Stop /ship's docs audit halting when the child writes prose after its result (GST-2349) (#4285)

* Stop /ship's docs audit halting when the child writes prose after its result (GST-2349)

The Step 14.5 documentation audit read only the child's last nonempty line.
Children that ended with their JSON result and then restated the documentation
section in prose had a valid "current, no edits" audit rejected with no format
repair, which spent an attempt each time and, after two, stopped /ship for an
owner risk grant (founderstar #837, harness-evals #393).

gstack-docs-candidate result now reads the last line that opens a JSON object,
ignores prose after it and says so with TRAILING_PROSE:. A result missing only
documentation_section keeps the existing free re-ask, whose prompt now tells the
child to put the section inside the JSON and write nothing after the line; the
re-audit retry note says the same. Content failures still count.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* Send any docs-audit result with text after it back for one free re-ask (GST-2349)

A cross-family audit of the previous commit showed that ignoring prose after
the result line let a contradicting verdict through: a blocked result followed
by a fenced "current" example was accepted as current, and a "Correction: the
audit is BLOCKED" after a current line became a repair that had to stay current.

Text after a JSON line carrying the audit id is no longer ignored and never
searched for another result. It is a format slip: FORMAT_REPAIR re-asks once
for the whole result line without spending an attempt, the re-ask counts
everything the audit wrote after its JSON line, and the reply is judged on its
own with nothing kept from the first line. A reply that itself has text after
its line counts as an attempt.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* Keep a blocked docs audit blocked through the free re-ask (GST-2349)

Review of the previous change found that the free whole-line re-ask could
turn a blocked audit into a passing one: a blocked result line, then a fenced
"current" example, then a "current" reply was accepted as current.

The re-ask now works one way. It is offered only when every JSON line that
carries the audit id is free of content problems; a content problem such as a
missing status or "current" with edits still spends an attempt. If any of those
lines said blocked or listed a blocker, the reply must stay blocked and keep
every blocker (it may add more), and the re-ask prompt says so.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

---------

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
This commit is contained in:
Garry Tanandcapy-ai[bot] authored and GitHub committed 2026-11-17 09:36:21 +00:00
1 parent bcece3fbcd
commit 2af3822f94
8 files changed
+244 -27

No files matched your search

+6 -1
View File
@@ -7289,7 +7289,12 @@ not an attempt: send `gstack-docs-candidate repair-prompt <output> --audit-id <i
to one fresh agent with no tools (never the audit child, which can still edit), save its one-line
reply and run `result <reply> --audit-id <id> --out-dir <dir> --repair-of <output>`.
The reply passes only if every value the first line gave correctly is
unchanged, and each audit gets one repair. Otherwise run Step 14.5's one
unchanged, and each audit gets one repair. Text after a JSON line that carries
the audit id (`text follows the JSON result line`) gets the same free re-ask for
the whole line, but only when every such line is free of content problems (a
content problem counts as an attempt). A blocked status and its blockers on any
such line must survive the reply, which may add blockers; nothing else from the
earlier line is kept or accepted. Otherwise run Step 14.5's one
remaining re-audit with the printed `RETRY_NOTE:` text appended to its prompt,
word for word; it quotes the exact keys. With no attempt left, Step 14.5 stops
and reports this code and the keys it names