MCPRun a probe investigation

Run a probe investigation

Select the exact deployment, choose the lowest-cost probe, collect bounded evidence, and clean up.

Select the target deployment

Run list_services with the project and environment. Use the exact reported service ID, confirm it is online, and inspect its latest safety state and capability list.

Choose the lowest-cost probe

Question	Probe
What values exist at this line?	`snapshot`
Did this line execute?	`counter`
How does a numeric value change?	`metric`
What diagnostic message should be emitted?	`log`

Prefer counters and metrics on hot paths because agents pre-aggregate them. Start snapshots with a one-hit limit and a narrow watch list.

Confirm the deployed commit

All set-probe MCP tools require commit_hash. Confirm it against the service-reported commitSha, then inspect source at that exact revision before choosing an executable line.

git cat-file -e "<DEPLOYED_COMMIT>^{commit}"
git show "<DEPLOYED_COMMIT>:path/to/source-file"

Commit metadata is an honesty signal

Neither the MCP server nor the runtime cryptographically proves that loaded code matches the Git revision. A mismatch warning is a strong reason to stop and verify the deployment.

Collect bounded evidence

Create the probe with a short TTL and explicit hit limit. Use get_probe_data with wait_seconds up to 30 rather than repeatedly polling. Review redaction and truncation markers when interpreting snapshots.

Clean up

Remove the probe as soon as the question is answered. If no evidence arrives, inspect list_probes for line-not-found, suspended, expired, or hit-limit-reached state before creating another probe.