Eve library

The job adapter, live — 2026-09-12

Guide

Revision 1

On this page

Evidence for 182.C.15 (submit, read, list, cancel) and 182.C.26 (likeness authorizations and their withdrawal). Five responses from a BFF this session started against the local Postgres, not from a schema.

Method#

apps/oshun/bff/src/server.ts under tsx on :4010, with OSHUN_V1_DATABASE_URL pointed at 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 finding that unblocked everything else#

The previous capture's README said the environment's prisma CLI "is 6.19 and cannot regenerate the v7 schema". That was wrong, and it was the only reason five adapters reported unbound. The root pnpm exec prisma resolves to nothing; libs/oshun/persistence/node_modules/.bin/prisma is 7.8.0 and regenerates the schema in 1.5 seconds. With the client regenerated, the boot line changes from one bound port to seven:

text
human-video api: 32 endpoints live; bound: preset catalogue, project drafts,
idempotency records, upload sessions, identity packs, audio takes, jobs;
answering 503 until bound: candidates, reviews (queue and decisions), releases,
erasure, project listing

boot-line.txt carries it verbatim.

The 500 that was hiding behind the 503s#

With the idempotency port bound for the first time, the first POST of every idempotent operation answered 500:

text
new row for relation "human_video_idempotency" violates check constraint
"human_video_idempotency_response_only_when_complete"

Prisma writes a JS null into a Json? column as the JSON value null, and jsonb 'null' is not IS NULL. The reservation row therefore carried a response while its status was in-flight, which the CHECK refuses — correctly.

The fix is to omit the column rather than write null to it: an omitted column takes its default, which is SQL NULL. Prisma's DbNull sentinel would also work, but it lives behind @prisma/client/runtime/client, which the BFF's vitest alias maps into a file path — so importing it makes every spec in the app unloadable. Omission needs no sentinel and no import.

The two values are genuinely different facts: "no body was kept" against "the body was the value null", and a response body can legitimately be null. The row type now says so — response?: StoredResponse | null — and rowFor spreads the key in only when there is one.

Measured after the fix, on the same database:

text
   key    |  status   | response_is_sql_null
 k-omit-1 | completed | f
 k-ev-1   | abandoned | t

An abandoned key holds SQL NULL; a completed one holds the body it owes a retry.

This could not have been found by a test against a double. It needed a real client and a real constraint, which is the argument for this capture existing. A double that fires on the same condition now stands in human-video-api-store.spec.ts so it cannot come back.

Idempotency, end to end#

POST /jobs twice under Idempotency-Key: k-omit-1 returned 202 both times with the same jobId (3d49ccf4…), and the job table held three rows, not four. The second call replayed the first answer rather than submitting again.

The four captures#

file request status
submit-202.json POST /jobs, real pack, live consent 202 — job written, authorizationIds: [auth-ama]
submit-403-consent-withdrawn.json the same after a whole withdrawal 403 hv.consent.missing, terminal
submit-403-scope-exceeded.json the same after a withdrawal of proj-live only 403 hv.consent.scope_exceeded
list-200.json GET /jobs 200 — keyset page, hasMore: false

The withdrawal, measured#

One authorization covering proj-live and proj-other, withdrawn for proj-live only:

  • POST /jobs for proj-live → 403 hv.consent.scope_exceeded, "proj-live" is not one of the 1 project(s) named.
  • POST /jobs for proj-other → 202, authorizationIds: ["auth-ama"].

The store takes the withdrawn projects out of projectIds and returns what is left. A withdrawal that empties the list drops the scope entirely rather than returning one with no projects — an empty projectIds makes coversUse throw dimension_unspecified, and the caller turns that into "this record is incomplete", which is a complaint about our bookkeeping in place of the fact that somebody took their permission back. Dropped, the answer is hv.consent.missing: true, terminal, and the sentence the creator needs.

What the submitted row actually holds#

text
 job_id  | stage      | cast_ids   | authorization_ids | unmeasured | attempts
 a38ede84| validating | {cast-ama} | {auth-ama}        |          5 |        0

Five named unmeasured facts and zero attempts, both stored as arrays because the migration requires them to be. unmeasuredFacts empty would say every pre-generation check ran; absent would say the same thing more cheaply. The five are real — no frame exists before a render, no reference embeddings were supplied, the pack records no viewing angle, and no approved audio is attached.

What is still unbound#

candidates, reviews, releases, erasure and project listing. A job can be submitted, read, listed and cancelled; nothing yet produces a candidate from one, because there is no provider adapter behind the render.

What a real end-to-end render is blocked on, measured rather than assumed#

182.C.40.01 asks for at least one production-configured render demonstrated end to end. As of this commit the CODE path is complete and proven hermetically in human-video-render-end-to-end.spec.ts: a planned job goes through the request builder, the provider transport, the download, the container probe and the candidate store, with only the four boundaries a test may replace standing in — the database driver, the object store, the HTTP transport and ffprobe.

Two things stop that from being run for real from this machine, and neither is a money problem. ~$4.35 of the approved canary budget remains and is not the constraint.

  1. A consented reference. A human-video render needs an identity pack, and a pack needs photographs of a real person who agreed to this use. That is 182.C.33.01's corpus, which is blocked. Rendering a stand-in face would be the exact thing this phase refuses.
  2. A reference URL the provider can fetch. placeReferences hands the provider a URL per reference, and the provider fetches it. The local object store is MinIO on a private port; a presigned URL to it is unreachable from OpenRouter. A render with no references is not a likeness render, so it cannot demonstrate the item either.

So 182.C.40.01 is blocked on a consented corpus and on a publicly reachable object store, not on the pipeline. That is a different sentence from the one this item carried before today, when the pipeline itself did not exist.