Home
# Review: DialRush 5+ minute feed, commit e8ae0a3
Reviewer: Fable reviewer thread, Nov 19 2026. Read cold, read only. No rewrites made.
## Rules that hold
- The sold step and dialrush_pull_calls are untouched. The learning step runs only after a clean sold step.
- Nothing goes to call_reviews.
- Layer 1 + names, Layer 2 and the ingest's own mask are all applied.
- Logs, the ledger and the cron answer carry code words, call ids and counts only.
- 3 new calls per run and the daily cap are enforced.
- The migration adds one table and 2 functions, service role only. The undo drops only those. Both files end with sql_run_mark.
## Changes, most important first
1. **Listing load.** Every 5 minute run lists yesterday and today with no GAP_MS between list requests (the header says there is one). Yesterday had 605 such calls, so about 4 pages a day and about 8 back to back requests a run, even after the daily cap is reached. A refusal there starts the shared 30 minute back-off and counts toward the 3 a day alert, so this step can pause the sold pull. Add the gap between list requests, and skip the listing when the cap is already used or the ledger still holds due rows.
2. **Tests use a fake ingest.** The handed body was never checked by the real validateLearning (readymode_call_id "dialrush:<id>", the placeholder recording_url, disposition rules, call_date window). Add one test that passes the handed body through the real validator. Today a mismatch would mark every call bad_row for good, with only a code word in the log.
3. **bad_row is permanent.** A systematic refusal burns every call silently. If the first calls of a run all come back bad_input, stop the step and log one distinct code instead of marking more.
4. **Which calls get picked.** Selection is oldest first with a default cap of 20 a day against about 600 calls, so the cap fills with yesterday's earliest calls and today's are never reached. State in the header which calls are meant to be picked, and confirm the intended cap with Matteo before switching on.
5. **Time budget.** Up to 3 calls, each a 12 s gap + up to 45 s transcript + a Layer 2 read, can start as late as 150 s into a 280 s lease. held() before the ingest keeps it correct, but check one call's worst case against LIST_RESERVE_MS, or start only when more time is left.
6. **"already" uses the cap.** A status of "already" marks done and uses a slot of the daily cap. After an Undo and re-run, re-listed calls would eat the cap. Do not count "already" against room or doneToday.
7. **Header wording.** The migration header says it adds only, but dialrush_learn_plan deletes ledger rows older than 30 days. Harmless (ids only), but say so in the header and the undo note.
8. **Unlabeled transcripts** are counted on the sold path and not here. Add the count or note why it is not needed.
## Not checked
The test file line by line, a live run, and the Learning store's own daily cap behaviour.