reconcile: agami's answer may be several statements (ACE-135)

[sandeep-agami]

Spec: ACE-135 (agami-sdlc projects/agami-core-extension/specs/reconcile-evidence/ACE-135-several-statements.md)

Stacked on #320 (ACE-134). Round 9 of the reconcile stack.

The ask

"we need to handle when agami fires multiple queries to answer a question and show it appropriately in the report." The cold client's reply was read by _first_json_object and sql had to be one string: a list under sql, two statements in one string, or a second JSON object lost everything but the first object without a word, or read as unreadable. The card had one sql_agami and no place for the others.

The change

• golden_run.py: GeneratedSql.statements (every statement, in order; sql is the last). _spawn accepts sql as a string or a list; a string is cut at its top-level semicolons (outside quotes and comments; text splitting only, no regeneration per ACE-093); a list carrying anything but statements is unreadable as a whole. The question prompt gains: "If answering takes more than one query, put them in order in a list under sql; the last must be the statement whose result answers the question." The golden run reads sql as before.
• run_golden_eval.py --ask / --ask-file: the payload carries statements (a generator that did not say gives [sql]).
• SKILL.md 2b: only sql goes to agami.sql and runs; the others go into the record's agami_statements (2d, after words) and are never run: the read-only rule refuses anything but a SELECT, so an earlier statement can only be a look at the data.
• reconcile.py report-items: sql_agami_steps when there was more than one; the rows and answer checks carry "agami ran N queries; the last one's result is compared". Renderer validates the field; the page's SQL block lists agami's statements numbered, the last marked "compared". CHANGELOG under Added.

Tests

test_golden_run.py: a list keeps every statement and answers with the last; two statements in one string are cut at the top-level semicolons and not inside a literal or a comment; one statement reads (sql,); an empty list, a list with a non-string, or a string of semicolons is unreadable; the prompt sentence. test_run_golden_eval_ask.py: the payload's statements (one existing pin gains the key). test_reconcile_report_items.py: the steps and the note, and their absence for one statement. test_render_reconcile_report.py: the SQL block and the field's validation. test_reconcile_learning_loop_skill.py: the 2b and 2d prose. 224 passed across the golden-run, ask-door, reconcile card, renderer, skill-pin and eval-skill suites (the client test is the known local failure).

🤖 Generated with Claude Code