A subject id is a string, whatever the row is keyed by

[czpython]

A subject id is a string everywhere it is read — the URL segment, the DBOS attribute, the dedup key, the event row — while the row behind one is keyed by an integer. Nothing named that, so every summary bridged it by hand:

id=str(note.id)

And a constructor written for that one field went on to hand-copy the rest, which is how NoteSummary.from_note ended up listing five fields to transform one.

The identifier is a type

SubjectId = Annotated[str, BeforeValidator(str)]

Takes either and is always the string — the same settlement GraphQL, JSON:API and Protobuf all reached independently:

• GraphQL's ID — "often numeric, it should always serialize as a String", and "any string (such as "4") or integer (such as 4) input value will be accepted".
• JSON:API — "The values of the id, type, and lid members MUST be strings."
• ProtoJSON — int64 becomes a decimal string, "either numbers or strings are accepted", because JSON parsers are "silently lossy if a number is an integer larger than 2**53".

Scoped to id rather than ConfigDict(coerce_numbers_to_str=True), which would widen every string field on the model: a label or a title arriving as a number stays a mistake.

What it removes

from_attributes on the header means a subject builds its own. The platform default on both bases is now SubjectSummary.model_validate(self), and a note's summary is its columns rather than a constructor listing them — from_note is gone, and field_notes' schemas.py no longer imports its models.py at all.

What remains composes, and the next commit shows how little of it there was.

Not a wire change

id=str(item.id) and id=item.id serialize identically. No response shape moves, so the frontend is untouched.

Ship's own shapes follow

WorkItemSummary and DashboardItem each declared every field and then assigned every field, which buried the only two bits of real work: a project name reached through a relation, and a set of links.

project_name: str = Field(validation_alias=AliasPath("project", "name"))

@computed_field
@property
def links(self) -> Links:
    return Links.for_work_item(repo=self.repo, pr_number=self.pr_number, ticket_url=self.ticket_url)

The relation is an alias and the links compute from fields the shape already carries, so both build with model_validate and neither constructor survives. Links keeps the one that earns it — the URL rule, in one place, called by both.

Both serialize byte-for-byte as before. ticket_url rides DashboardItem only to feed its links and stays off the wire via exclude=True, and source_id keeps its name through a validation alias.

ProjectRepoSummary.from_repo is deliberately untouched: it runs repo.get_status(), one database query per repo on GET /api/ship/projects — measured at 12 for 12 repos. Making it declarative would hide that, not fix it, and the fields it derives are the run state the platform now serves on the repo's own board. Its own ticket.