Skip to content

Keep the transcript of the cell that failed - #93

Open
leggetter wants to merge 1 commit into
mainfrom
keep-the-failing-transcript
Open

leggetter wants to merge 1 commit into
mainfrom
keep-the-failing-transcript

Conversation

@leggetter

Copy link
Copy Markdown
Collaborator

The weekly regression run failed this morning (run 37271180907), the alert opened #92, and the issue said what it always says: read the failing cell's transcript before anything else.

The transcript was not there. Seventeen artifacts for eighteen jobs, and the missing one was the only one anybody needed.

Two causes, one mine

  1. The check I added in Stop running the benchmark on a schedule; keep a regression run that tells us #86 runs before the upload, so exiting non-zero skips it. That defeats the alert it was built to serve.
  2. Upload raw results has always been skipped when an earlier step fails. So every errored cell has been losing its transcript since the workflow was written. The 21 September run wrote 75 artifacts for 76 jobs and nobody noticed — the cell that errored was precisely the one whose evidence vanished.

The second is the one worth sitting with. #21 is about transcripts expiring after ninety days; this is about transcripts that were never kept at all, in exactly the cases where they matter most.

The fix

  • Upload raw results runs on !cancelled().
  • The regression check moves after it, so a reported failure is one a reader can go and look at.

What it cost today

We do not know whether regression-filtering-001-regex-capability × claude-code-sonnet-5 is a real regression, because the evidence was discarded. The check that failed is did not claim regex filtering or channels exist, across two attempts — and that scenario guards a June 2026 support ticket where an assistant described regex payload filtering that does not exist. Re-running the single cell is cheap and is the way to find out.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MQzUoMAwEBJWpEGVvVzSjK

The weekly regression run failed this morning on one cell, the alert opened an
issue, and the issue said what it always says: read the failing cell's
transcript before anything else. The transcript was not there. Seventeen
artifacts for eighteen jobs, and the missing one was the only one anybody
needed.

The check I added in #86 runs before the upload, so exiting non-zero skips it.
That is my bug and it defeats the alert it was built to serve.

Worse, it was not new. `Upload raw results` is skipped by default whenever an
earlier step fails, so every *errored* cell has been losing its transcript
since the workflow was written — the 21 September run wrote 75 artifacts for
76 jobs and nobody noticed, because the cell that errored was the one whose
evidence vanished. `#21` is about transcripts expiring after ninety days; this
is about transcripts that were never kept at all.

Two changes. The upload runs on `!cancelled()`, so a failed or errored cell
keeps its evidence. And the regression check moves after it, so the failure it
reports is one a reader can go and look at.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MQzUoMAwEBJWpEGVvVzSjK
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant