EPS
← All tasks·#1337Completed

workflow-fix: step9c scratch oracle for file-anchored scan tests

kind: infra#wf-fix#wf-fix-fp:4d6cb4e2252b
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #1318 (emitting agent: /issue orchestrator).

Goal

Allow the step9c scratch pristine oracle for file-anchored scan-set tests (per-test anchoring allowlist) instead of refusing all GLOB_SCAN_TESTS nodes.

Workflow gap

  • Bug observed: step9c_baseline.py compare --run-pristine exited 2 (MF-4c indeterminate) on #1318's gate because the R-F eligibility rule class-refuses the scratch-worktree oracle for EVERY GLOB_SCAN_TESTS node, although the failing node (tests/test_shared_vm_thread_caps.py::test_no_new_torch_before_dotenv_vm_entrypoints) anchors its scan via Path(__file__).resolve().parents[1] (tests/test_shared_vm_thread_caps.py:871) and a manual scratch run resolved it trustworthily (identical offender set; positive control: the root's untracked strays absent from the scratch scan output).
  • Why it is a workflow gap: R-F's rationale ("repo_root()-anchored live-tree scanners read the MAIN root from any cwd") is correct only for repo_root()-anchored scan tests; applying it as a class rule makes any code dirt on the busy shared root (30+ concurrent sessions parking untracked scripts/*.py) an unavoidable exit-2 for every diff-linked scan-set failure — the #1318 gate had to resolve it with a hand-run oracle + recorded override, exactly the improvisation the mechanical compare exists to prevent.
  • Confidence (emitter): high
  • verified-at-filing: grep -rln "GLOB_SCAN_TESTS" scripts/step9c_baseline.py scripts/select_step9c_tests.py .claude/skills/issue/SKILL.md -> 3 files (target scripts/step9c_baseline.py: R-F rule at 2 sites — the test_file not in ctx.sel.GLOB_SCAN_TESTS eligibility conjunct in _resolve_pristine_bucket + its docstring; select_step9c_tests.py owns the GLOB_SCAN_TESTS map itself, not the refusal; SKILL.md documents the behavior) (2026-07-15). Live repro: /tmp/step9c-compare-issue-1318.json (rc=2, MF-4c, scan_set=True) vs /tmp/issue-1318-manual-oracle.log (identical 9-offender set at clean origin/main 9880b7f28b).

Proposed change (candidate diff sketch — refine in planning)

  • In scripts/step9c_baseline.py, add a module constant, e.g.
  • FILE_ANCHORED_SCAN_TESTS: frozenset[str] = frozenset({
  • "tests/test_shared_vm_thread_caps.py",  # Path(__file__).parents[1] anchor, verified :871
  • })
  • and relax the R-F conjunct in _resolve_pristine_bucket:
  • and test_file not in ctx.sel.GLOB_SCAN_TESTS  # R-F

  • and (test_file not in ctx.sel.GLOB_SCAN_TESTS

  •      or test_file in FILE_ANCHORED_SCAN_TESTS)  # R-F': __file__-anchored scanners scan their own tree
  • (alternative: a mechanical anchoring probe — grep the test file for Path(__file__)-derived root vs repo_root() — instead of a hand-list; planner decides). Update the docstring + the SKILL.md step-1d exit-2 reason list; extend tests/test_step9c_baseline*.py with a dirty-root + file-anchored-scan-test fixture asserting scratch resolution instead of exit 2.

Scope / surfaces

  • Primary target: scripts/step9c_baseline.py
  • Grep the workflow surface for the pattern before editing (grep -rln 'GLOB_SCAN_TESTS' .claude/ CLAUDE.md scripts/) and update every hit that documents the R-F class rule; list them in the plan.

Constraints / invariants

  • Workflow-surface only — never experiment code, configs/, or tasks/.
  • scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; if workflow.yaml or CLAUDE.md change, they stay consistent with the rule file.
  • Fail-closed direction preserved: a scan test NOT provably file-anchored keeps the R-F refusal verbatim; the relaxation is opt-in per verified test.
  • This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).

Provenance

  • workflow_fix_target: scripts/step9c_baseline.py
  • fingerprint: 4d6cb4e2252b

target_file: scripts/step9c_baseline.py bug_observed: compare --run-pristine exited 2 (MF-4c) on 1318 gate because the R-F rule class-refuses scratch resolution for every scan-set node although the failing test anchors its scan via file and a scratch run resolves it trustworthily why_workflow_gap: R-F's repo_root()-anchoring rationale does not apply to file-anchored scan tests, so shared-root code dirt forces exit-2 indeterminates the scratch oracle could safely resolve proposed_change: allow the step9c scratch pristine oracle for file-anchored scan-set tests (per-test anchoring allowlist) instead of refusing all GLOB_SCAN_TESTS nodes diff_sketch: |

  • FILE_ANCHORED_SCAN_TESTS = frozenset({"tests/test_shared_vm_thread_caps.py"})
  • and test_file not in ctx.sel.GLOB_SCAN_TESTS  # R-F

  • and (test_file not in ctx.sel.GLOB_SCAN_TESTS

  •      or test_file in FILE_ANCHORED_SCAN_TESTS)  # R-F'

confidence: high related_task: #1318

Activity