MCP Server v0.0.1-beta.16

This release strengthens protocol compatibility, catalog refresh, documentation
search, and connection handling while preserving existing Convert permissions
and authentication. It includes the session-hardening fixes shipped to the
hosted service after beta.15.

Modern MCP, Existing Clients Preserved

• The hosted endpoint supports MCP 2026-07-28 alongside the existing protocol
  adapter. Requests use the appropriate SDK implementation; unsupported versions
  are not silently rewritten to another version.
• Tools, prompts, resources, schema validation, permission checks, and
  destructive-action confirmations remain shared across both adapters.
• The npm package remains a local stdio server with the same API-key/HMAC setup.
  Worker-only catalog and runtime files are excluded from the installed package.
• Existing OAuth clients, grants, callbacks, and account/project scopes are
  preserved. A protocol or catalog update does not require new consent.

Catalog Refresh And Session Reliability

• Modern clients can subscribe to catalog-change notifications. Delivery is
  isolated to the same authorization grant, deduplicated, and protected against
  stale observations during rolling deployment.
• Staff and customer catalog fingerprints are independently cached, avoiding
  spurious refresh notifications between differently decorated catalogs.
• Subscription buffers and lifetimes are bounded, with authorization revalidation
  and explicit refresh as the fallback. Client support still determines whether
  notifications trigger an automatic refresh or resubscription.
• Stronger Durable Object session ownership, lifecycle diagnostics, and startup
  recovery preserve authentication boundaries. Writes are never automatically
  replayed as a recovery mechanism.

Documentation Search And Cancellation

• Hosted search uses an isolated public-knowledge compute boundary and a compact
  prebuilt lexical index, reducing repeated parsing across request/session
  contexts while preserving semantic retrieval and support-first ranking.
• The 9,738-chunk knowledge corpus, document-ID compatibility, and direct fetch
  support remain intact. Local semantic search retains its native vector store;
  hosted semantic search retains Workers AI and Vectorize.
• Request cancellation propagates through API transport, retry waits, recovery
  reads, asset loading, and documentation search. For native model inference
  that cannot itself be interrupted, the server stops waiting and discards the
  result rather than starting further vector work.

Compatibility And Rollout

No customer configuration change is required. Hosted customers keep the same
endpoint. Local users configured with npx -y @convertcom/mcp-server@latest
receive the published version on their next launch, subject to their client's
normal package caching; already-running local processes must restart.

Notification settings/history remain disabled by default pending their separate
backend rollout. This release does not enable those tools or change Convert UI
notification settings. Explicit account-wide grants, staff browse-only
impersonation, approval requirements, and two-step deletion confirmation remain
unchanged.

Verification

The implementation passed the deterministic release gate, local installed-binary
package checks, strict Worker builds, native workerd protocol/lifecycle tests,
independent review, and authenticated staging checks. Publication additionally
requires the committed beta.16 release gate.

Production acceptance covered both protocol generations, account/project reads,
docs search/fetch, catalog discovery, and modern subscriptions. One documentation
search in a legacy session held across the initial production deployment timed
out; later legacy sessions and repeated same-session reads passed without OAuth
reauthorization. This is not a promise of interruption-free transport handover.
See the dated production rollout record for exact versions, timestamps, and
observability limits.