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
{{ message }}
Repository navigation
Add Waitcheck triage and shared sanitizer skill foundation - #608
Adds Waitcheck planning and saved-evidence interpretation through AORTA, with a small shared sanitizer skill foundation. The skill selects one verified compiled kernel entry, prepares an explicitly bounded recipe and explains findings, no-finding scope and incomplete checking.
The installed skill contains its instructions, governance card, evaluation dataset, shared workflow, Waitcheck guidance/procedure, exact-entry template and offline report reader. Development controls, build/run helpers, cap/isolation/invalid-selector recipes and longer example guides live under examples/sanitizers/waitcheck/; their tests and upstream license notice are preserved. The reader requires the matching AORTA Python environment.
The AORTA reporting changes distinguish diagnostic totals, retained details, omitted details and native scan completion. Native checker exit, AORTA CLI exit and reader success remain separate outcomes. Legacy reports remain readable.
Validation:
330 local Linux CPU tests passed across eleven helper, recipe, reporting, parser and model test modules (one warning).
Generic skill validation and pinned Skillscope structural checks passed for the package, governance card, dataset and affected documentation. The authored dataset has twelve focused cases: seven positive and five negative; Skillscope also includes five shared negatives. Internal references were checked; external URLs were not fetched.
Pre-commit formatting/YAML checks and the staged whitespace check passed.
No paid model routing/behavior evaluation, native checker scan, GPU job or cluster command was run for this refactor. Retained bounded checker evidence keeps its original scope and limits.
The execution route remains draft. Native acquisition/build provenance, complete loaded-agent activity and final execution qualification remain open; retained bounded development evidence does not qualify arbitrary kernels. Additional checker integrations will arrive in separate AORTA PRs before one combined AMD Skills submission. Catalog owner/license approval, intake and publication remain separate work.
The reason will be displayed to describe this comment to others. Learn more.
🟢 Approval recommended
No active review issues are recorded for the draft sanitizer triage skill, while qualification and release gates remain pending.
No open issues 🎉
What changed in this PR
Adds a draft ROCm sanitizer triage skill that guides planning and evidence interpretation through AORTA’s existing checker integrations. All execution routes remain draft; qualification, agent evaluation, release gates, and dedicated skill CI remain pending.
Changes:
Defines shared guidance and checker-specific references for Waitcheck, ASan, ConSan, and Triton FpSan.
Packages Waitcheck finding, corrected-control, and digest-refusal examples with preparation and offline verification support.
Adds portable tests, evaluation prompts, and maintenance guidance.
The reason will be displayed to describe this comment to others. Learn more.
🟡 Changes recommended
Control verification can accept diagnostics from the wrong object, target, or image, while report-only validation misses hazard exits without supporting findings.
A matching summary header plus a diagnostic naming a different object, target or image is accepted as “expected result verified.” The parser copies the requested identity into the finding metadata, so those comparisons do not verify the diagnostic's origin. Validate the raw diagnostic prefix against the selected identity, as waitcheck_report.py already does, before accepting the control.
The reason will be displayed to describe this comment to others. Learn more.
🟡 Changes recommended
Validation gaps let the control verifier accept inconsistent diagnostics or metadata and the report reader overlook native exit 4 with no findings when optional diagnostic metadata is absent.
The reason will be displayed to describe this comment to others. Learn more.
🟡 Changes recommended
The report reader can accept warning results without retained diagnostics, reject valid nullable or omitted metadata, and misrepresent lower-bound counters as exact totals.
The reason will be displayed to describe this comment to others. Learn more.
🟡 Changes recommended
The reader incorrectly accepts a legacy pass/native-0 report despite a positive raw diagnostic count when all diagnostics are omitted, returning exit 0 and falsely reporting agreement.
The reason will be displayed to describe this comment to others. Learn more.
🔵 Needs a closer look
Human assessment is needed because critical validation gaps can accept failed scans as consistent evidence and publish exact totals that contradict retained findings.
Scope the skill entry point, references and evaluations to Waitcheck while
preserving the shared workflow and existing skill name. Defer other checker
procedures to their own AORTA integrations.
Keep development controls, recipes and long guides under repository examples,
with their license, helper paths, tests and CI coverage preserved. Retain only
the required catalog files and useful Waitcheck runtime support in the skill.
oyazdanb
changed the title
Add draft ROCm sanitizer skill and Waitcheck evidence validation
Add Waitcheck triage and shared sanitizer skill foundation
Nov 13, 2026
The reason will be displayed to describe this comment to others. Learn more.
🟡 Changes recommended
Saved reports can misrepresent truncated diagnostics as exact totals and accept invalid native exits as completed scans without flagging either contradiction.
Reject unsupported native exits in completed Waitcheck reports and flag
per-entry emitted records above a known requested limit in report-only mode.
Match CheckResult's exact-count/truncation invariant to KernelCheckResult.
Add regressions for all three review findings and preserve valid policy
failures, legacy unknown metadata and multi-entry aggregation.
The reason will be displayed to describe this comment to others. Learn more.
🔵 Needs a closer look
Human assessment is needed because a moderate-severity finding shows that matching raw logs cause valid completed policy-failure results to be incorrectly rejected.
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
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.
Adds Waitcheck planning and saved-evidence interpretation through AORTA, with a small shared sanitizer skill foundation. The skill selects one verified compiled kernel entry, prepares an explicitly bounded recipe and explains findings, no-finding scope and incomplete checking.
The installed skill contains its instructions, governance card, evaluation dataset, shared workflow, Waitcheck guidance/procedure, exact-entry template and offline report reader. Development controls, build/run helpers, cap/isolation/invalid-selector recipes and longer example guides live under
examples/sanitizers/waitcheck/; their tests and upstream license notice are preserved. The reader requires the matching AORTA Python environment.The AORTA reporting changes distinguish diagnostic totals, retained details, omitted details and native scan completion. Native checker exit, AORTA CLI exit and reader success remain separate outcomes. Legacy reports remain readable.
Validation:
The execution route remains draft. Native acquisition/build provenance, complete loaded-agent activity and final execution qualification remain open; retained bounded development evidence does not qualify arbitrary kernels. Additional checker integrations will arrive in separate AORTA PRs before one combined AMD Skills submission. Catalog owner/license approval, intake and publication remain separate work.