Actions Fleet Monitor has crashed on every run since 2026-09-21: legacy_attr_file is used before it is assigned #2279

Problem

The daily "Actions Fleet Monitor" workflow (actions-fleet-monitor.yml) has ended in failure on every scheduled run since 2026-09-21. Its last successful run was 2026-09-20 11:36 UTC. No fleet monitor report has been produced for three weeks.

From run 38141887715 (2026-10-11 13:06 UTC):

  petry-projects/.github
scripts/fleet_monitor.sh: line 221: legacy_attr_file: unbound variable
jq: error: writing output failed: Broken pipe
##[error]Process completed with exit code 1.

Cause

scripts/fleet_monitor.sh runs under set -euo pipefail. legacy_attr_file is appended to at line 221, inside the first per-repository loop (lines 177 to 308), but it is only assigned at line 553 (legacy_attr_file=$(mktemp)). The first repository that has a non-ingress workflow with runs aborts the script. Introduced by commit a1ccf3b ("Retarget run attribution to job/role level", #1789), merged 2026-09-20, which matches the first failing day.

No test runs the script far enough to reach line 221, so CI stayed green.

Why it matters now

• "A weekly Claude cap parses as 'unknown reset'" (A weekly Claude cap parses as 'unknown reset', so the retry cron re-dispatches every 2 hours against a block that can last 7 days #1863) is waiting on the report's "Unparsed rate-limit messages" section, which has not been produced since before that section could be read.
• "Three monitors misread a paused run" (Three monitors misread a paused run: the initiative canary passes, the pr-review canary fails, and the fleet monitor counts it as a success #2266) changes the same script; its fix cannot be observed live while the script crashes first.
• The failure itself went unnoticed for three weeks: a monitor that fails every day raised no alert.

Scope

1. Create legacy_attr_file before the loop that appends to it; check every other temp-file variable in the script for the same ordering fault.
2. A regression test that runs the script end to end against a stubbed gh with at least one non-ingress workflow that has completed runs, and fails on an unbound variable.

Acceptance criteria

1. The next scheduled run of "Actions Fleet Monitor" after the fix ends in success and posts its report.
2. The regression test fails when the late assignment is restored.
3. A follow-up is filed (or a decision recorded here) on why a daily monitor could fail for three weeks without an alert.