v9.5.2

v9.5.2 - Kimi K3 Tool-Call Security

Added

• Kimi K3 provider support now defaults kimi / moonshot to kimi-k3 while keeping legacy Kimi models selectable.
• KIMI_API_KEY is accepted alongside MOONSHOT_API_KEY.
• red-team --gpt-red --k3-long-context expands GPT-Red context for Kimi K3 runs with package scripts, CI workflows, deployment config, docs, MCP/tool manifests, and prior findings under hard caps.
• Agentic tool-call security checks for Kimi K3 / OpenAI-compatible agents:
  • dynamic tool definitions loaded from prompt/RAG/tool-result context
  • model-selected tool names executed without an allowlist
  • forced tool choice with untrusted user input
  • tool-result replay without preserving the assistant tool-call message

Changed

• Kimi reasoning output is treated as internal reasoning and is never parsed as final Ship Safe JSON output.
• README, docs, provider UI, and AI-readable files now document Kimi K3 and long-context GPT-Red mode.

Verification

• npm test - 261/261 passing.
• npx tsc --noEmit in webapp - passing.
• git diff --check - passing.
• npm pack --dry-run --json - package dry run passed for ship-safe@9.5.2.
• Clean fixture smoke test: red-team --gpt-red --k3-long-context --json --no-ai --no-deps exited 0 and ran GPTRedAgent.