[FEATURE] 资源管理支持从标准 Agent Skills / Agent Plugins 仓库批量导入

[FenjuFu]

Problem

Astron Agent already supports Skill folders in Resource Management and the standard Agent read_skill / run_skill runtime added by #1427. The missing step is repository-level ingestion: operators must clone a collection locally and upload every file through the browser before any Skill becomes selectable.

K-Dense-AI/scientific-agent-skills is a concrete standards-based fixture:

plugin.json
skills/
  scanpy/SKILL.md
  rdkit/SKILL.md
  literature-review/SKILL.md
  ...

It is an Agent Plugins 1.0.0 package with 160+ immediate skills/<name>/SKILL.md directories. The current /skill-file/upload-directory path accepts browser-provided multipart files and paths, but it has no repository URL/ref, import plan, collection metadata, source provenance, repeatable update path, or per-skill selection. Files rejected by the upload allowlist can also be skipped without giving an operator a repository-level compatibility report.

Proposed Solution

Add a dry-run-first external Skill repository importer to Resource Management.

Suggested MVP:

1. Accept a public HTTPS Git repository URL, a required tag/commit SHA, and an optional skills subdirectory (default skills/).
2. Fetch into an isolated temporary workspace with strict URL, redirect, timeout, total-size, file-count, and per-file limits. Reject private/reserved network targets; do not accept embedded credentials, submodules, Git LFS downloads, or arbitrary Git config.
3. Read root plugin.json only as optional collection/source metadata. Do not import hooks or execute repository code.
4. Discover immediate child directories containing SKILL.md, parse the standard frontmatter (including nested metadata.version), and show a dry-run table with name, description, version, license, path, file count/size, unsupported files, and validation errors.
5. Let the operator select a reviewed subset before import. Import each selected Skill as a normal Resource Management folder so existing authorization, storage, audit, and Agent attachment flows remain authoritative.
6. Persist provenance per imported Skill: repository URL, requested ref, resolved commit SHA, source path, content fingerprint, and imported timestamp.
7. Define deterministic collision behavior (skip, replace-unattached draft, or rename); never silently overwrite an existing folder. Return per-skill results so one invalid Skill does not hide successful imports.
8. Keep scheduled synchronization and automatic upstream updates out of the MVP. A later sync must produce a diff and require explicit approval.

Security Boundary

• Import is inert: no script, hook, package install, or Skill instruction runs during discovery/import.
• Archive/path traversal and symlink escapes must be rejected before storage.
• Unsupported or oversized files must be reported, not silently treated as a complete import.
• Repository credentials and private repository support are out of scope for the first slice.
• Imported Skills still use the existing sandbox behavior when an Agent later invokes run_skill.

Acceptance Fixture: Scientific Agent Skills

Use K-Dense-AI/scientific-agent-skills@v2.64.0 as an integration fixture without downloading model/data dependencies:

• The importer resolves and records the immutable commit behind v2.64.0.
• Dry-run discovers scanpy, rdkit, and literature-review as separate Skills from skills/.
• Nested metadata.version is preserved for each Skill.
• Every included file is either planned for import or reported with an explicit incompatibility; nothing is silently dropped.
• Selecting only those three Skills imports three independent Resource Management folders.
• Re-running the same ref is idempotent and reports unchanged content.
• An imported Skill appears in the existing Agent Skill selector and works through the existing read_skill path; runtime execution behavior remains governed by current sandbox configuration.
• Malicious URL, redirect-to-private-IP, traversal, symlink escape, oversized archive, and duplicate-folder tests are included.

Alternatives

1. Manual folder upload: works for a few Skills but loses repository/ref provenance and does not scale to a reviewed collection.
2. Publish through SkillHub first: useful for governed team distribution and documented in K-Dense PR Feature/fe refactor #241, but Astron Resource Management still lacks a direct repository intake path.
3. Treat the whole Agent Plugin as one Skill: rejected because each immediate child has its own trigger, metadata, license, dependencies, and lifecycle.

Related

• Runtime Skill support: 标准 Agent 新增 Skill 模块（read_skill / run_skill） #1427 / feat: standalone agent Skill capability (read_skill / run_skill) #1424
• Scientific Agent Skills: https://github.com/K-Dense-AI/scientific-agent-skills
• K-Dense SkillHub/Astron guide: docs: add SkillHub/Astron governed publishing guide K-Dense-AI/scientific-agent-skills#241
• SkillHub external repository importer design: [Feature] Import & sync skills from community skill repos (agent-skills style layouts) skillhub#516