Tune E2E CI Time Estimates

Use only complete GitHub Actions artifacts produced by the repository's ci-fetch-log skill. Produce an auditable CSV and, unless the user requests CSV-only output, update eligible est_time literals.

Accept a GitHub Actions run URL, job URL, or ci-fetch-log artifact directory. After full-run collection is verified, optional suite, job-name, or partition filters define the selected evidence set used for the CSV and edits. They never narrow log collection, and every grouping rule below applies across all rows in that selected set.

Obtain complete logs

For a URL, follow ci-fetch-log in complete-log collection mode for every job in the run. Treat a job URL only as a hint identifying the run. Never pass --only, --job, --job-name, a conclusion filter, or another selector that narrows collection. Do not fetch logs independently or use pasted excerpts.

For an existing artifact directory, require a manifest.json with:

• producer="ci-fetch-log".
• schema_version=1.
• artifact_kind="complete-gha-job-logs".
• mode="complete-log collection".
• A positive run_attempt that matches the attempt recorded in run.json.
• An existing run.json whose full job-ID set exactly matches the job-ID set in manifest.json, proving that no selector narrowed collection.
• A manifest repo that matches one of the current checkout's configured GitHub remotes; stop on a repository mismatch or an ambiguous checkout identity.
• A recorded head_sha and current HEAD. A mismatch is allowed when head_sha is an ancestor of HEAD; otherwise default to report-only mode unless the user explicitly approves applying cross-revision edits.
• Every job selected for parsing marked log_complete=true.
• An existing, non-empty log_path for each selected job whose actual line count matches line_count.

Stop and report the failed identity, scope, or completeness check. Apply any CSV/edit-view filter only after obtaining and validating the complete-run manifest, then parse every job admitted to that selected evidence set.

Parse scheduled tests and outcomes

Consider only scheduled files under tests/e2e. Ignore tests/fast, tests/fast-gpu, unit tests, setup spans, and other runtimes.

For each selected complete job log:

• Recover suite, partition_id, and partition_size from the job name when available.
• Parse the Enabled N test(s) block for scheduled files and logged estimates. A job can print the plan during both a list-only step and the real run, so deduplicate identical plan entries and prefer the block associated with the execution step.
• Parse End (...) timing records containing filename='tests/e2e/...', elapsed=<seconds>, _estimated_time=<seconds> and the corresponding [i/N] <test_file> PASS|FAIL|TIMEOUT elapsed=<seconds>s attempt-status lines.
• Tolerate ISO 8601 timestamp prefixes and ANSI escapes while parsing; do not rewrite the saved log.
• Parse the final PASSED: and FAILED: summary sections.
• Classify every scheduled file using the precedence below so each CSV row receives exactly one status.

Associate each End (...) record with its attempt-status line by filename and attempt order, not by the displayed index: legacy End indices are zero-based while attempt-status indices are one-based. For retries, exclude earlier failed or timed-out attempts and use only the timing associated with the passing attempt.

1. passed: The associated attempt has PASS, an End (...) timing record, and the file appears under the final PASSED: section. A passed test remains valid evidence when another test later fails the same job.
2. timeout: A file-level attempt has TIMEOUT; this takes precedence over the same file appearing under the generic FAILED: section.
3. unknown: An End (...) record exists but the final summary is missing, so the evidence cannot establish a successful sample.
4. failed: The file appears under FAILED: or has an associated failure record not already classified as timeout or unknown.
5. cancelled: The job was cancelled before a more specific file result became available.
6. not_run: The file was scheduled but has no timing record or final result and no job-level cancellation or timeout explains the absence.

For a job-level timed_out conclusion without a file-level result, classify the affected scheduled row as timeout before considering cancelled or not_run. Outside the explicit missing-summary rule above, use unknown only as the final fallback for contradictory or insufficient evidence.

Write the evidence CSV

Write one row per scheduled tests/e2e file to the requested path or <artifact_dir>/e2e-runtime.csv by default. Use these columns:

workflow_run_id,job_id,job_name,suite,partition_id,partition_size,test_file,status,elapsed_s,logged_est_s,current_est_s,proposed_est_s,action,source_log,timing_line,status_line,note

Populate logged_est_s from the complete log and current_est_s from the current checkout. Cite the saved log path and exact line numbers supporting both timing and status. For scheduled-only rows, cite the relevant Enabled N test(s) or summary line.

Compute one grouped action per test_file after parsing every selected job, then copy that action to every row for the file. Fill proposed_est_s only on passed rows. Use these actions:

• raise: the proposal is greater than the current estimate and edits apply.
• lower: the proposal is smaller and edits apply unconditionally.
• keep: the proposal equals the current estimate.
• report_only: passed evidence exists but edits were disabled.
• skip: no passed evidence exists, the file has multiple register_cuda_ci calls, or est_time is missing or non-literal.

Leave proposed_est_s empty for non-passed and skip rows.

Compute estimates by test file

Group rows by test_file across all selected jobs, partitions, retries, and reruns represented in the run. Never group by (job_id, test_file).

Use only passed rows as runtime evidence:

max_elapsed_s = max(elapsed_s for passed rows of this test_file)
raw_s = ceil(max_elapsed_s * 1.25)

if raw_s <= 200:
    proposed_est_s = ceil(raw_s / 10) * 10
else:
    proposed_est_s = ceil(raw_s / 100) * 100

Always round upward to the bucket. Never use raw_s directly or round down. raw_s == 200 proposes 200; raw_s == 201 proposes 300.

Failed, timed-out, cancelled, not-run, and unknown rows never contribute to max_elapsed_s. They never cancel a raise or block a lower. When the bucket-rounded proposal is below current_est_s, apply the lower unconditionally even if sibling rows did not pass. Timeout tuning is outside this workflow unless the user requests it separately.

Apply only literal estimate edits

For an eligible file, replace only the numeric literal assigned to est_time in its single register_cuda_ci(...) call. Preserve all surrounding formatting and code.

Do not change suites, labels, disabled flags, timeouts, skips, assertions, test logic, imports, or runner configuration. If the call is ambiguous or est_time is computed rather than literal, record skip instead of editing around it.

Do not stage, commit, push, or open a pull request. Leave edits in the working tree.

Verify and report

Verify that:

• The CSV parses successfully.
• Every passed row has elapsed_s, source_log, timing_line, and status_line.
• Grouped proposals exactly follow the formula and action rules.
• The diff changes only literal est_time values.

Run the registry's list-only command for affected suites and partitions, for example:

PYTHONPATH=. python tests/ci/run_suite.py --hw cuda --suite <suite> --match-all-labels --auto-partition-id <id> --auto-partition-size <size> --list-only

Confirm the new estimates appear and partition totals remain sensible. Do not run GPU E2E tests solely to verify timing metadata unless the user explicitly asks.

Report the CSV path, raises and lowers as max_elapsed_s -> proposed_est_s, skipped files and reasons, verification commands, and that all edits remain uncommitted.