[PavelRavvich]

Refs agbossjs/agboss#771

What changed

• Migration 0023_cost_entries_pruned adds:

  • agboss_cost_entries.made_at with an index. Rows written before 0023 get 'infinity', so they are never pruned and never refused; their pass's retry finds them held.
  • agboss_cost_entries_pruned, a one-row table holding the pruning mark, seeded at '-infinity'.
  • A COMMENT ON TABLE on agboss_cost_entries.

  0022 stays unchanged because migrations are checksummed. 0023's header corrects 0022's: a ledgerId is the pass's own id, and a step run again counts again. The README's row for 0022 now says the same.

• saveAll is one statement. It works like this:

  • It reads the mark with FOR SHARE. If a prune's UPDATE of the mark is in progress, this read waits and then sees the new mark, because a locked row is read at its latest version.
  • An entry made before the mark is never inserted.
  • If an entry made before the mark is not held, the statement adds nothing at all: no entries, no counters. It answers the refused ids, and the store throws CostLedgerEntryExpiredError.
  • If the mark row is missing (deleted by hand), the save adds nothing and throws PostgresLedgerMarkMissingError (postgres.ledger-mark-missing). Without that check, a save would silently drop its entries.

• deleteEntriesByMadeAtBefore is one statement: UPDATE … SET pruned_before = greatest(…) (the mark only goes up), then DELETE … WHERE made_at < mark.

• The test reset (truncateAll) seeds the mark again, as the migration does.

How it was tested

• npm run check and npm run test:faults green against PostgreSQL 16 under the lock, at 2a5b4641 with the core packed at a8fe1124.
• The core's costLedgerContract now includes the 4 #771 pruning cases. They run here through storeConformance against the real database.
• tests/stores/cost-ledger.test.ts:
  • 40 lost answers are retried on two instances while a third instance prunes 10 times with rising cutoffs. Every entry counts once ($5) and none is left.
  • Bite: with FOR SHARE removed, the same test counts $5.875 for $5, i.e. 7 entries counted twice. With the mark filter removed, it fails too.
  • With the mark row deleted, a save adds nothing (counters included) and names the row.
• tests/migrations/cost-entries-migration.test.ts: 0023 run on a database at 0022. An entry held from before is dated 'infinity'; a prune at 2030 keeps it, and a save of it adds nothing.
• Every migration test that lists the later migrations now includes 0023: versions 23, 24 for the "newer package" cases, and the CLI's lines.

Wiki to update after merge

None here. agbossjs/agboss Wiki → FinOps "Pruning the entries (#771)" and Your-own-store already describe this store (published with agbossjs/agboss#783).

Deviations from the spec

None

Not auto-verifiable

None

🤖 Generated with Claude Code

https://claude.ai/code/session_012Kh8JmysuhZyQRSpNyJdPg