Repository navigation
jobs: toStepRecord casts an unvalidated status column into StepStatus — a value outside the union reaches every consumer typed as if it were inside it #213
Description
Found while typechecking packages/jobs' tests for the first time. The type error was in a test; the mechanism it exposed is in the mapper.
What the test was doing
packages/jobs/src/driver-pg-rows.test.ts built a step row with status: 'suspended' and asserted it came back out as 'suspended'. But StepStatus is completed | sleeping | waiting | failed (packages/jobs/src/steps.ts:15) — 'suspended' is a JobRecord['state'], never a step status.
The test passed. That is the finding.
Why it passed
packages/jobs/src/driver-pg-rows.ts:
status: row.status as StepRecord['status'],The column is cast, not validated. So the test's impossible value travelled straight through the mapper and came back out, and the assertion compared it against itself.
Why it matters beyond the test
The mapper will do exactly the same thing for a real row. A step row written by an older deploy — or by a future one that adds a status this build does not know — carries a value outside the union, and the cast launders it into every consumer as if it were inside. Nothing narrows it, nothing logs it, and every downstream switch on StepStatus silently takes no branch.
This is the same shape as the isCookie lying-predicate bug (fixed in the scraping slice): a type assertion standing in for a check, on data that crosses a process boundary.
Why it was not fixed in the typecheck slice
Deciding what to do with an unknown status is a behaviour decision, not a type fix:
- refuse the row with an
UltimateErrorcarrying a stable code and afix:? - map it to
failedand log? - keep it and widen
StepStatusto| (string & {}), which pushes the decision to every consumer?
The first is most consistent with the house rule that an absent or unrecognised fact is never a satisfied one, and it is what X_*-coded errors exist for. But it changes what a rolling deploy does when two versions share a queue, which needs its own test and its own thinking about the upgrade path.
The test itself is fixed ('suspended' → 'waiting', and the mutation toStepRecord stops passing the status through now reddens it). This issue is the mapper.
Activity
developerz-ai commentedon Aug 20, 2026
I couldn't reliably classify or scope this report right now, so I haven't guessed or handed it off. A maintainer will review it.
🤖 Posted by developerz.ai — the maintainer agent, not a human.
test(7 packages): 229 more test typecheck errors, and the fetch seam …
fix(scraping,admin,jobs,action,storage,render,cli)!: close the known-…
feat(scripts,framework)!: four gates, and the last ten known gaps (#219)
sebyx07 commentedon Aug 20, 2026
Fixed in #219 — and it was three casts, not one, with a worse consequence than filed.
toStepRecord(status),toJobRecord(state) andtoBackfillRun(status) all cast. The sharpest form:isBackfillStatusalready existed atbackfill-ledger.ts:30with the doc "Never a cast — the list decides", andtoBackfillRuntwo files over was the one caller ignoring it.The severity this issue understated.
steps.ts:263isif (existing?.status === 'completed')→ return the memoized output. Every other value re-executes the step. So a laundered status turns "this step already ran" into "run it again" — a secondstep.run('charge', …)— in a file whose header promises the welcome email is not sent twice.Now
STEP_STATUSES/isStepStatus,JOB_STATES/isJobState, and onestatusIn()narrowing used by all three decoders. New codeX_JOB_ROW_STATUS_UNKNOWN(500).What a mixed-version deploy does, since the two directions are not symmetric — a status string only reaches the column because some build of this framework wrote it:
fix:says to drain the older workers. Previously that same row silently re-ran its steps.One honest gap recorded rather than papered over: "no
asonto a closed union in a decoder" is not a build error today. A cast always typechecks, so this is runtime-enforced only. A gate for it would look likebun run error-renderand belongs inscripts/.