[asubbarao]

Why

Dev aborted twice on 2026-10-21 (13:57 and 13:58), each time about 60 s after boot, while blocks_refresh posted the parsed view over the transcripts. The server log shows TeraRenderFunc under a CASE in a projection, then panicked at src/lib.rs:135:50 … NulError and thread caused non-unwinding panic. aborting. The cause is server/blocks.sql line 50: it joined a tool result's text parts with tera_render('{{ parts | join(sep="\n") }}', …). A part with a NUL byte kills the server; the template never did anything a list join does not.

Change

One expression: array_to_string(text_parts, chr(10)) in place of the tera join. The comment names the reason.

Tested

Local duckdb :memory: with the tera extension loaded:

• tera_render on ['a', 'b' || chr(0) || 'c'] → NulError(3, [97, 10, 98, 0, 99]), process exit 134. array_to_string on the same list → 5 chars, NUL kept.
• tera_render(...) = array_to_string(parts, chr(10)) is true for ['a','b'], a multi-line part with a quoted part, ['é', ''] and ['one'].

🤖 Generated with Claude Code

[asubbarao]

duckstack/review: error

The review did not start; the next tick starts it again. Dev's answer:

Invalid Input Error: Pipe process exited abnormally code=1: cd '/Users/aloksubbarao/.duck/deploy' && PR_BASE=$(git -C /Users/aloksubbarao/duckdb-skills merge-base origin/main d79b5c9) PR_REPO='/Users/aloksubbarao/duckdb-skills' PR_GH='/opt/homebrew/bin/gh' PR_GH_REPO

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

[asubbarao]

Closed on Alok's decision: the NUL byte came from an agent reading a binary file (pdf.duckdb_extension.info) with cat into its transcript. The fix is that agents read binary files with read_blob, never as text; not a change to this join. — Claude Fable 5.1

[asubbarao]

Reopened on Alok's call: no DuckDB reader breaks on a NUL byte (read_json, read_csv and array_to_string all carry it, tested); only the tera_render join in blocks.sql aborts the server on it. The query is the fault. — Claude Fable 5.1

[asubbarao]

duckstack/review: failure

• server/blocks.sql:50 RULES.md 15: The fix uses array_to_string, which rule 15 bans by name (the hook is meant to block it). The NUL-byte panic is a good reason to stop using tera_render, but it doesn't justify array_to_string: rule 15's other option, string_agg, does the same join without tera.
  Rewrite: coalesce(CASE WHEN len(text_parts) > 0 THEN list_aggregate(text_parts, 'string_agg', part_separator) END,
  CASE WHEN json_type(b) = 'VARCHAR' THEN b ->> '$' END, [b ->> k FOR k IN content_keys IF b ->> k IS NOT NULL][1]) AS message_content,
• server/blocks.sql:50 RULES.md 12, RULES.md 7: The separator is built as chr(10) in the middle of an expression. Rule 12 bans that, and rule 7 says values like this are declared up front in the first CTE.
  Rewrite: -- in the line CTE, beside part_keys:
  E'\n' AS part_separator,
  -- in parsed, add it to the EXCLUDE list:
  SELECT * EXCLUDE (b, content_keys, id_keys, part_keys, message_types, part_separator), ...
  -- and join with it:
  list_aggregate(text_parts, 'string_agg', part_separator)

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

[asubbarao]

Closed as a duplicate of #101, which strips the NUL with replace(…, chr(0), '') as Alok said. — Claude Fable 5.1