Safe diagnostics

A useful diagnostic report is small, reproducible, and content-free. Capture identifiers that join operational evidence, not the underlying private data.

Include

• UTC timestamp and deployment origin.
• Product area, SDK name/version, or MCP client name/version.
• HTTP method and route template or MCP tool name.
• HTTP status and error.code.
• correlationId and, when available, receipt, run, authorization request, or deletion job ID.
• Account, Source, or dataset UUID only when the support channel is authorized for that account.
• Source state, expected schedule, and lastSuccessAt for freshness issues.
• Whether the call was a retry and whether it reused the original idempotency key; report that a key existed, not its full value.
• Minimal reproduction steps and the expected versus observed behavior.

Redact

• Bearer tokens, cookies, API keys, OAuth codes, enrollment secrets, and provider credentials.
• Authorization request URLs and secret environment values.
• SQL parameter values, returned rows, captured Desktop content, messages, documents, and screenshots containing Source data.
• Service-role keys, internal storage paths, signed URLs, database connection strings, and job credentials.
• Full idempotency keys when a short hash or statement of reuse is sufficient.

Minimal REST reproduction

Prefer a request with identifiers and content values replaced by placeholders. Keep the actual correlation ID in the report.

UTC: 2026-08-16T12:34:56Z
Client: @combined/context 0.1.0
Operation: querySql (POST /v1/sql)
Status/code: 422 query_rejected
Correlation: cor_01K2A000000000000000000101
Receipt: <receipt UUID if returned>
Retry: no
Observed: query rejected after catalogue refresh
Expected: bounded SELECT over one granted logical relation

For Combined Desktop, use the in-app diagnostics export, review its contents locally, and apply the same redaction rules before sharing. For REST or MCP, do not enable body-level logging merely to diagnose authentication or routing.

If the issue is repeatable, include the shortest reproduction and one recent correlation ID. Multiple blind retries create noise and can obscure the original failure.