Atmosphere 4.0.54

Added

• Rich human-in-the-loop approval payloads. A reviewer can now resolve a tool
  approval with more than approve/deny: approve-with-edited-arguments (the
  tool runs with the reviewer's arguments) or respond (the reviewer answers on
  the tool's behalf — structured JSON or free-form text — and the tool does not
  run). Wire protocol: /__approval/<id>/approve {"arguments":{…}} and
  /__approval/<id>/respond {…}. Fail-safe: a malformed edited-args payload denies.
  Session-scoped in-memory (not crash-durable). The legacy boolean resolution path
  is unchanged.
• Eval flywheel. JournalDatasetPromoter turns a recorded CoordinationJournal
  interaction into an EvalCase dataset row (trace→dataset), and SampledLiveScorer
  grades a configurable fraction of live turns into EvalRun verdicts (online
  scoring). Both are wired into EvalController with admin REST routes
  (/api/admin/evals/dataset, /dataset/promote, /score).
• OAuth on-behalf-of credential vault. OAuthOnBehalfOfCredentialStore is a
  concrete CredentialStore that performs an RFC 8693 token exchange — swapping a
  user's stored subject token for a short-lived access token scoped to a downstream
  tool, so an agent calls external APIs as the user. Fail-closed (no token →
  no fallback credential), token-cached until expiry. Opt in with
  atmosphere.ai.identity.oauth-obo.enabled=true.
• Realtime voice bridge. VoiceBridge + the RealtimeVoiceProvider SPI bridge
  client audio frames over the existing WebSocket broadcaster to a speech-to-speech
  provider, fanning synthesized audio (Content.Audio) and transcripts back to the
  client. A dependency-free LoopbackVoiceProvider ships as the runnable reference
  (echoes audio); OpenAI Realtime / Gemini Live providers implement the same SPI.
• Content-safety moderation guardrail. ModerationGuardrail blocks turns
  whose request and/or response is flagged for hate / harassment / self-harm /
  sexual / violence / illicit content, on the existing fail-closed guardrail
  pipeline. Pluggable detector: zero-dep RuleBasedModerationDetector (default)
  or cross-runtime LlmModerationDetector. Fail-closed by default (a detector
  outage blocks the turn; .failOpen() is the explicit opt-out). Opt in with
  atmosphere.ai.guardrails.moderation.enabled=true
  (...detector=llm for the model tier).
• Self-healing structured output. @AiEndpoint(structuredOutputRetries = N)
  (or the ai.structured.retry request-metadata key on the AiPipeline path)
  re-prompts the model with the schema-validation error as feedback when a typed
  response fails to parse, up to N extra attempts, then fails closed. Works
  identically on the websocket and channel-bridge paths.
• OpenAPI → governed tools. OpenApiToolImporter turns an OpenAPI 3.x spec
  (JSON or YAML, with local $ref resolution) into ToolDefinitions whose
  executor performs the HTTP call. The imported operations ride the same
  policy-admission and plan-and-verify path as hand-written @AiTool methods;
  approvalForWrites routes mutating verbs through the HITL gate.
• MCP client depth. McpClientOptions adds per-server tool filtering and
  display-only renaming (the executor still calls the server's original tool
  name), plus elicitation/sampling callback handlers advertised during
  initialize. McpServerRegistry aggregates several servers into one
  collision-free tool list (first-wins) and owns their lifecycle.