Skill Studio for DeepSeek Harness

A Codex-compatible skill system for DSH, with a first-class page in the left sidebar. It reads and writes plain SKILL.md folders, so the same skills work in Codex, DSH, and the shared .agents directory.

What it does

• One Skill page in the sidebar — browse every Codex, DSH, and shared skill, search it, read its instructions, and see which copy is effective.
• Create, edit, duplicate, delete, and restore — writes the Codex SKILL.md format with YAML frontmatter. Deletions go to a recoverable .trash folder inside the same scope.
• Enable or disable per surface — disable-model-invocation controls the model catalog and the skill tool; user-invocable controls direct user invocation.
• Codex scopes — ~/.codex/skills and <project>/.codex/skills are bridged into DSH's skill registry, so existing Codex skills appear in the agent catalog without copying them.
• Invoke in conversation — type $skill-name in a chat to inject the full <skill_content> block into that turn, exactly like Codex. The existing /skill-name form also keeps working.
• Progressive disclosure — the model receives only names and descriptions; full instructions load on demand through the skill tool or $name.
• Codex presentation metadata — agents/openai.yaml interface and policy.allow_implicit_invocation, plus legacy config.toml [interface] fields such as displayName, shortDescription, brandColor, and defaultPrompt, are shown when present.

Scopes

Scope	Location	Writable	Notes
Codex user	~/.codex/skills	yes	Bridged into DSH by this plugin
Codex project	<project>/.codex/skills	yes	Nearest .git ancestor, or the workspace directory
DSH user	~/.dsh/skills	yes	Native DSH filesystem provider
DSH project	<project>/.dsh/skills	yes	Native DSH filesystem provider
Shared Agent	~/.agents/skills, <project>/.agents/skills	yes	Visible and editable when present
Codex built-in	~/.codex/skills/.system	no	Read-only; duplicate into a writable scope
Bundled/runtime	DSH packages	no	Read-only; duplicate when possible

DSH_CODEX_HOME (or CODEX_HOME) changes the Codex home; DSH_HOME and DSH_AGENTS_HOME change the DSH and shared roots. The Codex home is scanned under <home>/skills.

A project copy outranks a user copy; a DSH copy outranks a Codex copy of the same name. The page marks the losing copy Shadowed and shows both, so a stale duplicate is never invisible.

Skill format

---
name: release-checklist
description: Use when preparing and verifying a production release.
whenToUse: Before tagging a release.
---

# Release checklist

## Workflow

1. Run the project's own checks.
2. Inspect the final diff.
3. Report the result and remaining risks.

A skill is a directory bundle <name>/SKILL.md (recommended) or a flat <name>.md file directly under a scope root. The only required frontmatter fields are name (kebab-case) and description. Optional companion files can live beside SKILL.md (references/, scripts/, assets/, agents/openai.yaml, config.toml, icon.svg).

disable-model-invocation: true hides the skill from the model catalog and the skill tool. user-invocable: false removes it from the $name / /name gesture. Both default to allowed.

Using a skill in chat

$release-checklist verify the 2.4 release and list anything blocking it

The skill's instructions enter the current step as a <skill_content> block, and the catalog tells the model to follow it rather than load it again. Typing /release-checklist uses DSH's existing user-skill gesture and has the same effect.

The model can also load any model-invocable skill with the existing skill tool, which returns the same canonical <skill_content> shape.

Installation

Install directly from the public GitHub repository:

dsh plugin --profile <profile> add github:PopCapEA/dsh-skill-studio

Or install the release tarball:

dsh plugin --profile <profile> add "https://github.com/PopCapEA/dsh-skill-studio/releases/download/v1.0.0/dsh-skill-studio-codex-1.0.0.tgz"

Or, after the npm release, use the package name:

dsh plugin --profile <profile> add dsh-skill-studio-codex

Refresh the page after installation; the Skills icon appears in the left panel list.

Design notes

• The host half is plain Node ESM with only node: imports. It registers a codex-studio provider on ctx.skills, owns the authenticated /api/skill-studio-codex.* routes through ctx.connection.fetch, and installs one agent/pre-step listener for $name.
• The client half is a plain-JavaScript module loaded by window.__ModuleLoader__; it uses only React, theme tokens, and the public slot/locale services.
• File writes are atomic (temp file plus rename), names are validated, and every mutating route is authenticated by the existing browser session.
• Deleting a skill renames its bundle into <scope>/.trash/ with a sidecar record, so Trash → Restore works without a database.
• The Codex provider invalidates its catalog on a debounced filesystem watch, so edits made in Codex, VS Code, or a shell reach the next model step.
• This published edition uses its own Loader row, sidebar panel id, API prefix, locale namespace, and provider name. It can coexist with the local @local/dsh-skill-studio build instead of replacing it.

Limits

• Editing always updates SKILL.md frontmatter; comments on managed keys are not preserved when the key is rewritten. Unknown frontmatter keys and metadata: blocks are preserved.
• Renaming a skill is intentionally not done in place: the name in SKILL.md is its identity. Duplicate and delete the old copy instead.
• Symlinked skill bundles are visible but read-only.
• The page manages the default workspace for project scopes. Per-session workspaces with a different cwd are still discovered by the model, but the management page uses the process workspace.