deslop v0.3.4

deslop v0.3.4

This patch clarifies that removing local proof machinery must preserve supported stored formats and domain return values. A writer and reader changed together are not independent compatibility evidence. The existing reference now pairs supported-format preservation with explicit format retirement; invocation and authorization rules are unchanged.

The runtime change touches two existing files. Separate development additions provide JSON lifecycle cases, independently designed CSV/SQLite transfer cases, a rejection-coverage recognition revision, and calibrated supplementary probes for two scoring gaps. Frozen historical fixtures and scores are preserved.

The full GPT-6.1 Sol / xhigh comparison had 192 planned primary calls and one separately reported isolation repair. Results were mixed: micro 40/48 → 38/48, mini 6/9 (8 valid) → 8/9, JSON 6/6 → 6/6, transfer 12/12 → 12/12, edges 14/18 → 15/18, and read-only controls 3/3 → 3/3. A diagnostic on repeats of one known old-JSON fixture improved 0/3 → 3/3. This is a targeted policy fix, not a claim of overall model improvement. The comparison's known SQLite and legacy-JSON grader gaps are documented separately; raw trajectories remain in the maintainer's local archive.

Validation: independent GPT-6 Astra review and follow-up, 14 offline checks, strict Claude packaging validation, matching evaluated/packaged runtime hashes, Codex repository-copy discovery and explicitly selected Skill invocation, plus isolated Claude marketplace installation and a real /deslop:deslop audit invocation of the exact local plugin source. Smoke fixtures remained unchanged. These host checks are distinct from model-quality evidence.

• Full release notes and evidence limits
• 中文发布说明
• Installation and upgrades

The attached ZIP contains only the self-contained deslop/ Skill payload. Codex continues to use the standalone Skill; Claude Code can install deslop@deslop and invoke /deslop:deslop.

中文：本版是对持久化兼容性边界的定点修正，运行时只改动两个已有文件；评测与校准新增内容独立保存。完整对照结果有升有降，不宣称整体性能提升。独立审核发现的评分漏检已补充独立探针，原始评分保持不变。

deslop v0.3.3

deslop v0.3.3 is a patch release for evidence-backed cleanup boundaries. The shared skills/deslop/ installation path, explicit apply authorization, and Codex/Claude Code entry points are unchanged.

Changes

• Preserve supported public APIs and independently validated persisted metadata while deleting local self-proof chains.
• Retarget the last test of a surviving public contract to an independent result; keep rejection coverage without prescribing a single test layout or inventing an oracle.
• Add paired examples for public write/read operations and local versus publisher-enforced fingerprints.
• Consolidate the development corpus as dev-v3-evidence-edges-draft4: nine pairs, explicit retirement evidence, bounded coverage-repair calibration, and reader-loop recognition.

Validation

All seven offline repository checks, Skill validation, and five post-grade calibration states passed. Claude Code 2.1.280 passed strict Plugin and marketplace manifest validation. Frozen rc5 fixtures and historical scores are unchanged.

GitHub Actions passed for both the release tag and marketplace catalog.

The final payload has not been rerun in a model A/B. These are targeted guidance and regression improvements, not a claim of a measured performance increase.

Update

• Codex: install the v0.3.3 Skill, or move a linked checkout to the v0.3.3 tag.
• Claude Code Plugin: refresh the marketplace and run claude plugin update deslop@deslop, then restart the session.

---

本次为小幅补丁更新：保留公开接口、独立元数据校验及存活行为的必要测试，补充两组清理正反案例，并修正评测中的退役依据和 reader 循环识别。旧 rc5 与历史成绩保持不变。七项离线检查、Skill 校验、五个评分入口校准及 Claude Code 严格 manifest 校验通过；最终版本未重新进行模型 A/B，因此不宣称模型通过率已提高。

deslop v0.2.0

deslop v0.2.0 adopts the canonical skills/deslop/ runtime layout while preserving the Skill contents byte-for-byte from v0.1.0.

Highlights:

• The canonical runtime moved from skill/deslop/ to skills/deslop/.
• Standalone Skill installation remains fully supported.
• Runtime Skill policy, references, and evaluation semantics are unchanged.
• The repository now follows the plural skills/ layout expected by current Plugin packaging conventions.

Plugin status:

• Plugin distribution is not included in v0.2.0.
• In the tested Codex CLI 0.149.1 environment, a temporary valid Skills-only package passed static validation and local marketplace discovery, installation, and cache creation, but its bundled Skill was not registered in the native available-skills catalog after ambient Skill isolation and a fresh app-server session.
• This is an observed host registration issue in the tested version, not a claim about every Codex host or a confirmed upstream root cause.

Upgrade guidance for v0.1.0 installer and symlink users is included in Getting Started. See also Design, Evaluation, Field Trials, and the Plugin compatibility note.

v0.1.0 remains permanently available at its original immutable skill/deslop/ path.

deslop v0.1.0

First public release of deslop, a deletion-first Agent Skill for removing accumulated test-suite bloat, verification theater, and defensive or fallback machinery while preserving externally meaningful behavior.

Highlights:

• explicit read-only audit and apply authorization model;
• evidence-root and closed-justification-loop decision model;
• focused test, verification, and fallback cleanup policy;
• preservation-first evaluation;
• manually adjudicated real-world cluster-gpu-monitor field trial;
• standalone Skill distribution from an immutable Git tag.

Documentation:

• Getting Started
• Design
• Evaluation
• Field Trials