Skip to content

Changelog ​

The current hosted production private index offers Python 0.7.0, 0.8.0, 0.9.0 and 0.10.0, and npm 0.1.0 and 0.2.0. Its defaults remain Python 0.10.0 and npm 0.2.0; the wizard pins the available version.

Python yi-guard-sdk 0.11.0 and npm @yilabs/guard 0.3.0 ​

Immutable release artifacts were published on 9 November 2026. These renamed packages will become available from the hosted production private index after the control-plane deployment adopts their release records. They are not yet the production installation defaults.

Existing releases retain their package names. The Python module preworld_guard, CLI preworld-guard and environment variable names stay compatible; install old and new Python distributions in separate environments because they provide the same module.

Python preworld-guard-sdk 0.10.0 and npm @preworld/guard 0.2.0 ​

Released 4 November 2026. These versions keep their original package names.

  • Project identity is sent on every SDK request; mismatched project tokens are refused before tool execution.
  • Missing trusted subjects are handled only when the other rule conditions match. Invalid or oversized regular expressions are rejected when loading a policy.
  • Persisted audit events are retried after a client restart, including before a new event is produced. Refusals before execution are audited with bounded rate limiting.
  • The Python sandbox inherits a minimal environment; pass additional variables explicitly. Timeout cleanup ends the process group. Command hooks fail closed and record incomplete calls as unknown.
  • Gateway credential scanning covers escaped and streamed responses; MCP scans structured content too. These checks do not constitute general content inspection.

Python SDK 0.9.0 and npm @preworld/guard 0.1.0 ​

On the private index since 15 October 2026.

  • Five more framework adapters: Microsoft Agent Framework, LlamaIndex, Agno, DSPy and CrewAI, all on the shared decision path. See Framework adapters.
  • TypeScript SDK @preworld/guard 0.1.0, with an adapter for the Vercel AI SDK (ai@7). See TypeScript SDK.
  • SUBJECT_REQUIRED. An action without a trusted subject that carries an argument a rule compares with session.subject_id is now decided like a deny of that rule, under the policy's mode. Before, such a rule silently stopped matching, so a deny rule could fall away. This holds on every path: adapters, @guard.tool, the MCP proxy, egress, fallbacks and previews.
  • Identity failures fail closed. A resolve that raises or returns something other than an identity gives GUARD_IDENTITY_UNAVAILABLE. The model never sees the error.
  • One retry for connection errors to the control plane, only where it can't repeat a side effect. A request can now take up to twice the client timeout.
  • Event queue per agent. Agents started from the same folder no longer share a queue, which could get one agent blocked with 403.
  • Sets are ordered by their elements' canonical JSON, so list rules and approval digests no longer depend on the process.
  • openai-agents is allowed from 0.20 up to (not including) 0.24.

0.8.0 ​

On the private index since 12 October 2026.

  • One decision path for all adapters, with one payload for the model (the MCP proxy's): status, reason_code, message, policy_version, and the ticket fields for approvals. Anything that prevents a decision raises GuardUnavailable, and the tool doesn't run.
  • New adapters: LangGraph (guard_tool_node), Pydantic AI (guard_hooks), Google ADK (GuardPlugin).
  • guard_resume(ticket_id) lets an adapter run an approved call once, when the model proposes it again.
  • A used, expired or mismatched ticket is now a denial (APPROVAL_NOT_CONSUMABLE). A ticket that meets a newer policy is POLICY_CHANGED. New errors GuardSubjectRequired, GuardFallbackUnavailable, GuardRequestRejected. New methods approval_status, cancel_approval, prepare_fallback, execute_fallback, aexecute_fallback.
  • OpenAI Agents: by default a denial or pending approval is now the tool's result and the run continues. on_deny="raise" keeps the old behaviour.
  • Claude Agent SDK: records started before an allowed tool and failed calls via PostToolUseFailure. Needs claude-agent-sdk>=0.1.26.
  • LangChain: GuardMiddleware returns approval tickets to the model and runs fallbacks for degrade rules.
  • import preworld_guard no longer imports any framework.

0.7.0 ​

  • The Claude adapter fails closed. GuardHooks never raises. Any failure to decide denies the tool with GUARD_UNAVAILABLE. When Guard allows, your own permission rules now still apply (on_allow="allow" restores the old behaviour).
  • Explicit proxy and CA for control-plane traffic: GuardClient(proxy=..., ca_bundle=...), or PREWORLD_GUARD_PROXY and PREWORLD_GUARD_CA_BUNDLE. HTTPS_PROXY and SSL_CERT_FILE are still ignored.
  • Parallel tool calls share one policy refresh.
  • The policy is refreshed correctly after the host wakes from sleep.

0.6.2 ​

  • The MCP proxy's HTTP downstream checks Host and Origin (--allowed-host, --allowed-origin).
  • Progress from upstream servers is passed on. Unknown tools and missing resources return JSON-RPC -32602. Embedded resource text is scanned too.

0.6.1 ​

  • The authorization bridge used by the managed gateway.

0.6.0 ​

  • Per-user OAuth to MCP upstreams, kept in an encrypted vault on your side: preworld-guard mcp-auth connect, status, disconnect. The oauth extra.
  • MCP sampling, form and URL questions from upstream servers can be passed on, each checked and audited.

0.5.0 ​

  • preworld-guard connection-check for a connection package.
  • Asset manifests and automatic inventory of wrapped tools.
  • Offline policy replay (preworld-guard policy-replay). Cloud previews say "not enough information" instead of guessing.

0.4.0 ​

  • Outbound connections check every DNS answer and connect only to checked addresses. Private networks need an explicit range.
  • The HTTP broker handles model requests, including streaming.
  • The identity extra: JWT checks for the broker and the MCP proxy. Upstream credentials are chosen per trusted subject.

0.3.0 ​

  • GuardClient.start(): executor registration, heartbeats and confirmation of the loaded policy.
  • Approval snapshots and resume on your side (@guard.tool(..., pending=True), guard.pending.resume).
  • preworld-guard diagnose.
  • The gateway extra for the HTTP broker.