Architecture review due (10 PRs since arch-review/2026-10-18) #302

Activity

1. github-actions

   added

   ready-for-agentFully specified, ready for an AFK agentFully specified, ready for an AFK agent

   on Oct 23, 2026
2. bitbiter-dev commented on Oct 23, 2026

   

   bitbiter-dev

   on Oct 23, 2026

   Owner

   More actions

   Reviewed arch-review/2026-10-18..8547e1d (10 PRs, #284–#300). Every finding was re-checked against master at 8547e1d. Tagged arch-review/2026-10-24 on 8547e1d and pushed.

   Filed

   • fix(worker): A file overwritten with content stored elsewhere leaves its row stale #303 (medium, needs-design): a file overwritten with content the config already stores at another path. cp makes it DuplicateCopy, mv makes it PathTaken (a warning on every crawl), and either way the path's row keeps its old hash, metadata and proxies. This is Update the row when a file's content changes at the same path #250's problem by a route Update the row when a file's content changes at the same path #250 left out, and IngestionClassifierTests pins it. Decided with the maintainer: its own issue, design first. It blocks Soft-delete missing files after a crawl, and auto-restore them when they return #33 and Auto-restore a soft-deleted asset when its file is back at its path or its content turns up elsewhere #37.
   • chore: Medium and low findings from the 2026-10-24 architecture review #304 (medium and low, grouped):
     • DuplicateCopyTally is a process-wide singleton taken only when a crawl finishes, so a cancelled, released or failed crawl's copies go into the next report. Count them per run instead.
     • Worker.Work and Worker.Crawling depend on each other through WorkerState.
     • IngestionShutdown now covers an admin's cancel as well, and its comment names a crawl loop that is gone.
   • Folded into feat(worker): Similar photos join bursts — OpenCV feature matching on thumbnails #298: burst repair departs from the spec: Bursts — group photos of the same scene by camera label or feature similarity, show one cover per burst #296 design note (section 5, step 1). Memberless bursts are loaded and EF-removed in every config's run instead of being deleted in one statement, so two per-user Workers repairing at once fail the second item with DbUpdateConcurrencyException. feat(worker): Similar photos join bursts — OpenCV feature matching on thumbnails #298's body has a new acceptance line, and the session working on feat(worker): Similar photos join bursts — OpenCV feature matching on thumbnails #298 has been told. Decided with the maintainer.
   • Noted on feat(worker): Orphaned proxy file sweeper (proxies ↔ DB reconciliation) #174: StorageConfigSchedules<TPayload> uses the crawl interval for every per-config kind, and feat(worker): Orphaned proxy file sweeper (proxies ↔ DB reconciliation) #174's kinds should not inherit it silently.

   Dropped, with reasons

   • IWorkItemStore has nine members spanning executor, scheduler and retention: they are one table's state machine, all PostgreSQL-only and contract-tested together (ADR 0008).
   • IWorkHandler and IWorkProgress live in Core although only the Worker implements them: Core holds the payloads and WorkProgress, which IWorkItemReader returns, and work-handlers-live-in-worker keeps the handlers out. Moving the interface gains nothing.
   • One item at a time per Worker, so maintenance waits behind a long crawl: decided in spec: Work queue — durable API/Worker work items with scope routing, leases and a live admin view #276, and crawls were sequential before it.
   • Result is an untyped jsonb summary served as JsonElement: decided in spec: Work queue — durable API/Worker work items with scope routing, leases and a live admin view #276 (user story 8 and the column list).
   • WorkJson.Options builds new serializer options on every call: this follows CLAUDE.md's rule against static readonly initialisers, which cause mutants that flip between runs.
   • BurstDetection is both enqueued after a crawl and scheduled: decided in the spec: Bursts — group photos of the same scene by camera label or feature similarity, show one cover per burst #296 design note (section 3), and one dedup key coalesces them.
   • CrawlPayload and BurstDetectionPayload have the same shape: eight lines each. IStorageConfigPayload<TSelf> already shares what the scheduler needs.
   • Raw SQL naming WorkItems: checked as CLAUDE.md asks. It appears only in EfWorkItemStore (Worker) and EfWorkItemCanceller (Core), and neither inserts. architecture-exceptions.json is empty.

   The Epic 4 milestone order now puts #304 and #303 after #250 and ahead of #33.

   🤖 Generated with Claude Code
3. bitbiter-dev

   closed this as completedon Oct 23, 2026