workflow-fix: 6d.0-bis smoke gate: name data-ingestion tiny-
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 /daily Step-C parked-candidate sweep from a candidate parked on task #1113 (park_form: recursion-guard).
Goal
Extend the Step 6d.0-bis smoke-evidence standard to name data-ingestion tiny-real probes (real-corpus streaming builders) the way it names the #906 GPU-driver class, and correct the implementer memory's WildChat moderation-shape line.
Workflow gap
- Bug observed: Step 6d.0-bis's tiny-real smoke-evidence standard does not name real-corpus data-ingestion probes, and the implementer memory misstates WildChat moderation fields as 'per-turn openai_moderation' keys instead of per-turn-ALIGNED top-level list columns.
- Why it is a workflow gap: The #1092 P0 failure class (a streaming filter chain rejecting 100% of real rows while synthetic smokes stay green) is exactly what the smoke-evidence standard exists to force evidence for; the memory line seeds the wrong field-shape assumption.
- Confidence (emitter): medium (Alternatives critic + Methodology critic, #1113 Phase 2)
Proposed change (candidate diff sketch — refine in planning)
(1) SKILL.md 6d.0-bis: add real-corpus streaming/data-ingestion builders to the named tiny-real probe classes (bounded tiny-real streaming probe, kept>0 per dataset, per-filter reject counters). (2) memory: 'per-turn openai_moderation' -> 'per-turn-ALIGNED top-level list columns (not keys inside each turn dict)'.
Scope / surfaces
- Primary target:
.claude/skills/issue/SKILL.md, .claude/agent-memory/experiment-implementer/feedback_real_corpus_streaming_filters_tiny_real_probe.md - Grep the workflow surface for the pattern before editing (
grep -rln '<pattern>' .claude/ CLAUDE.md scripts/) and update every hit; list them in the plan.
Constraints / invariants
- Workflow-surface only — never experiment code,
configs/, ortasks/. scripts/workflow_lint.py --check-askspasses; ruff on touched files passes; ifworkflow.yamlorCLAUDE.mdchange, they stay consistent with the rule file.- This session runs under
EPM_WORKFLOW_FIX_SESSION=1and carries aworkflow_fix_target:Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).
Provenance
- workflow_fix_target: .claude/skills/issue/SKILL.md, .claude/agent-memory/experiment-implementer/feedback_real_corpus_streaming_filters_tiny_real_probe.md
- origin: parked candidate on task #1113 at 2026-07-07T18:29:07Z
parked — running under workflow_fix_target recursion guard (see .claude/rules/workflow-fix-on-bug.md § Recursion guard). TWO candidates surfaced by Phase 2 critics, logged not routed: (1) target_file: .claude/skills/issue/SKILL.md — extend Step 6d.0-bis smoke-evidence standard to name data-ingestion tiny-real probes (real-corpus streaming builders) the way it names the #906 GPU-driver class. Source: Alternatives critic, confidence medium. (2) target_file: .claude/agent-memory/experiment-implementer/feedback_real_corpus_streaming_filters_tiny_real_probe.md — one-line correction: 'per-turn openai_moderation' → 'per-turn-ALIGNED top-level list columns (not keys inside each turn dict)', matching the fact-checked shape. Source: Methodology critic, confidence medium. routed: parked (recursion guard); surface on a future non-workflow-fix orchestrator pass.