Verification is one query whose rows are the answer

[asubbarao]

#57 merged a verification template that wrote both sides to the agent log as parquet and read them back with read_parquet. The comparison doesn't need that round trip: it is one query.

• agent-dispatch section 3: the template runs the proposal (one shellfs uv run --no-project python - reader, which holds the inputs once) and the DuckDB version over the same rows. It returns one row per input: duckdb_answer, python_answer, python_error, differs. Logging the run is the agent-log COPY wrapped around this same query, and nothing reads it back to get the answer.
• useful_queries.sql: removes the read_parquet "Verification: both sides" readback entry.

Verified on dev 2026-10-18: T4's proposal against a DuckDB expression on 4 inputs. The +05:00 midnight, the year-1 OverflowError and an input containing ' all come back as differs = true in the query result.

🤖 Generated with Claude Code

[asubbarao]

duckstack/rules: success

• warn sql_file_too_long in skills/query/references/useful_queries.sql: Keep a job or loader file to 30-80 lines and one concern; split it, or state the reason and line count in the PR.
• warn sql_vague_cte_names in skills/query/references/useful_queries.sql: Name every CTE for what it holds (content_first_100, self_dispatch), not t / x / sd / tmp.

duckstack review queue, rule gate (hooks/hook_checks.sql)

[asubbarao]

duckstack/review: error

The reviewer gave no verdict. Its answer, cut at 3000 characters:

Opus 5.5, duckstack review queue (duckstack-review skill)

[asubbarao]

duckstack/rules: success

• warn sql_vague_cte_names in skills/query/references/useful_queries.sql: Name every CTE for what it holds (content_first_100, self_dispatch), not t / x / sd / tmp.
• warn sql_file_too_long in skills/query/references/useful_queries.sql: Keep a job or loader file to 30-80 lines and one concern; split it, or state the reason and line count in the PR.

duckstack review queue, rule gate (hooks/hook_checks.sql)

[asubbarao]

duckstack/review: failure

• skills/duckstack-review/queue.sql:0 review input (How to run it: the review runs on the lines the PR adds): The headless review prompt contained the skill, RULES.md and SQL-MANIFESTO.md but none of the PR's added lines, and the session had no file, shell or MCP tools to fetch them. With no code to read, none of checks 1–10 could run, and a clean verdict would have nothing behind it. Line 0 means no line number could be given.
  Rewrite: Have queue.sql put the PR's added lines in the prompt it sends to the headless review, as the receipt of a gh pr diff <n> read through shellfs (one dispatched statement per queued PR), placed after the skill text. Then run this review again on the same PR head.

Opus 5.5, duckstack review queue (duckstack-review skill)

[asubbarao]

duckstack/review: failure

• skills/agent-dispatch/SKILL.md:176 RULES.md 3 and 5 (author's reason does not hold): The stated reason for keeping the inputs in the script ("a self-dispatched Python side returns receipts (status, body as text), not the script's columns") contradicts RULES.md 5: what you build in the self-dispatch comes back as a column when you unnest the array. RULES.md 3 also says self-dispatch is the escape hatch for exactly this column-into-reader binder limit, so the reason does not justify the pattern.
  Rewrite: Replace lines 174–177 with: "The inputs are rows in the first CTE. Each one renders a one-input Python statement, which is self-dispatched beside its duckdb_answer, and the receipts are unnested back into columns (r.receipt.*). An input the proposal raises on comes back as its own receipt with the error."
• skills/agent-dispatch/SKILL.md:161 check 1 / RULES.md 6 and 4: A Python for loop with a per-row try/except runs the proposal once per input. That is a loop in a script doing per-row work that should be one dispatched statement per row, where a failing input comes back as its own receipt with its error kept.
  Rewrite: Make the Python a one-input program: print(json.dumps(dict(python_answer=<proposal>({{ raw | json_encode() }})))), rendered per intent row and sent with array_agg(selfdispatch(sql)). Do not write a for raw in [...] loop or a try/except that builds error rows. An input the proposal raises on comes back as a failed receipt that carries the exception, joined to its raw and duckdb_answer in the same struct.
• skills/agent-dispatch/SKILL.md:158 RULES.md 7 / check 4 (no intent CTE): The first CTE is python_side, and the inputs are hidden as [<inputs>] inside a Python heredoc. A reader of the SQL cannot see what is being compared, and the DuckDB side only gets its inputs back from whatever Python re-serialises as raw.
  Rewrite: WITH intent AS ( -- the inputs, declared once as rows
  SELECT unnest(['', '', '']) AS raw
  ), rendered AS (
  SELECT raw, AS duckdb_answer,
  tera_render($tera$
  SELECT * FROM read_json($cmd$/opt/homebrew/bin/uv run --quiet --no-project python - <<'PY'
  import json
  print(json.dumps(dict(python_answer=({{ raw | json_encode() }}))))
  PY
  |$cmd$, format := 'newline_delimited')
  $tera$, i::JSON, autoescape := false) AS sql
  FROM intent AS i
  ), dispatched AS (
  SELECT array_agg({raw: raw, duckdb_answer: duckdb_answer, receipt: selfdispatch(sql)}) AS receipts FROM rendered
  )
  SELECT r.raw, r.duckdb_answer, r.receipt.*
  FROM dispatched CROSS JOIN UNNEST(receipts) AS t(r)

Opus 5.5, duckstack review queue (duckstack-review skill)

[asubbarao]

duckstack/rules: success

• existing sql_file_too_long in skills/query/references/useful_queries.sql: Keep a job or loader file to 30-80 lines and one concern; split it, or state the reason and line count in the PR.
• existing sql_vague_cte_names in skills/query/references/useful_queries.sql: Name every CTE for what it holds (content_first_100, self_dispatch), not t / x / sd / tmp.

duckstack review queue, rule gate (hooks/hook_checks.sql)

[asubbarao]

duckstack/review: failure

• skills/agent-dispatch/SKILL.md:166 RULES.md 7 / check 4 (intent CTE), RULES.md 16-17: The proposal is the one thing line 185 tells the reader to swap in, but it is not declared in the first CTE. It is buried in the third quoting layer: $py$ wraps $cmd$, which wraps a heredoc, which wraps Python. To change it, the reader has to edit inside nested quotes instead of changing a value in intent.
  Rewrite: WITH intent AS ( -- the inputs and the proposal under test, declared once
  SELECT unnest(['2026-10-18', '2026-02-30', $v$O'Brien$v$]) AS raw,
  'datetime.date.fromisoformat(value).isoformat()' AS proposal
  ), rendered AS ( -- per input: the DuckDB answer, and the proposal as a one-input statement
  SELECT raw, try_strptime(raw, '%Y-%m-%d')::DATE::VARCHAR AS duckdb_answer,
  tera_render($tera$SELECT * FROM read_json($cmd$/opt/homebrew/bin/uv run --quiet --no-project python - <<'PY'
  import json, datetime
  value = {{ raw | json_encode() }}
  print(json.dumps(dict(python_answer={{ proposal }})))
  PY
  |$cmd$, format := 'newline_delimited')$tera$, i::JSON, autoescape := false) AS sql
  FROM intent AS i
  ) -- 9 lines before, 10 after: the template only substitutes; the swap is one value in intent

Opus 5.5, duckstack review queue (duckstack-review skill)