v0.6.0 - Security hardening

Highlights

• Removed the experimental in-process embedding-provider loader, eliminating
  the path that could import and execute a user-selected Python module or file.
• Limited bundled retrieval to Agent-ranked host input and deterministic,
  zero-dependency lexical search over a user-provided local corpus.
• Replaced the runner adapter's dynamic sibling-module loader with a fixed
  import and changed upstream dependency probing to distribution metadata only.
• Aligned the skill contract, public documentation, contributor guidance, and
  automated regression checks with the new data-only retrieval boundary.

Upgrade note

The experimental embedding backend and its --embedding-provider, --model,
--provider-version, and --fallback options have been removed. Existing
workflows should use host for rankings produced by the current Agent or
runner, or lexical for deterministic retrieval inside a local corpus.

Release verification

• 87 unit tests passed.
• 120/120 isolated offline contract-canary runs passed across 12 cases with
  zero illegal transitions.
• Skill validation, Python compilation, pinned skills CLI discovery/rendering,
  dangerous dynamic-loader pattern scanning, and git diff --check passed.

Known limitations

• Retrieved sources and external runner output remain untrusted data and still
  require the existing prompt-injection, citation, and artifact-validation
  boundaries.
• skills.sh security results are generated externally and may continue to show
  the previous release until the repository is re-indexed.
• Co-STORM remains a prompt-native preview with simulated participants and does
  not claim upstream CoStormRunner parity.

Install or update

npx skills add lizhouai/storm-research-skill

For an existing project-local installation:

npx skills update storm