Four responses from a BFF this session started against the local Postgres, and
the boot line they came from. The morning's capture
(../2026-09-12-live-surface-capture/) had ONE port bound and eleven answering
503; this one has eleven bound and two left.
Method#
apps/oshun/bff/src/server.ts under tsx on :4010 with OSHUN_V1_DATABASE_URL
against oshun_dev. Secrets were sourced through a throwaway /tmp script,
deleted at the end; none is in this directory, and the captures were grepped for
key material before committing. The server was torn down and :4010 confirmed
free. Tokens were minted with createDevAuthToken.
The boot line#
boot-line.txt, verbatim:
human-video api: 32 endpoints live; bound: preset catalogue, project drafts,
idempotency records, upload sessions, identity packs, audio takes, jobs,
candidates, reviews (queue derived from candidates), project listing (a project
is a draft), revocation reach (jobs and authoring drafts); answering 503 until
bound: releases, erasure
Two of those entries say something an operator needs and a list of names would not.
"queue derived from candidates". The review queue is not a table something
enqueues into. A written queue is only as true as the last thing that remembered
to write to it, and a missed enqueue is indistinguishable from a clean
evaluation — forever, with nothing reporting it. waiting reads the candidates
and works out which failed a hard gate, so a missing row is impossible by
construction.
"a project is a draft". There is no projects table. One would be a second row that could disagree with the draft about whether a project exists.
The captures#
| file | request | status |
|---|---|---|
projects-200.json |
POST then GET /projects |
201 then 200 with the created project |
reviews-200-empty.json |
GET /reviews |
200, reviews: [] |
candidates-200-empty.json |
GET /jobs/{id}/candidates |
200, candidates: [], comparisonSets: [] |
boot-line.txt |
— | what a 503 means on this deployment |
What the empty pages mean, exactly#
They are not placeholders. Each is a true answer about this database:
reviews: []— no candidate has a measurement, so nothing has failed a hard gate, so nothing is waiting for a person. A candidate nobody evaluated is waiting for an EVALUATOR, and the store deliberately leaves it out: putting it in front of a reviewer would ask them to judge a take with no numbers beside it, whichdecideActionrefuses anyway.candidates: []— the job in this database was submitted and never rendered. Nothing has produced a take for it.
What the project listing proved#
POST /v1/human-video/projects answered 201 and
GET /v1/human-video/projects returned it with a real createdAt of
2026-09-12T05:12:33.579Z — the column added by
20260912080000_human_video_project_draft_created_at, written on the draft's
first save. Before that migration the draft table had only updated_at, and a
listing keyed on the last edit loses rows: a draft somebody touches moves to the
front of a page a caller is part-way through, and the keyset skips whatever slid
past the cursor.
What is left#
releases and erasure. A release needs a proof store and a gate record for
the final bytes; erasure needs something that can actually delete from an object
store and report, per locator, whether the target is gone. Neither has an
adapter yet, and both fail closed.