v0.1.0-preview.5 — OpenCode skills and subagents

OpenCode joins the library with 16 shared skills and six native subagent profiles. Export once from the canonical sources, then install the whole library or choose individual skills.

From a source clone, install Humanizer into an existing project:

python3 tools/kit.py install --target opencode --project /existing/project --skill mkl-humanize

Omit --skill for all skills and agents. Skills use .opencode/skills/; agents use .opencode/agents/, with embedded workflows and no model or permission overrides. See docs/opencode.md for invocation, updates, removal and discovery limitations.

The five attached client ZIPs contain exports, a license and installation instructions. The Python installer and Skill Watch runtime are in the source archive/clone, not the client ZIPs. SHA256SUMS.txt covers all five bundles. Each ZIP was checked against its source exports and a byte-identical repeat build.

Also includes the Skill Watch baseline-size safeguard merged after preview.4: an oversized write preserves the existing readable baseline. OpenCode files now appear in source-change impact reports.

Validation: 94 unit tests, source and generated-output checks, staged-index guard, executable bugfix and offline Skill Watch fixtures, Ruff, and CI. OpenCode live-client discovery and model task outcomes have not been evaluated; no paid API calls were used. This remains a preview.

v0.1.0-preview.4 — Install only the skills you need

Try one skill with Python, then add others when you need them. This preview adds repeatable --skill NAME selection to installation, updates, and removal for Codex, Claude Code, and Cursor.

Install just Humanizer

With Git and Python 3.11+, clone this release and choose an existing destination project:

git clone --branch v0.1.0-preview.4 https://github.com/00200200/maintainer-skills-lab.git
cd maintainer-skills-lab
python3 tools/kit.py install --target codex --project /existing/project --skill mkl-humanize --dry-run
python3 tools/kit.py install --target codex --project /existing/project --skill mkl-humanize

Replace /existing/project with your project path. Use --target claude or --target cursor for the other coding clients. No Node.js, third-party Python package, or model API key is required.

• Repeat --skill to select several skills. Each selection includes its supporting resources.
• Adding a selection preserves previously installed workflows and their ownership records.
• Use the same selection with uninstall to remove only unchanged owned files for those skills. Local changes within the selection block the operation; unowned files remain unowned.
• Omitting --skill keeps the existing full-library behavior. Native agents embed workflow copies and are preserved during selective updates; use a full-library update to refresh them too.

Installation and updates

Downloads

The library still contains 16 skills and 6 agent profiles. The four client ZIPs contain instructions and agent exports, with updated installation guidance. Grok Bot remains a manual setup recipe.

The Python installer is in the source archive or repository clone, not the client ZIPs. Extract client bundles into a temporary directory and follow their INSTALL.md. Download SHA256SUMS.txt beside all four bundles and verify with shasum -a 256 -c SHA256SUMS.txt on macOS or sha256sum -c SHA256SUMS.txt on Linux.

Verification

All 90 repository tests passed, including 11 new selection tests covering three targets, supporting resources, additive selection, ownership preservation, local edits, dry runs, conflicts, retired skills, adjacent names, and real CLI calls. Source/export checks, the bugfix fixture, the offline Skill Watch example, Ruff lint/format checks, and the staged export guard passed.

All four ZIPs passed integrity and export-content checks and matched a separate deterministic rebuild. These checks establish installer and fixture behavior, not live-client discovery or agent quality. No model was invoked. Pull request and CI.

Changes since preview.3

v0.1.0-preview.3 — Review source changes

A changed documentation page does not mean every dependent instruction is wrong. This preview adds a shared workflow for reviewing the actual claims, identifying supported corrections, and preserving instructions that remain valid.

The library now contains 16 skills and 6 agent profiles for Codex, Claude Code, Cursor, and Grok Bot.

New since preview.2

• mkl-review-source-change works from a supplied diff or existing Skill Watch tools. It checks version applicability, traces the affected owner claim, and leaves incomplete evidence unresolved instead of guessing a fix.
• mkl-source-reviewer combines source review with fix verification when a candidate behavioral fix and reproduction are available. Documentation-only reviews do not require inventing a runtime test.
• Worked review and acceptance cases cover a changed checkpoint expectation, an unaffected instruction, a truncated diff, and a version mismatch. The offline fixture now contains the concrete owner expectation discussed in the review.

Worked review · Skill source · Skill Watch guide

Get the files

Choose one of the four client ZIPs below for the skills and agent exports. Extract it into a temporary directory and follow its INSTALL.md. Grok Bot files are manual setup recipes. Download SHA256SUMS.txt beside the four bundles to verify their checksums with shasum -a 256 -c SHA256SUMS.txt on macOS or sha256sum -c SHA256SUMS.txt on Linux.

The client ZIPs contain the instructions. Use the source archive or clone the repository at v0.1.0-preview.3 for the Skill Watch runtime and runnable examples. The CLI uses Python 3.11+; the optional MCP server installs its SDK separately. Installation and updates.

After installing the skill, try:

Use mkl-review-source-change to review this documentation diff against the affected skills. Identify supported corrections and instructions that remain valid. Flag missing evidence; return a review before making edits.

Validation

All 79 repository tests passed. PR CI passed on Linux and macOS with Python 3.11 and 3.13, including source/export consistency, installation tests, offline examples, and bundle builds. Skill validation, Ruff lint/format checks, and the staged source/export guard passed. The MCP SDK 2.2.0 stdio check passed in automatic and legacy modes with no model invoked.

Built from ef088ed. All four archives passed ZIP integrity checks and matched their canonical exports, including the new skill and agent.

The review scenarios are authored acceptance examples, not independent live-agent evaluations or a review-quality benchmark. Native agent exports inherit the host's model and execution policy. Grok Bot setup remains manual.

Changes since preview.2

v0.1.0-preview.2 — ML diagnostics and Skill Watch

Debug a training run and find agent instructions that need review when their documentation changes. This preview adds ML diagnostics and Skill Watch, and updates all four client bundles to 15 skills and 5 agent profiles.

New in this preview

• ML training diagnostics: mkl-debug-ml-training and the mkl-ml-investigator agent guide investigations of misleading losses, NaNs, disconnected gradients, and reproducibility issues in PyTorch, Lightning, and TensorFlow/Keras. A runnable CPU example compares loss, gradients, and an optimizer update against an analytical result. Example and recorded framework checks.
• Skill Watch CLI and MCP server: compare selected documentation with a saved baseline and locate affected skills, dependent agents, and generated files. Checks preserve the baseline; acceptance requires an explicit current hash. An interrupted HTTP response is an error even if the selected text arrived before the interruption. Usage, MCP setup, and limits.

Download a client bundle

Asset	Contents
maintainer-skills-lab-codex.zip	15 skills and 5 native agent profiles for Codex
maintainer-skills-lab-claude.zip	15 skills and 5 native agent profiles for Claude Code
maintainer-skills-lab-cursor.zip	15 skills and 5 native agent profiles for Cursor
maintainer-skills-lab-grok-bot.zip	15 skill recipes, 5 agent recipes, and first-task setup guides

Extract a bundle into a temporary directory and follow its INSTALL.md. Manual copies have no installer ownership manifest. Grok Bot requires manual setup in the Bot app. Every ZIP includes the MIT license. Installation, updates, and removal.

Download SHA256SUMS.txt alongside the four ZIPs, then verify them with shasum -a 256 -c SHA256SUMS.txt on macOS or sha256sum -c SHA256SUMS.txt on Linux.

Try Skill Watch without a model

Skill Watch and the executable examples are in the repository/source archive, outside the client bundles. With Git and Python 3.11+:

git clone --branch v0.1.0-preview.2 https://github.com/00200200/maintainer-skills-lab.git
cd maintainer-skills-lab
python3 examples/skill-watch/run.py

The authored offline example changes one documentation sentence, finds the dependent agent, and verifies that repeated checks preserve the baseline until explicit acceptance. It uses a disposable project and makes no network requests or model calls.

The CLI uses Python's standard library. Optional MCP integration installs SDK 2.2.0 separately; no MCP client configuration or scheduler is activated automatically.

Verification and limits

Built from bb8e7bc, with passing release-commit CI on Linux and macOS, Python 3.11 and 3.13.

• Re-ran all 79 repository tests, source validation, generated-file checks, bug-regression and offline Skill Watch examples, Ruff lint/format checks, and four bundle builds locally.
• Exercised actual MCP stdio client/server communication with SDK 2.2.0 on Python 3.12.11, in automatic and legacy negotiation modes; all three tools were discovered and checked without invoking a model.
• Every attached ZIP passed integrity checks, matched the checked-in provider files, contained the new ML skill and agent, and was byte-identical to a separate rebuild.

The three ML framework runs are previously recorded local fixture checks linked above; they were not repeated for packaging this release. Source/export validation, fixture execution, and MCP protocol checks do not establish live-agent diagnosis, writing quality, model performance, or client discovery. Windows and live Grok Bot execution remain untested. Compatibility and evidence.

v0.1.0-preview.1 — Humanizer and maintainer workflows

Try Humanizer on a draft, or bring the full maintainer library into your coding project. This first preview includes 14 skills and 4 agent profiles maintained from shared sources.

Start with one skill

From your project directory, with Node.js 22.20.0+ and Git:

npx skills@1.5.26 add 00200200/maintainer-skills-lab --skill mkl-humanize --agent codex --copy

Use --agent claude-code or --agent cursor for the other coding clients. This command uses the third-party Vercel Skills CLI and follows the repository's default branch. Use the attached archives for the exact content of this release.

Humanizer example and usage · Installation and removal

Choose a bundle

Download	Contents
maintainer-skills-lab-codex.zip	14 skills and 4 native agent profiles for Codex
maintainer-skills-lab-claude.zip	14 skills and 4 native agent profiles for Claude Code
maintainer-skills-lab-cursor.zip	14 skills and 4 native agent profiles for Cursor
maintainer-skills-lab-grok-bot.zip	14 skill recipes, 4 agent recipes, and first-task setup guides

Extract a bundle to a temporary directory and follow its INSTALL.md. Grok Bot requires manual setup in the Bot app. ZIP copies do not have the Python installer's ownership manifest. Every bundle includes the MIT license.

SHA256SUMS.txt covers all four ZIPs. Download it alongside the archives and run shasum -a 256 -c SHA256SUMS.txt on macOS or sha256sum -c SHA256SUMS.txt on Linux.

The source archive also includes the Python library installer, the worked bug-regression example, and an optional Git pre-commit hook for contributors editing shared skills. These tools are in the repository, outside the client bundles.

Verification and limits

Built from 73fc1f3.

• All 45 tool and fixture tests passed, with source validation, generated-file checks, lint/format checks, and four bundle builds. Release-commit CI passed on Linux and macOS with Python 3.11 and 3.13.
• Each attached ZIP passed integrity checks, matched the source exports, and was byte-identical to an independent local rebuild.
• Skills CLI 1.5.26 discovered all 14 source skills. Humanizer installation, listing, and removal were checked in separate temporary projects for Codex, Claude Code, and Cursor, with telemetry disabled.

This is a preview. These checks do not evaluate live-client loading, writing quality, model performance, Windows support, or Grok Bot execution. Recorded evidence and known limits.