the SPA's delegated click listener…).
• Step-by-step request-flow narratives (A redirects to B which mounts C which calls D…) — these belong in docs/design/*.md, not in release notes.
• "Why" explanations of internal mechanics (…to prevent the React SPA from intercepting the click) — readers need to know the observable contract, not the trick used to deliver it.

Useful in per-audience sections (the things readers do adapt against):

• The exact env vars, query params, metadata fields, error messages they will see or set.
• Defaults, validation ranges, fail-fast errors.
• "Old behaviour vs new behaviour" framed as observable effects, not internal causes.
• Conditions under which a feature self-disables or activates.

Test the section by asking: if a downstream consumer pasted this section into their team's "what changed" channel, would they know what to do, without anyone needing to ask "but what does that mean for us?" If the answer is no, the section is too implementation-flavoured.

Structure dense sections as bullets

Per-audience sections frequently end up holding 3+ distinct points (a behaviour change, a config knob, a fallback condition, a related caveat). Rendered as one wall-of-text paragraph these are hard to scan in release notes — readers skim for the points that affect them and lose them in dense prose.

When a per-audience section has 3+ distinct points, structure them as a bullet list under the bold audience label. Keep one short framing sentence on the same line as the label if useful; put each distinct point on its own bullet. One short bullet per point is much easier to scan than a long paragraph that mixes them.

Two- or one-point sections can stay as inline prose under the label — the bullet structure is for the dense case, not a mandatory shape.

Bullet lists survive the 2-space indent that @changesets/changelog-github applies to changeset body lines (see the ## heading restriction in Body structure above), so they render correctly in the generated CHANGELOG.md and on GitHub release pages.

Example

---
'ePDS': minor
---

Longer sign-in codes, optionally mixing letters and numbers.

**Affects:** End users, Operators

**End users:** OTP codes may now be longer than 6 digits and may include uppercase letters if the operator has opted into the `alphanumeric` charset. Codes of length 8 or more are displayed with a space in the middle (e.g. `1234 5678`) for readability; copy-paste still yields the flat code.

**Operators:** two new environment variables on the auth service — `OTP_LENGTH` (integer, range 4–12, default 8) and `OTP_CHARSET` (`numeric` (default) or `alphanumeric`). Values outside the range cause the service to fail on startup. The OTP input form fields adapt automatically; no template changes required.

A denser change with several distinct points per audience is better as bullets:

---
'ePDS': minor
---

A second app sign-in in your browser now skips the email code step.

**Affects:** End users, Client app developers, Operators

**End users:**

- After signing in once, a second app asking you to sign in skips the email code step.
- You either land on the "approve this app" screen or on a chooser where you confirm which identity to reuse.
- A "Use a different account" link on the chooser takes you back to the email form.

**Client app developers:** no client-side changes required.

- When a previous sign-in's cookies are present, the user lands on the account chooser to confirm which identity to reuse.
- To force the email code form instead, set the standard OIDC `prompt` parameter to `login` on the authorization request your OAuth library sends.

**Operators:** no new required configuration.

- ePDS auto-detects whether the auth service shares a parent domain with the PDS (`AUTH_HOSTNAME` ends with `.<PDS_HOSTNAME>`) and broadens the device-session cookies to that parent. On unrelated hostnames the feature self-disables.

What the **Affects:** line feeds

After changeset version runs, scripts/changelog-audience-summary.mjs post-processes the newly-written release section of CHANGELOG.md. For every changeset bullet it reads the **Affects:** line, groups summaries by audience, and prepends a "Who should read this release" block at the top of the release section. Each summary bullet in the block is a markdown link to an HTML anchor on the corresponding detailed entry below, so a reader can click through from their audience to the specific adaptation instructions.

You do not need to do anything special to feed this — as long as each changeset has a valid **Affects:** line, the block is generated automatically on release. If a changeset omits the line, the script fails and the release workflow stops, which is intentional: it means "forgot to tag your changeset" is a loud, fixable error, not silently-missing release notes.

Publishing

See docs/PUBLISHING.md for how changesets are consumed at release time and how the release workflow (manual phase 1 to open the Release PR, automatic phase 2 to tag + publish the GitHub Release on Release-PR merge) produces the git tag and GitHub Release.

Key files

• .changeset/config.json — Changesets configuration.
• Root package.json — the "workspaces": ["."] field is what puts Changesets into single-package mode. See docs/PUBLISHING.md for why this coexists with pnpm-workspace.yaml.
• .changeset/*.md — existing changesets are the best reference for naming and format.
• scripts/changelog-audience-summary.mjs — post-processor that reads **Affects:** lines and builds the "Who should read this release" block.
• docs/PUBLISHING.md — full publishing / release workflow guide.
• .github/workflows/release.yml — the workflow that consumes changesets.