Skip to content

[AutoFyn] autofyn/2026-04-10-f84f65#72

Open
kiwi0401 wants to merge 9 commits intoautofyn/2026-04-10-9e09ccfrom
autofyn/2026-04-10-f84f65
Open

[AutoFyn] autofyn/2026-04-10-f84f65#72
kiwi0401 wants to merge 9 commits intoautofyn/2026-04-10-9e09ccfrom
autofyn/2026-04-10-f84f65

Conversation

@kiwi0401
Copy link
Copy Markdown
Member


Branch: autofyn/2026-04-10-f84f65 · Run: da37b4b2-cb5e-4bbd-89b7-6e74bbcf4a2d · Generated by AutoFyn

AutoFyn Bot and others added 9 commits April 10, 2026 20:54
…x patterns

- Add step 3b JOIN TYPE DECISION: LEFT JOIN default with 3-condition gate for INNER JOIN
- Add cumulative/running-total model rule: full spine range must appear in output
- Improve current_date warning with concrete fix patterns (4 common variants)
- Exclude calendar/spine models from pre-computed data advice
- Add scan_package_date_spine_models scanner function
- Fix MCP gateway URL for Docker networking (host.docker.internal)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add scan_nondeterministic_ordering() to detect ROW_NUMBER with non-unique ORDER BY
- Inject warning into agent prompt when non-deterministic patterns found
- Lower over-filter threshold: any row loss now triggers compare_join_types
- Add anti-rationalization rule: "Do NOT use semantic reasoning to justify INNER JOIN"
- Remove dead scan_package_date_spine_models function

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Remove exception that kept current_date for age/days_since calculations
  (f1003 proved replacement works; zuora001 needs it replaced)
- Add guidance against unnecessary ROW_NUMBER dedup on mapping tables
- Strengthen WHERE filter rule with anti-dedup examples

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Previously only flagged ratio < 0.5 as over-filter. Now flags 0.5-0.95
as INFO with explicit instruction to call compare_join_types. This
catches cases like flicks001 where 17% row loss was rationalized as
"expected" when it was actually wrong JOIN type.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…cement

Spine/calendar models were excluded from the "query pre-computed max date"
guidance, causing agents to cap calendar spines at max transaction date instead
of the pre-computed spine end. For xero001 this produced 354 rows vs 1170 gold.
Now spine models with pre-computed DB entries get the correct hint plus a
SPINE CRITICAL urgency note.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…rigger

- Step 3b: Allow INNER JOIN when verified by compare_join_types, not just
  semantic reasoning. Prevents over-counting in apple_store001-type tasks.
- Step 4b: Add LPAD zero-padding guidance for hour_*/month_*/day_* columns
  to fix format mismatches (reddit001, netflix001).
- Step 4d: Trigger compare_join_types on row over-count (fan-out) as well
  as under-count. Both directions indicate potential JOIN issues.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The general instruction said "Get max date from get_date_boundaries" which
contradicted the per-model hint saying "use pre-computed table max instead."
Added explicit EXCEPTION line so the agent knows model-specific hints take
precedence. Fixes pendo001 spine being capped at source max (2021-07) instead
of pre-computed spine max (2024-09).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The previous COALESCE(0) guidance was too broad — it told the agent to use
COALESCE for all SUM aggregates, but the evaluator expects NULL for monetary
columns when no data exists (NULL != 0). Salesforce001 regressed because
opportunities_lost_amount got 0 instead of NULL on 7 rows.

Now COALESCE is restricted to explicit count columns only.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The softened anti-rationalization rule ("alone") caused flicks001 regression —
agent rationalized INNER JOIN without calling compare_join_types. Restoring
strict "Do NOT use semantic reasoning" while keeping the bidirectional audit
trigger in step 4d for catching fan-out cases.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant