server@0.68.0

Minor Changes

• c409a84: Assistant runtimes can now export agent traces (turns, tool calls) over OTLP
  to any OpenTelemetry-compatible backend such as Sentry, Datadog, or Honeycomb.
  Export is enabled by configuring an OTLP endpoint for assistant runtimes, with
  gRPC and HTTP transports supported; traces are tagged with the assistant and
  project they belong to.
• bedfe84: Add backend risk policy bypass request workflow support for the risk-owned request URL flow, backed by current-state request records and principal grants.
• 1dda609: Add prompt-based (LLM-judge) risk policies. Risk policies gain a policy_type discriminator (standard | prompt_based) plus prompt and model_config fields. A new llm_judge evaluation is wired into the realtime enforcement scanner (scoped to tool-call messages) and the batch analyzer, with findings flowing into risk_results. The feature is gated behind the gram-prompt-policies flag.
• cc9d8ee: Add optional name (display name) and logo_asset_id to remote session issuers across both the project-scoped (remoteSessionIssuers) and organization-scoped (organizationRemoteSessionIssuers) services. On create, name is trimmed and stored as NULL when empty; on update it follows the same three-state semantics as the nullable endpoint fields (omitted keeps, empty string clears). logo_asset_id is set-only for now (no clear path, no upload UI yet). The dashboard renders the display name as the primary label with the issuer URL as the secondary line, exposes an optional Display name input on the attach and modify sheets, and renders a logo when one is present. On the attach sheet the Display name auto-derives from the Issuer URL hostname until the operator edits it, matching the existing Slug behavior.

Patch Changes

• 06b1f0d: Add generic access webhook event names for audit logs. Shadow MCP approval requests now emit audit_log.access_request_event_v1, and access rules emit audit_log.access_rule_event_v1; the previous Shadow MCP-specific event names remain in the webhook catalog with deprecated descriptions for compatibility.
• ba8bdd4: Direct assistant MCP authentication prompts to the assistant's owner instead
  of whoever happened to trigger the assistant. Slack onboarding now records the
  owner's Slack identity in the assistant's instructions, runtime guidance
  delivers OAuth links to the owner (ephemeral or DM) and tells anyone else that
  the owner has to complete the connection, and prompts shown when the owner is
  unknown now say explicitly that authentication is for the owner — so an
  unexpected auth message is no longer mistaken for a failed setup.
• 9d59f83: Assistants now connect to all of their MCP servers in parallel when a thread
  starts, so startup time no longer grows with the number of servers and one
  slow or unreachable server cannot stall the rest. Connection attempts are
  bounded by connect and handshake timeouts, so a hung server fails fast instead
  of blocking the assistant.
• 9a78d97: Default the device-agent command list in the generated observability plugin
  identity.sh to speakeasyd, the binary the daemon actually ships as. The
  previous default (device-agent,speakeasy-device-agent) never resolved on a
  standard install, so identity enrichment was skipped and hook events reached
  Gram anonymously (no user_email). The fix applies to the Claude Code, Cursor,
  and Codex plugin templates. Installs that still use a differently-named binary
  can override via GRAM_DEVICE_AGENT_COMMANDS.