[Iamsdt]

This brings the docs in line with the security fixes landing in three repos:

• Core: 10xGraph/10xGraph fix/fetch-ssrf-file-tool-root
• API: 10xGraph/10xgraph-api fix/reserve-core-config-keys
• Client: 10xGraph/10xgraph-client fix/security-review

Several of those changes break existing code or deployments, so the guides, references and examples are updated to match. Pages that showed APIs that no longer exist (the graph-level memory API) or unsafe code (eval() in a calculator tool) are corrected. Merge this when those three PRs are merged and released.

The PR also adds one new blog post and extends an existing one.

Core (Python library)

• Memory.
  • The graph-level memory API (MemoryIntegration, create_memory_preload_node, memory_tool) is removed from every page. Memory is configured only with Agent(memory=MemoryConfig(...)).
  • Documented the per-scope collection (a name or a function of the run config) and its precedence.
  • Documented the reserved shared owners from system_user_id(name), and that every store operation needs a non-empty user_id.
  • Preload memories are now described as a fenced <user_memories_...> block in the latest user message, not a system message.
  • The Qdrant pages now require qdrant-client and server 1.10 or newer.
• RAG. RAGAgent docs cover store_config_builder, the default knowledge-base owner, and the request-only <retrieved_documents_...> block.
• Tools.
  • create_file_tools(root=...) is the way to set the file root; the run-config root is deprecated.
  • Documented the default deny list for secret files, max_parallel_tool_calls, and the generic "internal error (reference ...)" text a model now sees.
  • Handoffs need a registered handoff tool.
• Media. MediaFetchPolicy and compile(media_policy=...) for client-supplied media URLs.
• Interrupts. resume_interrupt_id, response_schema validation, InterruptMismatchError and ResumeValidationError, with their codes on the error-codes page.
• Prebuilt agents. SupervisorTeamAgent, PlanActReflectAgent and StructuredOutputAgent limits are per turn.
• Observability. user_id_mode on the OTEL, Logfire and LangSmith setup, and Kafka TLS/SASL options on the publishers page.
• Examples. The graceful-shutdown example and pages that copied it use safe_calculator instead of eval(). The RAG example is updated.

API server

• Thread ids. Documented the runs.thread_ids: "server" default: start a new thread by omitting the id, and read meta.thread_id from the response.
• Busy threads. Documented the per-thread run lock: a 409 "thread is busy", the on_busy option, and how this works on WebSocket and AG-UI.
• Resume. Resume with interrupt_id: 409 on a mismatch, 422 when the schema check fails.
• Client input rules. Client messages must have the user role. Built-in state fields can't be written by clients. The server generates message ids. Request size limits are listed. Media URLs are checked before the run.
• New routes. PUT /v1/threads/{id}/messages/{message_id} (feedback) and DELETE /v1/files/{file_id}. Message import (POST .../messages) needs the checkpointer:import scope.
• Auth. Documented __system__ ids being refused, JWT user_id and scope-claim validation, the bearer token on HTTP coming from the Authorization header only, the token re-check on /v1/graph/live, and owner isolation as the default for custom authorization backends.
• Production. Covered MODE validation, STRICT_PRODUCTION_CHECKS on by default, the 300 s runs.timeout default, hidden error text in stream chunks, and observability.user_id_mode. The production checklist and troubleshooting pages are updated.
• Deployment. Docker, compose and Kubernetes pages match the locked-down generated files: non-root uid, read-only filesystem, compose bound to 127.0.0.1, and Secrets for Kubernetes.

TypeScript client

• New options. on_busy, resume and interrupt_id on invoke, stream and wsStream.
• New error. ConflictError on the errors reference and error-handling page.
• New methods. deleteFile and updateThreadMessageFeedback.
• Deprecation. addThreadMessages is deprecated, with the admin-scope note.
• fixGraph. Corrected what the repair does.

Blog

• New post: "Moving past vibe checks: deterministic agent evals in CI" (deterministic-evals-for-ai-agents.mdx). It shows how to gate pull requests on tool trajectories with eval sets and pytest, with no judge model.
• Updated: "hot-and-cold-agent-memory" gains a section on the PgCheckpointer Postgres schema and queries.

Checks

• npm run build succeeds (search index, llms.txt, markdown twins)
• Code samples checked against the released packages
• Links checked