Durable Insight

The durable-insight MCP server gives you READ-ONLY access to the execution history that AWS Lambda durable functions emit to a configured destination. Use it to investigate failures, timing, and step-level errors. The destination is chosen by the operator through DURABLE_INSIGHT_* environment variables; you do not pick it and must not assume which one is in play.

The one rule that makes this skill correct: ask the server for the schema

This skill deliberately contains ZERO destination-specific schema facts — no field names, no column casing, no dialect syntax. That is not an omission; it is the design:

• Drift becomes impossible, not merely detectable. Any schema prose copied here would silently rot the moment the server's schema changed. There is nothing to keep in sync because there is nothing duplicated.
• Token economy. One destination's guidance alone is over ten thousand characters. Inlining every destination's schema would load all of it on every invocation, whether or not you are on that destination.
• Correctness. Only the running server knows the configured destination. A static file would have to hedge across every backend and would be wrong for six of them at any given moment.

So: call describe_schema before you write a query. It returns the record fields and query idioms for whatever destination is actually configured, authoritatively. Writing a query without it is guessing.

The five tools, and when to reach for each

• test_destination — Run first. Confirms the destination is reachable and its configuration is complete. If required environment variables are unset it names them and returns without any AWS call. Stop and report if it fails.
• describe_schema — Call before any query. Returns the configured destination's record schema, query engine/dialect, the table or log group in play, and the row cap. Makes no AWS call, so it is safe even before setup is complete.
• list_executions — The common case: list executions filtered by any of status, functionName, since, and until, with no SQL required. Prefer this over a hand-written query — it cannot be got wrong and costs fewer tokens.
• get_execution — Fetch a single execution record by its execution ARN. A record that does not exist is a success with found=false, not an error.
• query — Escape hatch for questions the structured tools cannot express. Only after describe_schema, so field names and syntax match the destination.

Guarantees and limits you can rely on

• Read-only. For the SQL destinations, any statement that is not a SELECT/WITH is refused before any AWS call is made; the CloudWatch Logs query language has no write forms at all. You cannot mutate data through this server.
• Row cap. Every result is capped at a fixed maximum row count (MAX_ROWS, currently 1000). A truncated flag tells you when the cap was hit — narrow your filters rather than assuming you saw everything.
• Log destinations need a lookback window. CloudWatch Logs queries have no "all time"; pass a lookback window (hours) or accept the 24-hour default. This is ignored by the SQL destinations, which are not time-windowed.

Suggested order for a failure investigation

1. test_destination — confirm reachability and configuration.
2. describe_schema — learn this destination's fields and idioms.
3. list_executions — narrow to the executions of interest (start with FAILED).
4. get_execution — drill into a specific record for step-level detail.
5. query — only when the structured tools cannot express the question.