onnxruntime

094d02ef - Add /ort-ci skill for triaging and re-running CI checks (#29721)

Add /ort-ci skill for triaging and re-running CI checks (#29721) ### Description Adds a new `/ort-ci` agent skill that documents how to trigger, re-run, and unblock ONNX Runtime CI checks on a GitHub pull request, plus a helper script that safely re-runs only genuinely failed/cancelled GitHub Actions workflows. The skill emphasizes diagnosing each failure before re-running, since most CI failures require a code change and re-running them just fails again. ### Key Changes | File | Change | |---|---| | `.agents/skills/ort-ci/SKILL.md` | New skill documenting ORT CI triage and management: classifying a check's provider, gathering PR/check context, per-failure triage (download log -> classify -> fix or re-run), and targeted playbooks. | | `tools/scripts/rerun_failed_ci.sh` | New helper that re-runs a PR's failed/cancelled GitHub Actions runs for the current head commit, skipping workflows that already have a newer queued/in-progress run to avoid duplicate runs. Supports `--dry-run` and auto-detects `owner/repo`. | | `AGENTS.md` | Adds a `CI` section pointing to the new `/ort-ci` skill. | The skill covers the current ORT CI topology (as of 2026-07): nearly all checks run as GitHub Actions; only `Linux Android Emulator QNN CI Pipeline` remains on Azure Pipelines, plus the bot-driven `license/cla` and `Python format` checks. It includes step-by-step playbooks for: - Re-running transient GitHub Actions failures (via `rerun_failed_ci.sh`) - Unblocking `license/cla` - Fixing the `Python format` required check - Triggering the Azure Pipelines QNN Android Emulator check (`/azp run` and direct fallback) - Fixing the `Windows GPU Doc Gen CI` check when operator docs are out of date - Safety rules and post-trigger verification ### Testing Notes - Documentation- and tooling-only change; no product code or runtime behavior is affected. - `rerun_failed_ci.sh` was exercised in `--dry-run` mode against live PRs (e.g. `./rerun_failed_ci.sh <pr> --dry-run`) to confirm it correctly lists failed/cancelled workflows and triggers nothing. - The script requires an authenticated `gh` CLI and uses only `gh`'s built-in `--jq` (no external `jq` binary needed).