[mason] tracing: UC-only and opt-in via mason tracing setup

[jamesbxwu]

UX Change

• Off until configured. mason tracing setup --trace-location <catalog.schema> records a UC schema in agent.toml. Tracing is disabled until then.
• Never blocks. mason deploy and mason dev displays a message to remind users to set up tracing using mason tracing setup

Key changes

• dev.py / deploy.py: share provision_trace_experiment (returns None when unconfigured; no gate).
• pyproject.toml: add opentelemetry-exporter-otlp-proto-http to the runtime/runtime-openai extras - UC trace export uses HTTP OTLP.
• runtime/tracing.py: reverted to the shipped gate (has_destination and has_experiment); we always wire both, so no runtime change is required.
• Removed the managed [dev] experiment machinery (dev_experiment_name, MasonClient.ensure_experiment, the [dev] fallback in tracing list).

Manual testing

Full CUJ against a real workspace, from a fresh venv with the CLI installed from this branch, for both LangGraph and OpenAI. Sequencing: no tracing first (confirm nothing is logged), then mason tracing setup, then confirm traces land - in both mason dev and a real mason deploy.

Stage	LangGraph	OpenAI
mason init	scaffolds, tracing off	scaffolds, tracing off
mason dev, no tracing	invoke OK, 0 traces, off-hint, no MLFLOW_* env	invoke OK, 0 traces, off-hint, no MLFLOW_* env
mason tracing setup --trace-location james_wu.mason	recorded in agent.toml	recorded in agent.toml
mason dev, tracing on	invoke OK, trace in james_wu.mason	invoke OK, trace in james_wu.mason
mason deploy, tracing on	—	real rollout, UC experiment created, env wired
live deployed app	—	SP granted, app invoked, its trace in james_wu.mason

Warning message when tracing is off

mason dev with no mason tracing setup run - no MLFLOW_* env is wired, nothing is exported, and the startup panel nudges (never blocks):

✓ Starting agent
  Chat UI   http://localhost:8020
  Next steps   ($ = run in your terminal)
  • Open http://localhost:8020 to chat with your agent
  $ mason tools add mcp <service>      Give the agent a tool
  $ mason dev -m <store> -s <store>    Attach a memory / session store
  $ mason deploy cuj-langgraph         Deploy it to Databricks
  • Tracing is off. Run `mason tracing setup --trace-location <catalog.schema>` to enable it.

mason deploy on an unconfigured project shows the same one-line hint as a next step instead of gating, so the deploy still succeeds.

After mason tracing setup (tracing on)

mason dev now names the destination in its panel:

✓ Starting agent
  Chat UI   http://localhost:8020
  Traces    james_wu.mason
  ...

mason deploy wires it and creates the UC-linked experiment:

✓ Deployed agent 'mason-cuj-openai'
  URL              https://mason-cuj-openai-….staging.aws.databricksapps.com
  Traces           /Users/james.wu@databricks.com/mason-traces/cuj-openai (james_wu.mason)
  Next steps
  • Grant the app's service principal USE CATALOG + USE SCHEMA + MODIFY/SELECT on james_wu.mason so it can write traces.

Traces actually land

mason tracing list after dev invokes (both frameworks) and a live invoke of the deployed OpenAI app - all three in the configured UC schema:

Agent Traces · james_wu.mason
TRACE ID                                  STATUS   LATENCY (MS)   CREATED
trace:/james_wu.mason/cbeca5bccee60cf0…   ● Ok     1833           just now         ← deployed OpenAI app
trace:/james_wu.mason/b792fba9600c702c…   ● Ok     1695           12 minutes ago   ← dev (OpenAI)
trace:/james_wu.mason/31b138b84d6d2436…   ● Ok     1353           20 minutes ago   ← dev (LangGraph)
3 items

Notes

• Reading traces back (mason tracing list) uses the workspace default SQL warehouse; pass --warehouse-id (or mason tracing setup --warehouse-id) where no default warehouse resolves in the target workspace.
• The HTTP OTLP exporter dep ships in the extras here; to validate before this merges/publishes it was injected into the agent venvs, which is what the merged extras provide automatically.