You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
(issues #7627, #7630, #7636, #7637, #7639, #7640, #7641, #7642, #7643 — rubric finding: nine prior ci-scan-feedback runs filed this same proposed edit as fallback issues instead of PRs because git push to origin kept failing; this is the 10th attempt)
Rubric finding: of 67 ci-scan runs since window-start (2026-06-08T13:53:53Z), the recurring skip reason is stale build window (>14d) / no follow-up build yet, defer to next run. Runs hitting these conditions consumed 2.2–2.4M+ effective tokens before concluding with noop — approximately 10× the ~250K ET of a correctly-halting Step 1 exit.
Rubric finding (tally honesty): skip-only runs emit | 0 | 0 | 0 | 0 | (wrong) instead of | 0 | 0 | 0 | 1 | (correct). Hard Rule 10 fixes this by mandating a skipped-with-reason=1 tally on every skip-only run.
Proposed edits
.github/workflows/ci-scan.agent.md (Hard Rules section, after rule 9): Add Hard Rule 10 that names exact forbidden operations (AzDO timeline fetch, log download, Helix query) and mandates the literal tally row | 0 | 0 | 0 | 1 | — tied to the signal above (nine prior failed attempts + ~10× token variance).
.github/workflows/ci-scan.agent.md (Step 1 trailing sentence): Replace the inline restatement of the three skip reasons with a single reference to Hard Rule 10 so the constraint is stated once, authoritatively — eliminates the ambiguity that let previous scanner runs continue past the skip condition.
Expected behavior change
On any run where Step 1 yields a selection-time skip (no follow-up build yet, defer to next run, stale build window (>14d), or no failed build in 7d), the scanner will append the reason to the coverage file, print | 0 | 0 | 0 | 1 |, call noop, and stop immediately — without fetching any AzDO timeline, downloading any task log, or querying any Helix work item. This eliminates the observed ~10× token variance between correct low-ET runs (~250K ET) and high-ET runs (2.2–2.4M+ ET) on identical pipeline state, and ensures the tally row correctly shows skipped-with-reason=1 on skip-only runs.
The patch file is available in the agent artifact in the workflow run linked above.
To create a pull request with the changes:
# Download the artifact from the workflow run
gh run download 28146069117 -n agent -D /tmp/agent-28146069117
# Create a new branch
git checkout -b ci-scan-feedback/hard-rule-10-early-exit-2026-06-25-f983f0de80cc2582
# Apply the patch (--3way handles cross-repo patches where files may already exist)
git am --3way /tmp/agent-28146069117/aw-ci-scan-feedback-hard-rule-10-early-exit-2026-06-25.patch
# Push the branch to origin
git push origin ci-scan-feedback/hard-rule-10-early-exit-2026-06-25-f983f0de80cc2582
# Create the pull request
gh pr create --title '[ci-scan-feedback] ci-scan: add Hard Rule 10 to force early exit on no scannable build' --base main --head ci-scan-feedback/hard-rule-10-early-exit-2026-06-25-f983f0de80cc2582 --repo dotnet/machinelearning
Show patch preview (58 of 58 lines)
From 6cb381741cfa78efcb2b664d96901c00a2f4ec44 Mon Sep 17 00:00:00 2001
From: "github-actions[bot]" <github-actions[bot]@users.noreply.github.com>
Date: Thu, 25 Jun 2026 04:14:32 +0000
Subject: [PATCH] ci-scan: add Hard Rule 10 to force early exit on no scannable
build
10 consecutive ci-scan-feedback runs identified that the scanner
continues to Steps 2-7 (AzDO timeline, log download, Helix queries)
even after Step 1 determines there is no scannable build, consuming
2.2-2.4M+ effective tokens on skip-only runs vs ~250K ET when properly
halted.
Add Hard Rule 10 which elevates the no-scannable-build exit to the
same hard-constraint level as the issue-cap rule: append the skip
reason to the coverage file, print | 0 | 0 | 0 | 1 |, call noop,
and stop - without fetching any AzDO timeline, downloading any log,
or querying any Helix work item.
Also fixes Step 7 tally honesty: skip-only runs should emit
| 0 | 0 | 0 | 1 | (correct) not | 0 | 0 | 0 | 0 | (wrong).
Update Step 1's trailing sentence to reference Hard Rule 10 directly
instead of restating the skip-reason list inline.
Signal: issues #7627, #7630, #7636, #7637, #7639, #7640, #7641,
#7642, #7643 (nine prior failed PR push attempts carrying the same
proposed edit).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---
.github/workflows/ci-scan.agent.md | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/.github/workflows/ci-scan.agent.md b/.github/workflows/ci-scan.agent.md
index 0937c5f..4ba5b16 100644
--- a/.github/workflows/ci-scan.agent.md+++ b/.github/workflows/ci-scan.agent.md@@ -78,6 +78,7 @@ These invariants are not delegated to the shared file. Honor them even if a shar
7. **All state under `/tmp/gh-aw/agent/`;** each bash call is a fresh subshell.
8. **AzDO REST is anonymous;** stay on `https://dev.azure.com/dnceng-public/public/_apis/build/...`. Follow every rule in [Environment constraints](shared/ci-scan.instructions.md#environment-constraints) (pre-bind UR
... (truncated)
Triggering signals
stale build window (>14d)/no follow-up build yet, defer to next run. Runs hitting these conditions consumed 2.2–2.4M+ effective tokens before concluding with noop — approximately 10× the ~250K ET of a correctly-halting Step 1 exit.| 0 | 0 | 0 | 0 |(wrong) instead of| 0 | 0 | 0 | 1 |(correct). Hard Rule 10 fixes this by mandating askipped-with-reason=1tally on every skip-only run.Proposed edits
.github/workflows/ci-scan.agent.md(Hard Rules section, after rule 9): Add Hard Rule 10 that names exact forbidden operations (AzDO timeline fetch, log download, Helix query) and mandates the literal tally row| 0 | 0 | 0 | 1 |— tied to the signal above (nine prior failed attempts + ~10× token variance)..github/workflows/ci-scan.agent.md(Step 1 trailing sentence): Replace the inline restatement of the three skip reasons with a single reference to Hard Rule 10 so the constraint is stated once, authoritatively — eliminates the ambiguity that let previous scanner runs continue past the skip condition.Expected behavior change
On any run where Step 1 yields a selection-time skip (
no follow-up build yet, defer to next run,stale build window (>14d), orno failed build in 7d), the scanner will append the reason to the coverage file, print| 0 | 0 | 0 | 1 |, callnoop, and stop immediately — without fetching any AzDO timeline, downloading any task log, or querying any Helix work item. This eliminates the observed ~10× token variance between correct low-ET runs (~250K ET) and high-ET runs (2.2–2.4M+ ET) on identical pipeline state, and ensures the tally row correctly showsskipped-with-reason=1on skip-only runs.Note
This was originally intended as a pull request, but the git push operation failed.
Workflow Run: View run details and download patch artifact
The patch file is available in the
agentartifact in the workflow run linked above.To create a pull request with the changes:
Show patch preview (58 of 58 lines)