Revert "feat(sbd-v6): declare the axis a session graded on, and grade the conc sweep on it too" - #1523
Merged
Conversation
… the conc sweep on it too"
xiaofei-zheng
deleted the
revert-1520-feat/sbd-v6-agentx-grading-axis
branch
September 16, 2026 08:21
lishuoshuo-amd
added a commit
that referenced
this pull request
Sep 16, 2026
Picks up the three reverts (#1523, #1524, #1525), which drop #1520, #1511 and #1495 out of the release. #1525 is the one that changes this branch. #1495 both re-keyed the Recipe KB on compute-partition shape and removed --recipe-kb-strict-fingerprint, so reverting it restores the option: the parser carries it again, and the CLI delta against v1.1.0 is now one removal (--breakdown-include-transcripts) and one addition (--extend-hours). Its two changelog entries, which main deleted from [Unreleased], are dropped from the v1.1.1 section here -- the conflict was the whole section against an emptied [Unreleased], resolved by keeping the section and deleting those two. The --breakdown-include-transcripts entry loses the comparison it drew against the option that is now back. The release notes lose the same two claims: the partition-key highlight, the warm-start clause in the opening paragraph, and the upgrade-table row, leaving one removed option rather than two. #1520 and #1511 needed no entries because neither ships. The environment contract is untouched by the reverts: the seven variables are still absent from .env.template, and the five collective ones still have no read site.
8 tasks
xiaofei-zheng
pushed a commit
that referenced
this pull request
Sep 17, 2026
…c sweep on it too (#1527) * feat(sbd-v6): declare the axis a session graded on Reapply the changes from #1520 after the temporary revert in #1523. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(sbd-v6): address review feedback on grading axis wiring Allow degraded output-axis lifts to promote and record degrade_reason, read grading from resolved_grading instead of the environment in explore and report paths, thread session grading into GEAK axis selection, and sync docs/docstrings with the V6 outcome wire shape. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(sbd-v6): fail closed when AgentX grading cannot apply AgentX sessions no longer KEEP on the diagnostic output figure when the interactivity axis pair is missing. Explore and _lift_to_current_best refuse the promotion instead of degrading to throughput grading. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(sbd-v6): remove remaining AgentX throughput fallback paths Drop ANCHOR_DEGRADED, force REVERT at the grading resolver when degraded, check comparability before stack validation verdicts, and align explore tests with fail-closed AgentX grading. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(sbd-v6): thread the seeded grading into the AgentX workload spec ``build_agentx_workload_spec`` grew a ``grading`` parameter so the spec it publishes to GEAK would name the axis and band the session was seeded with, but nothing passed it: the sole call site is inside ``apply_agentx_switch``, which had no way to reach the live state. ``metric_basis`` and ``intvty_p90_veto_pct`` therefore still came from whichever process happened to render the recipe, which is the defect the seeded block exists to fix -- a resume or a re-baseline subprocess can have lost HYPERLOOM_PERF_METRIC, or hold a different one. ``grading`` now rides the config-materialization chain the way ``agentx_mode`` already does, and the three callers that hold the live state pass it: the baseline executor, explore, and the conc sweep. A caller that cannot reach the state passes nothing and keeps the environment-derived fallback byte for byte, so no existing recipe moves. Also drops the ``framework_is_scriptable`` import explore no longer uses after the switch to ``resolved_grading``; it was failing ruff's F401. Co-authored-by: Cursor <cursoragent@cursor.com> * test(sbd-v6): a degraded pair asserts REVERT, not a throughput KEEP These two suites still encoded the fallback the previous commits removed: ``test_report_performance_comparison_snapshots_effective_axes`` expected KEEP on every missing-axis case, and the stack-validation matrix expected KEEP whenever the output figure happened to move up. ``resolve_graded_comparison`` now returns REVERT on a degraded pair whichever way output moved, so both assertions were pinning the old rule and failing CI. The output figures are still recorded beside the refusal, which is what these cases check -- they are a diagnostic, not a verdict that could promote. Co-authored-by: Cursor <cursoragent@cursor.com> * docs(sbd-v6): state the fail-closed rule where the degrade narrative was The code has been fail-closed since the two previous commits, but the wire contract still described degradation as a working path: the ``V6Grading`` and ``_grading`` docstrings and two sections of the breakdown reference all said a session configured for interactivity "still grades an individual comparison on output whenever either side cannot supply the axis pair". That is no longer true at any layer -- such a comparison returns REVERT at the resolver. ``_degraded_adoptions`` and the ``validation.notes`` finding it fed go with them. The note counted adoptions carrying a ``degrade_reason``, and no adoption can carry one now: a REVERT never reaches the adoption writeback, so the finding could not fire. ``test_a_degraded_agentx_lift_is_refused`` already pins that invariant on the refusal itself, which is where it belongs, so the note's own test is dropped rather than rewritten against a condition the orchestrator cannot produce. CHANGELOG gains the entries this PR's own wire-shape changes were missing: the ``conc_sweep`` pair-column rename, flagged breaking because no alias is kept and the old names read a latency percentile as a throughput, plus ``metadata.grading``, the ``perf`` axis blocks and the ``graded_on`` fields. Co-authored-by: Cursor <cursoragent@cursor.com> * style(sbd-v6): ruff format the degraded-lift assertion ``ruff format --check`` is a separate CI step from ``ruff check``, and the assertion added with the fail-closed commit was never run through it. Formatting only. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: kong.haishuo <Kong.Haishuo@amd.com> Co-authored-by: Cursor <cursoragent@cursor.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Reverts #1520