v0.3.0

skilled 0.3.0 adds project setup for Codex and Claude Code and repairs evaluation scoring, execution boundaries, and long-running MCP workflows.

Install and activate

curl -fsSL https://github.com/DeadHead04/skillEd/releases/latest/download/install.sh | bash
export PATH="$HOME/.local/bin:$PATH"
skilled --version
cd /path/to/your/project
skilled setup --host codex

For Claude Code, use skilled setup --host claude-code. The self-contained CLI includes its Node runtime. skilled-improve is the bundled agent workflow, not a second terminal executable. Setup preserves existing Skills and project instructions; repeated setup, edited-file protection, and removal are covered by installation acceptance.

Evaluation repairs

• Correct negative-rule scoring, skipped assertions, gate failures and comparisons of actual rule outputs.
• Preserve sandbox and approval policy when Codex resumes a session.
• Run with/without-Skill benchmarks only with a verified Docker filesystem boundary. Native evaluations remain available; native and OpenSandbox benchmarks fail before execution. Grading uses a clean container and validated workspace snapshot.
• Start durable MCP jobs with skilled_start, recover by idempotency key, and use skilled_status, skilled_results, and skilled_cancel. Reconnection does not launch duplicate work. Concurrent jobs use separate report directories.
• Preserve requested HTML and JUnit reports alongside JSON. Synchronous MCP cancellation remains supported.
• Add explicit Docker Codex login-file forwarding with credential validation and cleanup.

Verification and limits

Real authenticated native Codex acceptance reproduced 13/18 baseline passes, 18/18 improved passes and 10/10 independent holdouts. An intentional regression failed and restoration passed. A 331-second evaluation survived MCP reconnection with a 67ms start response. Native resume acceptance denied an out-of-workspace write. A fresh installed Codex session discovered the workflow and validated the policy suite.

Source CI and release builds require actual Docker boundary tests. Release archives are executed on native Linux, macOS and Windows runners before publication; the installation jobs subsequently check the published version and latest installers. See the linked Actions run for exact results rather than treating an archive build as proof of every device/provider combination.

Final release acceptance passed all eight jobs, including published versioned/latest installers on Linux, macOS and Windows. Final source CI recorded 1,586 passes, zero failures and 41 explicit skips; the separate real Docker suite recorded 47 passes and two OpenSandbox skips. A fresh public download also passed setup for both hosts without GitHub login or system Node. One final authenticated Codex case through the published CLI passed after 42 seconds, recovered the same durable job after reconnection, and produced JSON, HTML and JUnit reports.

Optional Docker subscription-login model execution remains unverified: local Docker storage failed before that end-to-end test. Credential validation and private-file staging were tested, and authenticated native evaluations passed. Run a small Docker evaluation before relying on this optional login path. OpenSandbox benchmark isolation and every provider/OS/CPU combination are not certified.

Setup documentation · Verification scope

Release acceptance workflow

skilled v0.2.0 — integrity, comparison, and MCP

One installation for validating, comparing and improving agent skills from the CLI or a compatible MCP host.

What changed

• Corrected scoring: prerequisite gates no longer inflate judge assertion scores. A failed gate fails the case and marks unexecuted judging as skipped, with zero benchmark credit.
• Added skilled validate --skill <directory> for metadata and direct local-reference checks.
• Added installation selection checks and SHA-256 readback of installed files.
• Added skilled compare <before.json> <after.json> --fail-on-regression, with separate coverage and provenance warnings.
• Added skilled mcp --workspace <directory> with four tools, the skilled-improve prompt, and bundled workflow references and templates. Actual execution can be enabled with --allow-execution.
• Added cooperative evaluation cancellation, bounded MCP responses, and agent-assisted setup instructions.

Install or update

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/DeadHead04/skillEd/main/install.sh | bash
export PATH="$HOME/.local/bin:$PATH"
skilled --version

Windows PowerShell:

$installer = Join-Path $env:TEMP 'skilled-install.ps1'
Invoke-WebRequest https://raw.githubusercontent.com/DeadHead04/skillEd/main/install.ps1 -OutFile $installer
& $installer
$env:Path = "$(Join-Path $HOME '.local/bin');$env:Path"
skilled --version

The release bundles Node and application dependencies. No separate npm install or improvement-workflow download is required. The npm tarball is a downloadable asset; no npm registry package is published.

Agent-assisted setup · MCP setup · Comparison and scoring

Compatibility and verification

Refresh pre-0.2.0 baselines that mix expect gates and single-judge assertions. Their assertion scores use different semantics. Offline comparison reports observed changes; it is not autonomous drift monitoring or proof that a Skill edit alone caused a change.

The local cumulative gate passed 1,568 tests, with zero failures and 26 optional checks skipped. Real tests include files, processes, Docker installation/readback, SDK MCP client/server interactions, actual local engine runs, timeouts and cancellation. All 60 existing workflow assertion regressions also passed, and real Chrome report navigation/feedback completed without JavaScript errors. Local macOS ARM64 and Linux ARM64 Docker archives passed CLI/API/report/MCP acceptance without a system Node installation.

Native source CI and the release workflow verify their exact commits and accepted archives on Linux, macOS and Windows. Read the linked workflow evidence for job outcomes; native runner architectures do not certify every CPU variant.

Authenticated model engines, model judges and user simulators still require separately configured credentials and a spending limit. Those provider-backed checks were not run for this release and local process tests do not certify model quality.

Exact-commit CI · Release acceptance and published installation checks

All eight release jobs passed, including native archive execution and published install/update checks on Linux, macOS and Windows. An additional anonymous macOS ARM64 check passed pinned installation, latest reinstall, installed CLI validation/dry-run, and the installed MCP server with four tools and ten bundled workflow resources.

v0.1.0

skilled 0.1.0

First versioned release of the TypeScript CLI, with English documentation and installers for macOS, Linux and Windows.

Install

The repository and release assets are public. The CLI includes Node and application dependencies; no separate Node, Go, or npm installation is needed.

macOS and Linux

curl -fsSL https://github.com/DeadHead04/skillEd/releases/download/v0.1.0/install.sh -o install.sh
SKILLED_VERSION=0.1.0 bash install.sh
export PATH="$HOME/.local/bin:$PATH"
skilled --version

Windows PowerShell

Download install.ps1 from this release and run ./install.ps1 -Version '0.1.0'. Add $HOME\.local\bin to your PATH. No GitHub login or -UseGh option is needed.

See the installation guide for updates, manual extraction, and removal, and getting started for a first evaluation.

Included

• CLI and execution API, local/Docker/OpenSandbox environments, agent adapters and MCP integration.
• Evaluation rules, script/model judges, iteration and baseline runs, JSON/JUnit/HTML reports.
• skilled-improve, codex-skilled and dsh-skilled workflow/plugin archives.
• Six CLI archive targets, an exact npm tarball, SHA-256 checksums and source provenance.
• English setup, evaluation, configuration and reference guides.

Model-backed evaluations require the selected engine/provider and credentials. Installing the CLI does not configure a model account. Native CI checks the archive matching each Linux, macOS and Windows runner; other architectures are not certified merely by building them. Authenticated model scenarios remain outside package acceptance.

All standalone distributions retain LICENSE and NOTICE. The npm tarball is attached here; this release does not publish the unqualified skilled npm registry name or a container image.