#593 release: per-message cost/usage metadata + ai SDK upgrade (#592)

Summary

• Adds messageMetadata callback to runAgentStep → emits message-metadata chunks per assistant finish-step with {modelId, lastStepUsage, totalMessageUsage, lastStepCost, totalMessageCost, stepFinishReasons}
• Wire format byte-compatible with sandbox.recoupable.com's existing UI consumption (verified via Chrome DevTools MCP capture against open-agents production)
• Upgrades ai@6.0.0-beta.122 → ^6.0.190 (+ @ai-sdk/anthropic/gateway/google/openai/mcp to latest stable) so api's LanguageModelUsage matches open-agents' flat shape natively
• SRP cleanups: addTokenCounts, hasGatewayShape, GatewayProviderMetadata, AgentStepFinishMetadata each in their own file
• Deletes normalizeUsage (-82 LOC) — no longer needed after SDK upgrade

Test plan

• 29 new unit tests across 7 helper files
• Full suite 3106/3106 pass
• E2E on preview verified: 2 message-metadata chunks emitted, totals accumulate correctly across steps, cost extraction from gateway providerMetadata works ($0.006191 + $0.010118 = $0.016309)
• Cross-validated wire format against open-agents production — byte-for-byte compatible

🤖 Generated with Claude Code

---

Summary by cubic

Adds per-message cost and token-usage metadata to assistant turns and upgrades the ai SDK so usage data matches the open-agents wire format used by sandbox.recoupable.com.

• New Features

  • Emit message-metadata chunks from runAgentStep on each finish-step via a messageMetadata callback with { modelId, lastStepUsage, totalMessageUsage, lastStepCost, totalMessageCost, stepFinishReasons }.
  • Accumulates totals across steps and parses step cost from Vercel AI Gateway providerMetadata.
  • Extracted focused helpers (addLanguageModelUsage, addTokenCounts, extractGatewayCost, hasGatewayShape) and types; deleted normalizeUsage.
  • Updated runAgentStep and setupChatRequest to await convertToModelMessages.

• Dependencies

  • Upgraded ai to 6.0.190 and @ai-sdk/anthropic, @ai-sdk/gateway, @ai-sdk/google, @ai-sdk/openai, @ai-sdk/mcp to latest stable, adopting the flat LanguageModelUsage shape for byte-compatible UI payloads.