[matej21]

CI on main broke when oven-sh/setup-bun started installing Bun 1.4.4. rpc-handlers > query.execute automatically logs to history now fails, because a SELECT 1 is logged with rowCount: 0.

Cause

History and the session log record affectedRows ?? rowCount. Bun.SQL reports affectedRows for every statement, including 0 for a SELECT:

Bun	PostgreSQL SELECT	MySQL SELECT	SQLite SELECT
1.3.14	null	0	null
1.4.4	0	0	0

So a read was always logged as 0 rows on MySQL, and since Bun 1.4 on PostgreSQL and SQLite too.

Fix

statementRowCount() in @dotaz/shared/query-result counts what a statement returned when it produced a result set, and what it changed otherwise. This is the same split the result panel already makes on columns.length. It replaces the expression in all four places that summed it:

• backend history
• session log
• demo adapter
• frontend history entry

Verified

• New test: history counts returned rows for a read and changed rows for a write.
• The unit suite passes under both Bun 1.3.14 and 1.4.4 (1349 pass, 0 fail).
• lint, format and tsc are clean.

Status

• Review round 1 (751c9a9, independent Claude subagent): approve. Nits addressed in 9f9dc75; one older demo-driver bug is left as a follow-up.
• CI: green on 9f9dc75 (check + integration).
• Merged with a merge commit (7654870). feat: agent CLI — database access for AI agents, with approval-gated writes #19 is rebased onto it.

🤖 Generated with Claude Code

[matej21]

Review round 1 — 751c9a9

Reviewer: independent Claude Code subagent (Claude Opus). Ran real statements against PostgreSQL, MariaDB and SQLite on Bun 1.3.14 and 1.4.4.
Verdict: approve

The fix is correct for all three Bun.SQL drivers on both Bun versions. Cases checked:

• zero-row SELECT
• INSERT/DELETE … RETURNING
• EXPLAIN and SHOW
• SET and DDL
• error, cancelled and timeout results
• multi-statement runs

No other site sums the old idiom.

Findings

1. low — demo driver reports a stale affectedRows (predates this PR). src/frontend-demo/wasm-sqlite-driver.ts:104 uses this.db.changes(), which keeps the last write's count. After INSERT … VALUES (1),(2),(3), a zero-row SELECT, CREATE INDEX or PRAGMA reports 3, and demo mode shows "3 rows affected". The fix belongs in that driver.
2. low — test coverage. The new test runs on SQLite and covers history only. On Bun 1.3 it would pass without the fix. A zero-row SELECT case would pin the fallback branch.
3. nit — doc comment. statementRowCount says "produced a result set", but columns is built from the first row, so the check means "returned at least one row".
4. nit. SqlResultPanel.tsx still writes affectedRows ?? rowCount inline under columns.length === 0. This is equivalent to the helper's fallback branch.
5. nit. The src/shared/CLAUDE.md structure list does not mention query-result.ts.

[matej21]

Response to review round 1

1. low — stale affectedRows in the demo driver: won't fix here. This predates the PR and lives in wasm-sqlite-driver.ts, which this PR does not touch. It needs its own fix: report the change delta, or 0 for statements that return columns.
2. low — coverage: fixed in 9f9dc75. Added a zero-row SELECT case, which pins the fallback branch. CI runs Bun 1.4.4, where the test fails without the fix.
3. nit — doc comment: fixed in 9f9dc75. It now says "returned rows" and explains the zero-row fallback.
4. nit — SqlResultPanel inline idiom: won't fix. It sits under columns.length === 0, where it is equivalent to the helper.
5. nit — src/shared/CLAUDE.md: fixed in 9f9dc75.