Skill Loom v0.3.0

Skill Loom v0.3.0

Weekly observation can now select an explicit time interval instead of summarizing an entire transcript.

• Add optional --since (exclusive) / --until (inclusive) timezone-aware bounds to observe-codex and observe-claude, plus Python keyword arguments.
• Compare UTC instants, report skipped timestamps, and count only input files with matching objects in filtered coverage.
• Keep cumulative Codex token totals unknown for filtered intervals. No-option behavior and schema 1 remain compatible.
• Add synthetic dual-host commands, seven regression methods, dated research/source review and upgrade/rollback guidance.

Validation: local Windows/Python 3.12.8 ran 97 tests (95 passed, two platform skips); targeted release check passed for 93 files; isolated demo passed exact rollback, reapply and no-change checks. Synthetic comparisons preserve default reports versus v0.2.1 except generation time and verify both documented interval commands. Independent final review found no material blockers. A normal wheel installed in a fresh environment: all 14 runtime modules matched the frozen commit, pip check and 14 installed interface/window tests, demo and recommendation smoke passed.

Limits: this still scans full files and does not freeze live inputs, manage cursors, deduplicate logical events or establish continuous capture. Native Claude execution, native hooks, macOS, task quality and cost benefits remain unverified. Papers were not reproduced. Dependencies, global skills and host policies are unchanged.

Upgrade in a separate checkout/environment. To downgrade, retain private reports/journals, return to v0.2.1, and remove --since/--until from old CLI calls. No global transaction was performed.

Full commit: 8ccd8f59d5b5484910a14de5386df3bf04650390

Verified CI

Skill Loom v0.2.1

Skill Loom v0.2.1

This documentation PATCH makes prior maintenance decisions reusable when source, task or environment conditions change.

• Add a manual decision-history guide and two-record synthetic example. Reopening requests review; it never authorizes installation. No CLI consumes or enforces this format.
• Add the dated research supplement and two primary-paper readings, with limits on historical evidence and isolated-skill results.
• Recheck seven registered sources; preserve existing pins and keep source triage separate from adoption.

Runtime behavior, schema 1, dependencies and global skills are unchanged. The historical manuscript/PDF/TeX retains its original scope.

Validation: Windows/Python 3.12 ran 90 tests (88 passed, two platform skips); isolated demo passed exact rollback, reapply and no-change checks; targeted release check passed for 87 files; independent review found no material blockers. The ordinary wheel was checked in a fresh environment before publication. The exact publication commit and hosted CI links follow below.

Limits: the example is a manual convention, not a tested decision engine. Native hooks, macOS, model-task benefit and maintenance-cost gains remain unverified. Paper results were not reproduced.

Upgrade/rollback: use this tag in a separate checkout/environment, retain private journals and records, and return to v0.2.0 if needed. No global skill transaction was performed.

Full commit: c4454832b4fe33b565e78df457600a7be74b62e5

Verified CI

Skill Loom v0.2.0

Skill Loom v0.2.0 is a compatible MINOR release from v0.1.1.

• Add an environment-aware recommendation queue, explicit Claude transcript observations and per-host usage review. Recommendations install nothing.
• Include merged fixes from PRs #1 and #2: reject malformed usage reports, deduplicate repeated input filesystem objects, require one report to meet coverage thresholds, preserve positive signals, and keep staging arguments separate from shell text.
• Clarify Codex observation semantics with optional skill_signal_basis metadata. The legacy skill_md_reads field and load-signal status count syntactic tool-payload path references, including quoted code or failed commands. They do not verify file access, skill use or task success.
• Add a dated research supplement and source review. No new dependencies, upstream pin changes or global skill installation.

Validation: 90 local Windows/Python 3.12 tests (88 passed, two platform skips), 83-file release check, isolated install/rollback/reapply/no-change demo, 12/12 synthetic coverage contracts and citation consistency. A normal wheel is validated in a fresh environment before publication. Exact-head hosted checks are linked below.

The original manuscript and PDF retain their historical scope; the dated supplements contain subsequent engineering evidence. No claims of model quality, token savings, native hook portability or macOS validation are made.

Changes · Research supplement · Upgrade and rollback

For a clean install, use Python 3.10+ and a new virtual environment:
python -m pip install "git+https://github.com/Koooki3/skill-loom.git@v0.2.0"

Keep the old environment and observation reports for rollback to v0.1.1. Schema 1 remains compatible; new commands are unavailable in v0.1.1. The observation thresholds describe selected inputs, not complete capture or independent task opportunities.

Full commit: 331c103a2aba6782bc10e47d70e083a39e87674e

Verified CI

Skill Loom v0.1.1

Skill Loom v0.1.1 fixes a malformed evidence-receipt input contract.

A receipt whose top-level JSON value was a list previously produced an uncaught AttributeError in both the CLI and Stop-hook script. Receipt objects, check IDs, artifact collections, paths and digest fields are now validated before artifact bytes are read. Invalid input produces a structured CLI error (exit 2) or an explicit unverified hook result (exit 1). Existing valid receipt verdicts and schema 1 remain unchanged.

Six new regression tests cover malformed input, pre-read validation, valid/incomplete controls and both entrypoints. Local Windows Python 3.12 ran 49 tests: 47 passed and 2 skipped for platform/privilege limits. The isolated self-evolution demo verified exact rollback, reapplication and no-change behavior. Exact-commit Windows/Ubuntu CI is verified before this Release is created.

The patch adds a bounded engineering supplement and an incremental ClaimReceipt literature note. It does not implement that paper's protocol or establish model-quality, token-cost, semantic-evidence, native-hook or macOS gains. The original manuscript/PDF and twelve-case study retain their v0.1.0 scope.

Regression report · Changelog · Validation

Upgrade in an isolated environment, rerun the relevant gate fixtures and retain the prior version. No global skills or host policies were changed.

Full commit: 34efa6e539e2d1da829d05652d6c0df1c1433a7a

Verified CI

Skill Loom v0.1.0

Skill Loom v0.1.0 is a local toolkit for evidence-driven, reversible maintenance of agent skill libraries.

本版将需求档案、来源检索、能力缺口、候选评审、变更计划、验证与回退组织为可运行闭环。

• Python CLI with explicit public discovery, commit-pinned and Git blob-verified staging, capability gap/ranking/portfolio analysis.
• Exact reviewed file plans, content/mode drift checks, interrupted journals and rollback; task/candidate/time-bound evidence receipts and a bounded optional Stop-hook script.
• Codex and Claude Code adapter guides, a portable lifecycle skill, 39-entry anonymized capability catalog, environment/output management and upgrade procedures.
• Fifteen primary-paper research matrix, a six-page English systems technical report, Chinese technical blog, reproducible selected fault scenarios, icon, social card and data chart.

Verified

Windows and Ubuntu CI passed on Python 3.10 and 3.12. Each Linux job ran all 43 tests successfully; Windows jobs passed 42 and skipped the inapplicable POSIX permission test. Release checks also passed.

A separate real Windows Codex maintenance cycle updated two files in an existing maintenance skill, verified rollback, reapplied, and confirmed a no-change run. The local skill count stayed at 39. The selected deterministic experiment satisfied 8/8 mutation and 4/4 evidence contracts; its intentionally simple reference baselines do not support a claim of superiority over other products or academic systems.

Scope

This is a research and engineering prototype. Native Codex/Claude hooks, Claude Code local operation, macOS and downstream model-task improvements remain unverified. The paper is a technical report, not peer reviewed. The reading PDF was generated with ReportLab; standalone TeX is included, but the local built-in TeX compiler could not initialize its platform directories.

Start here

README · English · Validation · Research · Technical blog

Download the source archive, create a Python 3.10+ virtual environment, run python -m pip install -e ., then python -m skillloom demo --workdir ../loom-demo-001. The demo directory must be new; it is an isolated fixture, not an automatic global installation.

Code and original documentation: MIT. Referenced third-party materials retain their own licenses. No raw user traces, host credentials or real private profiles are included.

Full commit: c315324dbc447b88e45059798dee9839727c263a

Verified CI run