Terminal
Terminal/Skill
Groups recent failed Nx Cloud CI pipelines into recurring failure patterns, without changing anything, so you see what breaks most.
By Nx
Works with
npx skills add nrwl/nx-ai-agents-config --skill nx-cloud-ci-failure-reviewRequires the nx-cloud-api skill for all Nx Cloud reads. This workflow is read-only.
Reduce failed pipelines to a small set of diagnostic groups. A failed pipeline does not prove that an agent or platform caused the failure. Classify each group only at the level the evidence supports.
Use focused investigation mode when the user supplies one or more CIPE or run-group links. Fetch those exact records first. Follow their returned run-group, run, task, and workflow links. Compare the supplied executions directly. Do not run a broad time window or build a workspace-wide ranking unless the user asks for recurrence or scope. Keep cancellations that follow the named critical error as consequences, not separate root-cause groups.
Use review mode when the user asks for a workspace failure review. Review
failures from the last 24 hours by default. CIPE and run-group list filters use
createdAt, not failure time. Query a 30-hour creation window to cover the maximum
normal six-hour CI duration. State the intended 24-hour review period and the
30-hour creation window.
Query statuses=FAILED, statuses=TIMED_OUT, and statuses=CANCELED (the
cipes filter accepts only terminal statuses). State the
branches, CI contexts, statuses, endpoints, and record counts. Ask before widening
the window or mixing unrelated repositories/workspaces.
Use nx-cloud-api (npx nx-cloud api) for every read. Inspect the live
schemas with --describe for cipes, run-groups, run-groups/{runGroupId},
runs, runs/{runId}/tasks, and the linked steps and instances endpoints.
Use --list-operations to confirm the routes; the public API has no separate
workflow routes. Reach child records by following returned links. Start with supplied CIPEs in focused investigation mode. Start with failed CIPEs
and their linked run groups in review mode. Use only live documented filters and
fields. Fetch one narrow page first and check nextCursor before continuing.
Collect in this order:
--pages value. Rank groups workspace-wide
only after a zero exit with nextCursor: null; otherwise
report the data as incomplete.Use a semaphore of at most five narrow detail requests. Fetch logs only after a group reaches the top five. Tell the user before reading CI log content. Quote only redacted error lines.
Treat cancellations as their own category until their reason and timing show that they belong elsewhere. Do not call a cancellation a failure without that context.
Build groups from stable evidence. Prefer a task identifier plus normalized error signature. Normalize timestamps, GUIDs, retry counters, temporary paths, and random ports. Keep stack context, command, and error code that distinguish two otherwise similar messages.
Use these categories only when their evidence exists:
Call a group workflow or agent failure only when workflow/instance status, critical error, or agent evidence supports it and no task failure explains the pipeline. Do not infer infrastructure from the absence of a task log.
Reconcile the full pipeline picture before classification. A task can record
SUCCEEDED or a cache hit while the main job fails during artifact download,
validation, or restoration. Compare the run-group criticalErrorMessage, task
status, cache status, workflow status, and later cancellations. In that case,
classify the main-job cache restoration error as primary and the canceled runs as
consequences.
For every group, record its CIPE count, run-group count, affected task count, first and last occurrence, branches/contexts, and one or two representative IDs. Rank first by affected pipeline count, then recurrence, then blocked duration when it is documented.
CODEOWNERS and project ownership metadata at the recorded commit. Suggest a
likely review owner. Do not assign work or claim ownership when that evidence is absent.Do not make code, workflow, or allocation changes without explicit approval.
# Failed CI review
## Scope
- Window: …
- Data: …
## Failure groups
### 1. <evidence-level category> — <short signature>
- **Confirmed facts:** affected CIPEs/runs/tasks; representative IDs; log excerpt.
- **Assessment:** **Confirmed** / **Likely** / **Needs diagnostic run** — …
- **Impact:** …
- **Highest-leverage next action:** …
## Unclassified or incomplete data
- …
Limit a review report to five groups. In focused investigation mode, report the supplied executions as one comparison when they have the same normalized signature. Link each finding to exact CIPEs, run groups, runs, and tasks where available. State the data limit when pagination, missing assets, or missing context affects the conclusion.