feat(sdk): add config driven sink selection

[namrataghadi-galileo]

Summary

Added config-driven observability sink selection across the SDK and server.

Introduced shared telemetry sink-selection primitives so observability can use the default backend, registered SDK sinks, or named custom sink factories.

Expanded tests and README examples to cover the new sink-selection behavior and lifecycle handling.

Fixed the SDK mypy issue in shutdown handling by wrapping sink lifecycle awaitables in a concrete coroutine before passing them to asyncio.run().

Scope

User-facing/API changes:

• SDK agent_control.init() now accepts observability_sink_name and observability_sink_config
• SDK exports new sink registration helpers for external/custom observability sinks
• Server observability config now supports sink selection and sink config via environment/settings
• README includes a new example for registering and selecting an external sink

Internal changes:

• Added shared sink-selection models/registry in telemetry/
• Refactored SDK observability to resolve active sinks from config rather than always using the built-in batcher
• Refactored server startup to resolve an observability backend from sink selection
• Added server/SDK/telemetry tests for sink selection, lifecycle, and event delivery paths
• Adjusted SDK shutdown typing to satisfy mypy

Out of scope:

• No changes to core control evaluation logic
• No new built-in third-party sink implementation is included in this branch
• No UI work or rollout automation is included

Risk and Rollout

Risk level: medium

Rollback plan: Revert this branch to restore the previous default-only observability path. As a partial mitigation, deployments can keep observability_sink_name=default to stay on the legacy behavior even with this code merged.

Testing

• Added or updated automated tests
• Ran make check — could not run in this environment because uv was not available in the shell
• Manually verified behavior

Checklist

• Linked issue/spec (if applicable)
• Updated docs/examples for user-facing changes
• Included any required follow-up tasks

Follow-up tasks:

• Run make check in a full local/CI environment with uv available
• Confirm any downstream custom sink integrations against the new SDK/server registration APIs

[codecov]

Codecov Report

❌ Patch coverage is 93.00412% with 17 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...etry/src/agent_control_telemetry/sink_selection.py	86.48%	5 Missing ⚠️
sdks/python/src/agent_control/observability.py	96.66%	4 Missing ⚠️
server/src/agent_control_server/main.py	89.18%	4 Missing ⚠️
...er/src/agent_control_server/observability/sinks.py	88.00%	3 Missing ⚠️
sdks/python/src/agent_control/settings.py	83.33%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[lan17]

I like the direction of moving sink selection into a shared model, but I don't think this middle layer is ready to merge yet. The biggest issue is that the config surface is now shared between SDK and server, but the semantics still diverge. I also don't think the SDK should take ownership of caller-registered sinks on shutdown, and returning the best fanout result makes partial delivery failures too easy to miss.

[lan17]

Overall I think this is in good shape. The split between shared selection shape and runtime-specific semantics is the right direction, and moving the server onto its own sink env names fixes the biggest design problem from the earlier version. I left one follow-up comment on named-sink cleanup when switching back to default, but I don't think it should block this landing. Assuming this gets rebased cleanly on top of 175, I'm good with it.