Repository CI/CD lifecycle monitoring¶
MAC treats CI/CD as a continuation of repository work. Publication still completes when reviewed work is integrated and remotely verified; CI health is checked afterward and can create linked maintenance work without rewriting the historical publication result. This describes the generic post-publication monitor, not MAC's release gate: MAC candidate CI runs the full suite up front, and its image qualification and documentation publication wait for their required checks. See testing strategy.
Why this exists¶
A read-only audit of the authenticated GitHub notification feed on 2026-07-24 found 1,336 accumulated threads. Of those, 1,304 (97.6%) were CI activity and 1,302 were still unread. The alerts spanned 15 repositories from 2026-05-09 through 2026-07-24; 989 concerned a repository's main/default branch.
The largest backlogs were:
| Repository | CI notifications |
|---|---|
jordanhubbard/ACC |
616 |
jordanhubbard/mac |
412 |
NVIDIA-dev/Sim-Slim |
115 |
jordanhubbard/nanolang |
73 |
NVIDIA-dev/oss-tracker |
24 |
jordanhubbard/Aviation |
13 |
In that historical audit window, there were 19 new CI notifications in the preceding 24 hours. Representative failures included MAC's impact-selected coverage gate, NanoLang's backend matrix, and Aviation's application deployment workflow.
Notification read state is therefore not an execution ledger. The monitor keys checks to immutable repository, ref, and commit SHA identities and records the outcome in MAC.
Cadence¶
The monitor intentionally uses only two coarse latency buckets:
- Recent average completed-run duration at or below two hours: check at least every four hours.
- Recent average duration above two hours: check at least every eight hours.
This matches observed repositories without trying to predict every workflow: MAC and NanoLang generally finish in 10-20 minutes, while successful OrcaSlicer builds average about four hours and can approach six. A run that is still pending is re-polled; it is not treated as missing or failed.
Repositories may override the defaults with project metadata under
cicd_monitor. GitHub repositories are otherwise auto-detected, and
repositories with no Actions/check/status data are quietly classified as
having no CI.
Post-publication checks¶
After a Git publication, MAC records a durable
cicd.followup.scheduled observability event containing:
- the publication and source task IDs;
- the registered project and GitHub repository identity;
- the exact canonical SHA; and
- the first due time.
It checks that exact canonical SHA after the settle delay, rechecks pending CI, and records the outcome. A terminal failure is ordinary maintenance input: MAC creates at most one unfinished, low-priority cleanup task per repository and coalesces later failures into the same cleanup pressure. That task may fix a bounded issue directly, classify a transient or intentionally accepted failure, or split out narrower work only when that is genuinely useful.
The original repository task remains completed. A later CI failure is new, linked lifecycle work rather than a retroactive claim that publication did not happen. It does not hold dispatch, publication, deployment, or unrelated work.
Operation¶
The authoritative hub enables the monitor by default. Operators can inspect or trigger it through:
GitHub credentials are resolved through the existing secret-safe GitHub token provider. Reports and task metadata contain repository/run URLs and classified outcomes, never token values or authenticated URLs.