Agent Ops and Telemetry

Problem

The initial agent-context and provider-prompting work improved the runtime, but three follow-ons were still missing:

• admins had no first-class UI to author the durable agent_context business profile
• runtime prompt-caching and context-management behavior were invisible in cost views and tools
• prompt quality changes were not protected by focused regression fixtures

Solution

Workspace Context Authoring

Add a dedicated editor to Admin > Agents that reads and writes the typed agent_context tenant setting. This makes durable workspace business knowledge editable without modifying seeded agents or shipping a code change.

Runtime Telemetry

Emit ai.runtime.completed analytics events from the shared agent runtime with:

• model and source
• cache read/write token totals
• reasoning token totals
• whether prompt caching was configured
• whether context management was configured
• how many context edits were actually applied

Aggregate those signals into the admin cost surface and the getUsageStats tool.

Regression Coverage

Add fixture-backed prompt regression tests for:

• OCI workspace-business-context rendering
• the key operating-language expectations in seeded system-agent prompts

Outcome

• Workspace business context is now editable in the admin UI.
• Cost dashboards and tools can show cache reuse, cache writes, reasoning usage, and context-management activity.
• Prompt regressions now fail fast in tests instead of drifting silently.