Formerly "Ops with Skill", now officially upgraded. Fully dedicated to AI, we share both the why (fundamental insights) and the how (practical implementation). From technical operations to breakthrough thinking, we help you understand AI's transformation and master the core abilities needed to shape the future. ShugeX: boundless exploration, skillful execution.
This article analyzes Matt Pocock's 27-agent skill framework that imposes explicit exit criteria across eight development activities — from requirement clarification to retrospective — to prevent AI coding failures like misalignment, verbosity, broken code, and architectural decay.
This article dissects Milvus 3.0's default segment sealing behavior — 12% fill threshold, 24-hour TTL, and 16 compaction slots — explaining how sharding, large batches, and delete chains multiply segment counts, why compaction falls behind due to queue and slot bottlenecks, and the correct troubleshooting order: seal conditions first, compaction health second, manual force merge or clustering compaction last.
This article explains the Milvus 3.0 deletion pipeline from request to disk space reclamation, covering tombstone creation, query visibility via bitmaps, L0 segment materialization, two-phase compaction, and GC tolerance windows that delay space return by hours.
This article explains why newly inserted vectors in Milvus 3.0 may not be immediately searchable, revealing that the default Bounded consistency level sets a 5-second grace period (GracefulTime) during which recent writes fall outside the query's guaranteed timestamp (GuaranteeTs). It details the underlying mechanisms — TSO timestamps, timetick propagation, query node water-level waiting, and the four consistency levels — and provides a troubleshooting decision tree to distinguish expected behavior from actual failures.
The article explains that Milvus 3.0 search aggregation operates on ANN candidates, not full collections, so missing buckets stem from candidate budget limits (topK and TopHits.size), not the safe_factor parameter; it distinguishes candidate-signal counts from exact filtered counts and advises choosing the right query path for each need.
This guide explains how to select Matt Pocock's AI programming skills by analyzing your development workflow gaps, covering skill categories, invocation types, setup configuration dependencies, security risk interpretation, and a practical installation checklist.
Milvus 3.0's search iterator pagination fails after reranking because the cursor uses the last page's score as a boundary, but reranking replaces ANN distances with model predictions (e.g., XGBoost raw scores), causing the filter condition to exclude nothing and return duplicate pages. Milvus now explicitly rejects combining function chains with search iterators rather than fixing pagination, as the fix would require carrying original distances through the entire reduce pipeline—a feature-level change. Workarounds include using offset within single-search limits or moving reranking to the application layer.
This article analyzes why rolling back a Milvus 3.0 upgrade is far riskier than the upgrade itself, explaining how Storage V3, index version changes, WAL message semantics, and SDK/proto contracts create irreversible state boundaries that a simple image tag revert cannot undo.
DeepSeek Harness uses three distinct continuation mechanisms—Goal, Ralph, and Workflow—each with explicit trade-offs: Goal requires human resume due to non-persisted activation, Ralph runs fresh sessions with shared workspace, and Workflow lets models write orchestration scripts in isolated worker threads.
This article dissects DeepSeek Harness's approval mechanism, showing it is a fail-closed human-machine handoff rather than a security boundary, covering its four outcome states, dual approval paths, async Cordis flow, per-package vs future-version authorization, host/client execution split, and orthogonal approval-sandbox knobs.
