[alberthflores95]

Stacked on #91. The base is fix/persist-chats-artifacts so this diff shows only the chat repair; retarget to main once #91 merges.

Problem

Artifacts repair a rejected SQL statement once (#90). The chat did not: on ClickHouse 22.1 (GONAC PoC server) every dialect slip reached the user as Pipeline error. Errors seen in the server log on 2026-10-27, from the chat and artifacts:

Error	What the model wrote
SYNTAX_ERROR, Unrecognized token: '\xc3' (x3)	a Spanish alias with an unquoted ñ
ILLEGAL_AGGREGATION (x2)	SUM(x) AS x, then another SUM(x)
ILLEGAL_TYPE_OF_ARGUMENT (x2)	toDateOrNull(order_date) AS order_date on a column that already is a Date
NO_COMMON_TYPE (x2)	a String date compared with a Date
UNKNOWN_IDENTIFIER (x3)	a missing column, or arithmetic between window functions
UNKNOWN_TABLE (x1)	a table that does not exist in PRD (not fixable by a rewrite)

Change

• sql_repair.py (new): the repair, its dialect notes and its error hints move out of routers/artifact.py, shared by chat and artifacts. Three hints added for the errors whose text does not name the fix (non-ASCII alias, String vs Date, conversion on a Date). Connection, timeout and authentication errors are not sent to the model.
• routers/query.py: a rejected chat statement gets one repair (token phase sql_repair). A statement that fails twice keeps its second error; the chat shows the statement that ran.
• prompts/clickhouse.py: rule 7 asks for ASCII-only aliases; new rules 14 (no alias that shadows the column it reads) and 15 (window arithmetic in an outer query). Prose only: an example in a rules block acts as a template.

Not covered: Flash mode generates and runs its own SQL inside ask-intent-resolution.

Tests

• test_chat_sql_repair.py: 16 tests (one repair and its result, no repair when the statement runs, no second repair, original error when no usable repair, unreachable database not sent to the model, the three new hints with the real error texts, which errors are repairable). Checked against mutants without the chat repair and without the new hints: they fail.
• test_artifact_sub_sql_repair.py and test_artifact_env.py pass unchanged apart from patching SettingsCache where the repair now lives.

Measurement on the GONAC server (before/after on the same questions) to follow in a comment.

🤖 Generated with Claude Code

[alberthflores95]

Measurement on the GONAC PoC server (2026-10-27)

Same 12 Spanish questions, workspace resumen-de-ventas-gonac, env dev (ClickHouse 22.1, database dbt_qas_bi), mode precise, Claude Sonnet 5. Each arm in a one-off container with the live configuration; one run per arm, so read it as an indication, not a rate.

	SQL ran	Failed on the first statement	Rescued by the repair
A, sha-c5aed3e (before)	2 / 12	10	n/a
B, sha-b87fbd5 (this PR)	7 / 12	7	2

First-statement errors, by kind:

Kind	A	B
Unquoted non-ASCII alias (año)	2	0
Alias shadowing the aggregated column	3	0
Arithmetic between window functions	1	0
Column missing from the QAS table	4	6
if() branches of different types (Decimal vs 0)	0	1

The remaining failures are mostly not SQL. dbt_qas_bi.gold_mv_resumen was last modified on 2026-08-11 and has 19 columns; dbt_prd_bi.gold_mv_resumen has 27. Eight exist only in PRD (proyectado, nombre_destinatario, umb_pendiente, facturado_umb, objetivo_umb, proyectado_umb, umb_aa, unidad), and the semantic layer published to dev declares the first three. Any dev question that leads to them fails whatever the SQL; the repair rescued two by answering from the columns QAS has. That is a data refresh, not a code change.

The if() type mismatch is the one dialect failure the repair did not fix; a hint for it is a candidate follow-up, to be measured first.

Deployed to the PoC server as sha-b87fbd5.

[alberthflores95]

Same 12 questions on prod (dbt_prd_bi), sha-b87fbd5

12 / 12 ran. One first statement failed (arithmetic between window functions, UNKNOWN_IDENTIFIER) and the repair fixed it; the other eleven ran first time. The if() type mismatch seen on dev did not come up in this run.

dev is not the place to judge this workspace: its only data product has to be published to dev before prod, and dbt_qas_bi.gold_mv_resumen is an obsolete copy that lacks eight of the columns the semantic layer declares.

"Ran" means the database accepted the SQL; the figures in the answers were not checked against the source here.