[/tdd] This fix directly addresses a production outage (15 consecutive failures over ~8h) but ships with no regression test. A future degradation event could silently break the retry logic with no automated signal.

@copilot please address this.

2>&1 contaminates candidate_file on success: pr_list_output contains merged stdout+stderr, so any warning gh emits to stderr on a successful run silently corrupts the JSON written to candidate_file.

set +e
pr_list_stderr_tmp=$(mktemp)
pr_list_output=$(gh pr list --repo "$EXPR_GITHUB_REPOSITORY" \
  --state open \
  --search "is:pr is:open -is:draft sort:updated-desc" \
  --limit 30 \
  --json number,title,url,headRefOid,headRefName,updatedAt,author,mergeStateStatus \
  2>"$pr_list_stderr_tmp")
pr_list_status=$?
pr_list_stderr=$(cat "$pr_list_stderr_tmp"; rm -f "$pr_list_stderr_tmp")
set -e

[/diagnosing-bugs] The bare EOF token matches any error message containing that substring (e.g. Unexpected end-of-file, JSON parse errors, script errors), which could cause the code to silently retry what is actually a deterministic failure.

elif echo "$pr_list_output" | grep -qE "HTTP (5[0-9]{2}|429)|context deadline exceeded|connection reset|unexpected EOF|read: EOF"; then

@copilot please address this.

[/diagnosing-bugs] If gh pr list exits with a non-zero status but produces no output (e.g., hard kill, OOM, runner timeout), $pr_list_output is empty — the grep -qE match fails silently and the code takes the else branch, exiting immediately with a generic error instead of retrying.

elif [ -z "$pr_list_output" ] || echo "$pr_list_output" | grep -qE "HTTP (5[0-9]{2}|429)|context deadline exceeded|connection reset|EOF"; then

@copilot please address this.

Bare EOF match is too broad: the pattern EOF will match any output line containing that substring (e.g., "EOF in configuration", "PROOF", "SIZEOF"). This could cause legitimate non-transient errors to be silently retried rather than fail-fast.

Consider tightening to:

• \bEOF\b (word boundary) — avoids substring false positives
• or unexpected EOF — matches only the Go net/http error text that actually indicates a truncated response

@copilot please address this.

[/diagnosing-bugs] When the final attempt hits a transient error, the inner if [ "$pr_list_attempt" -lt "$pr_list_max_attempts" ] is false — so no log message is printed and no sleep is needed. But the fi just falls through, the loop condition fails, and control lands on the post-loop failure message. This is correct but the last transient error output is only shown in the post-loop message, which makes it hard to tell whether the last attempt was transient or not.

elif echo "$pr_list_output" | grep -qE "..."; then
  echo "Transient error on attempt $pr_list_attempt/$pr_list_max_attempts (exit $pr_list_status); ${backoff}s..." >&2
  echo "$pr_list_output" >&2
  if [ "$pr_list_attempt" -lt "$pr_list_max_attempts" ]; then
    backoff=$((2 ** pr_list_attempt))
    sleep "$backoff"
  fi

@copilot please address this.

Silent exit on final transient failure: when the 4th attempt hits a transient error, the inner guard [ "$pr_list_attempt" -lt "$pr_list_max_attempts" ] is false, so the if block is skipped — no log is emitted, no sleep, the loop simply ends, and the post-loop failure handler fires with no diagnostic output showing whether the last error was transient or not.

elif echo "$pr_list_output" | grep -qE "HTTP (5[0-9]{2}|429)|context deadline exceeded|connection reset|EOF"; then
  echo "Transient error on attempt $pr_list_attempt/$pr_list_max_attempts (exit $pr_list_status)" >&2
  echo "$pr_list_output" >&2
  if [ "$pr_list_attempt" -lt "$pr_list_max_attempts" ]; then
    backoff=$((2 ** pr_list_attempt))
    echo "Retrying in ${backoff}s..." >&2
    sleep "$backoff"
  fi

Observability gap on the last attempt: when pr_list_attempt equals pr_list_max_attempts, the transient-error branch is entered but the inner if [ "$pr_list_attempt" -lt "$pr_list_max_attempts" ] guard is false, so no diagnostic is logged and no exit occurs. The loop silently ends and the post-loop if catches it. This makes log triage harder because the retry count message and the final error output are separated, with nothing indicating the last attempt was also transient.

Suggestion — always log, only gate the sleep:

elif echo "$pr_list_output" | grep -qE '...'; then
  echo "Transient error on attempt $pr_list_attempt/$pr_list_max_attempts (exit $pr_list_status)" >&2
  echo "$pr_list_output" >&2
  if [ "$pr_list_attempt" -lt "$pr_list_max_attempts" ]; then
    backoff=$((2 ** pr_list_attempt))
    echo "Retrying in ${backoff}s..." >&2
    sleep "$backoff"
  fi
fi

@copilot please address this.