Second instance in one session of the same mistake, in the same file, so it
wants an entry rather than a shrug.
A gha caller stub is mostly commented-out inputs. Reading one and finding no
use-ai-config: line invites the conclusion that the feature is off. The file
supports the opposite conclusion: an omitted input means the callee's default
applies, and use-ai-config defaults to true in both claude.yml and
claude-code-review.yml at @v2.
Measured 2026-08-28. A session filed UCD-SERG/shigella#36 claiming the reviewer
there "works from whatever prose the repo happens to carry" because neither
caller passed the input. The shared ai-config corpus had been loading on every
run all along. Two PRs shipped against that premise, and the change is a
behavioural no-op --- worth having, since it records intent and survives a
default moving, but not the fix the issue described.
The recurrence is the useful part. The entry added in #2504 came from the same
session and the same @v2 file, also from assuming a default rather than
reading one (lint-markdown's companion toggles). That session had opened
claude-code-review.yml at @v2 three times that evening --- for the
ANTHROPIC_API_KEY secret, then for the companion defaults --- and never
scrolled to this input. So the failure is not a missing file; it is asking the
file only the question already in mind.
Proposed remedy in the entry: when a caller is under review, dump every input's
default at once rather than looking up the one you came for.
Second instance in one session of the same mistake, in the same file, so it
wants an entry rather than a shrug.
A gha caller stub is mostly commented-out inputs. Reading one and finding no
use-ai-config:line invites the conclusion that the feature is off. The filesupports the opposite conclusion: an omitted input means the callee's default
applies, and
use-ai-configdefaults totruein bothclaude.ymlandclaude-code-review.ymlat@v2.Measured 2026-08-28. A session filed UCD-SERG/shigella#36 claiming the reviewer
there "works from whatever prose the repo happens to carry" because neither
caller passed the input. The shared ai-config corpus had been loading on every
run all along. Two PRs shipped against that premise, and the change is a
behavioural no-op --- worth having, since it records intent and survives a
default moving, but not the fix the issue described.
The recurrence is the useful part. The entry added in #2504 came from the same
session and the same
@v2file, also from assuming a default rather thanreading one (
lint-markdown's companion toggles). That session had openedclaude-code-review.ymlat@v2three times that evening --- for theANTHROPIC_API_KEYsecret, then for the companion defaults --- and neverscrolled to this input. So the failure is not a missing file; it is asking the
file only the question already in mind.
Proposed remedy in the entry: when a caller is under review, dump every input's
default at once rather than looking up the one you came for.