memories: an omitted caller input carries the callee's default - #2509
Conversation
A gha caller stub is mostly commented-out inputs, so an absent line reads as a disabled feature. It means the opposite: the callee's default applies, and use-ai-config defaults to true in both Claude workflows at @v2. Measured on UCD-SERG/shigella#36, which claimed the reviewer worked without the shared corpus because neither caller passed the input. It had been loading all along, and the two PRs that shipped against that premise are a behavioural no-op. Recorded with the recurrence rather than alone: this is the second entry in this file from one session, both from assuming a default instead of reading one, and that session had already opened the same @v2 file three times for other questions. The remedy in the entry is therefore to dump every input's default at once when a caller is under review. Closes #2508 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HYDWsGL8cYC3ptLEL4Wayn
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
Three findings, all correct. The link read [#36] while pointing at pull/38 -- #36 is the issue, already linked earlier in the same paragraph, so a reader following the link would land somewhere the text did not promise. 'this file' and 'the same file' four words apart meant the memory file and the @v2 workflow file respectively. Named the second one. 'and a duplicate' is dropped rather than pinned down. Both use-ai-config commits on shigella's main (dc87f8e and 41bf0e7) report as belonging to PR #38, so there is no second PR to name, and the claim is not load-bearing for the entry. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HYDWsGL8cYC3ptLEL4Wayn
Dispositions --- all three Addressed in
|
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
|
Claude finished review — View run Confirmed: the merge commit (current PR head) only pulled in unrelated Since no content in this PR's actual diff changed since the last clean verdict, this is a confirming review with no new findings. Review: #2509Scope: Single new entry appended to The current head commit ( The most recent substantive review of this PR's content is the second round posted 2026-08-28T08:05:23Z (exact comment URL not independently resolvable from this environment, but identifiable by its timestamp and Since the file under review has not changed since that round, that verdict still stands and I have no new findings to add. VerdictReady for merge Reviewed commit: 35d2967 |
1 participant
Closes #2508
One entry appended to
memories/gha-reusable-workflows.md.A gha caller stub is mostly commented-out inputs, so finding no
use-ai-config:line invites the conclusion that the feature is off. The filesupports the opposite: an omitted input means the callee's default applies, and
use-ai-configdefaults totruein bothclaude.ymlandclaude-code-review.ymlat@v2.Measured 2026-08-28 on
UCD-SERG/shigella#36, which
claimed the reviewer there "works from whatever prose the repo happens to
carry" because neither caller passed the input. The shared
ai-configcorpushad been loading on every run all along, so the two PRs that shipped against
that premise are a behavioural no-op --- worth having, since they record intent
and survive a default moving, but not the fix the issue described.
Recorded with its recurrence rather than alone. The entry added in #2504
came from the same session and the same
@v2file, also from assuming adefault instead of reading one. That session had opened
claude-code-review.ymlat@v2three times that evening --- for theANTHROPIC_API_KEYsecret, then forlint-markdown's companion defaults ---and never scrolled to this input. The failure is not a missing file; it is
asking the file only the question already in mind, which is why the entry's
remedy is to dump every input's default at once when a caller is under review
rather than to look up one input more carefully.
scripts/semantic-line-breaks.py --base origin/mainreports 0 paragraphs wouldchange; no non-ASCII punctuation added;
check-memory-file-size.pyclean.