We use anthropics/claude-code-action's claude-code-review.yml workflow on the nedati-technologies/claude-wow-plugin repo. Across May 2026 (PRs #62-#74) we've observed two recurring failure modes producing ~70% review-failure rate on otherwise-passing PRs. Both are deterministic + reproducible.
Mode A: gh pr view invoked without --repo
In some workflow code paths, the action calls gh pr view <NUMBER> without --repo $OWNER/$REPO. When the action runs in a workflow context where the default repo isn't auto-detected (or where the working directory is not the repo root), gh errors out with:
no default remote repository or repository is not specified
The action then exits non-zero and no review is posted, despite the action succeeding at all other steps.
Suggested fix: always pass --repo "${GITHUB_REPOSITORY}" to gh pr view / gh pr list / gh issue invocations inside the action. The GITHUB_REPOSITORY env var is provided automatically in GitHub Actions context.
Mode B: haiku dedup skips reviews with no specific findings
The action's "haiku pre-check" stage is supposed to dedupe identical review outputs, but on small/trivial PRs (1-2 file changes, no real issues to flag) it appears to mis-classify the absence of findings as a "duplicate" and skips posting the review entirely. Result: PR author gets no review at all, even a "no issues found" stub would be useful.
Reproduction: PRs #72/#73/#74 on the cited repo are doc-heavy / config-light PRs where the previous PR's "no issues found" review and the current PR's "no issues found" review both look identical to haiku-dedup, so the second is suppressed.
Suggested fix: either (1) skip the dedup pass when the LLM's verdict is the no-issues stub, or (2) include the PR number / commit SHA in the dedup hash so trivially-identical outputs across distinct PRs aren't deduped.
Repro environment
- GitHub Actions: ubuntu-latest
- claude-code-action: latest at time of failure (May 2026)
- Repo size: ~120 source files, ~70 markdown files
Happy to provide failed-run URLs if useful; please ping back.
We use
anthropics/claude-code-action'sclaude-code-review.ymlworkflow on the nedati-technologies/claude-wow-plugin repo. Across May 2026 (PRs #62-#74) we've observed two recurring failure modes producing ~70% review-failure rate on otherwise-passing PRs. Both are deterministic + reproducible.Mode A:
gh pr viewinvoked without--repoIn some workflow code paths, the action calls
gh pr view <NUMBER>without--repo $OWNER/$REPO. When the action runs in a workflow context where the default repo isn't auto-detected (or where the working directory is not the repo root),gherrors out with:The action then exits non-zero and no review is posted, despite the action succeeding at all other steps.
Suggested fix: always pass
--repo "${GITHUB_REPOSITORY}"togh pr view/gh pr list/gh issueinvocations inside the action. TheGITHUB_REPOSITORYenv var is provided automatically in GitHub Actions context.Mode B: haiku dedup skips reviews with no specific findings
The action's "haiku pre-check" stage is supposed to dedupe identical review outputs, but on small/trivial PRs (1-2 file changes, no real issues to flag) it appears to mis-classify the absence of findings as a "duplicate" and skips posting the review entirely. Result: PR author gets no review at all, even a "no issues found" stub would be useful.
Reproduction: PRs #72/#73/#74 on the cited repo are doc-heavy / config-light PRs where the previous PR's "no issues found" review and the current PR's "no issues found" review both look identical to haiku-dedup, so the second is suppressed.
Suggested fix: either (1) skip the dedup pass when the LLM's verdict is the no-issues stub, or (2) include the PR number / commit SHA in the dedup hash so trivially-identical outputs across distinct PRs aren't deduped.
Repro environment
Happy to provide failed-run URLs if useful; please ping back.