Observed repeatedly in Lacaedemon/sparta (a quarto-publish/review REVDEPS consumer) during a
high-throughput PR session on 2026-07-02, tracked there as
Lacaedemon/sparta#590:
Symptom
A push-triggered claude-review run completes green but posts only a stub comment
("I'll wait for the background agents..." or "needs your approval") with no review verdict.
The job's exit status gives no signal that no review happened.
- sparta PR #583: both push-triggered runs stubbed; the two real reviews only came from
workflow_dispatch re-triggers.
- sparta PR #569: 3 of 5 runs stalled this way ("waiting for background agents" /
"needs your approval" stubs) while the job still passed.
- sparta PR #589: three consecutive stalls without a verdict; the fallback was a
self-review posted by the session driving the PR.
- sparta PR #578 (the original #590 report): same pattern, while sibling PRs reviewed fine.
workflow_dispatch re-triggers succeed far more reliably than push-triggered runs, which
suggests something about the push/synchronize event context (or a concurrency/approval
prompt in that path) rather than the review prompt itself.
Why it matters
Downstream automation treats "review job green" as "a review verdict exists". A green
stub breaks that: an ARDI loop waits on a verdict that never arrives (or worse, a
stale prior verdict gets read as current). A run that posts no verdict should fail, or
at minimum the workflow should detect the stub and retry once.
Possible root
May ultimately sit in anthropics/claude-code-action rather than this repo's reusable
workflow - the stub texts look like the action's agent stalling on an approval prompt.
Filing here first since the reusable workflow is where a stub-detect-and-retry (or
fail-on-no-verdict) guard could live regardless; if it reproduces cleanly against the
action itself, an upstream report there needs the owner's sign-off first.
Observed repeatedly in Lacaedemon/sparta (a quarto-publish/review REVDEPS consumer) during a
high-throughput PR session on 2026-07-02, tracked there as
Lacaedemon/sparta#590:
Symptom
A push-triggered claude-review run completes green but posts only a stub comment
("I'll wait for the background agents..." or "needs your approval") with no review verdict.
The job's exit status gives no signal that no review happened.
workflow_dispatch re-triggers.
"needs your approval" stubs) while the job still passed.
self-review posted by the session driving the PR.
workflow_dispatch re-triggers succeed far more reliably than push-triggered runs, which
suggests something about the push/synchronize event context (or a concurrency/approval
prompt in that path) rather than the review prompt itself.
Why it matters
Downstream automation treats "review job green" as "a review verdict exists". A green
stub breaks that: an ARDI loop waits on a verdict that never arrives (or worse, a
stale prior verdict gets read as current). A run that posts no verdict should fail, or
at minimum the workflow should detect the stub and retry once.
Possible root
May ultimately sit in anthropics/claude-code-action rather than this repo's reusable
workflow - the stub texts look like the action's agent stalling on an approval prompt.
Filing here first since the reusable workflow is where a stub-detect-and-retry (or
fail-on-no-verdict) guard could live regardless; if it reproduces cleanly against the
action itself, an upstream report there needs the owner's sign-off first.