Mastra

The retrieved chunks do not mention an “AI Gateway” specifically. Mastra’s top-level configuration sets a default per-task timeout of 300000 ms (5 minutes) via `defaultTimeoutMs`, which can be overridden per-tool or per-call [17]. The framework provides a default `RetryConfig` with configurable fields such as `maxRetries`, `retryDelayMs`, and `backoffMultiplier` [1][2][4][6], and automated retries are supported at the step or workflow level [6]. For model-specific retries, model fallbacks allow setting a retry count per fallback model [9]. The chunks do not specify default numeric values for `maxRetries`, nor do they document per-provider retry overrides in a gateway context.

Clarify in the configuration reference whether top-level timeout and retry settings apply to LLM provider calls, and document how (or if) users can override retry limits per provider or per model gateway route.

Where we looked: /reference/configuration, /reference/configuration, /reference/configuration, /blog/mastra-workflows-enhanced, /blog/model-fallback, /reference/configuration

The chunks confirm that Mastra agents have a default memory configuration and that the `lastMessages` option controls how much recent conversation history is included per call, but they do not specify the default numeric value for `lastMessages` or a default context-window cap before the agent compresses or drops earlier turns. For long-running conversations, the docs note that raw message history grows until it fills the context window, and Observational Memory can be enabled to compress old messages automatically.

State the default value of `lastMessages` (and any other default memory limits) explicitly in the Memory Configuration Guide or Agent Memory API docs.

Where we looked: /blog/agent-memory-guide, /blog/agent-memory-guide, /learn/how-memory-works, /docs/memory/overview

The retrieved chunks do not explicitly state whether Mastra guarantees exactly-once semantics for step processing in distributed or resumed executions. They document that Mastra supports automated retries at the step level when steps fail due to transient errors (for example, via `retryConfig` with `attempts` and `delay`), and that resumed workflows continue executing from where they last left off, but they do not specify whether steps may be executed more than once or whether step functions should be idempotent.

Document whether Mastra's workflow engine provides exactly-once, at-least-once, or best-effort execution semantics for steps, particularly in the context of retries and distributed execution. Include explicit guidance on idempotency requirements for step functions.

Where we looked: /blog/building-workflows, /ai-workflows, /blog/mastra-workflows-enhanced, /docs/workflows/error-handling, /docs/workflows/error-handling, /blog/introducing-temporal-workflows

The retrieved chunks do not specify a hard limit on the number of sequential steps a single Mastra workflow can execute, nor do they document a maximum runtime timeout enforced by the built-in orchestration engine. Mastra supports 'long-running workflows' with automatic state persistence and suspend/resume capabilities, and workflows can be deployed to external runners like Inngest that offer concurrency and rate-limiting controls, but no explicit step count or execution timeout ceiling is provided in the retrieved content.

Add a 'Limits and quotas' section to the workflow documentation that explicitly states whether there is a maximum number of sequential steps per workflow and any execution/runtime timeout enforced by the built-in orchestration engine.

Where we looked: /ai-workflows, /guides/deployment/inngest, /ai-agent-framework

The retrieved chunks show that Mastra supports PostgreSQL via `PostgresStore` using a `connectionString` (e.g., `postgresql://user:password@localhost:5432/myapp`), but they do not specify whether Mastra manages its own connection pool, what the default pool size is, or whether you must pass in an already-pooled client from your application.

Add documentation to the PostgreSQL storage reference explaining whether Mastra manages its own connection pool internally, the default pool size, and whether users should pass a pre-pooled client.

Where we looked: /blog/mastra-storage, /blog/mastra-storage, /blog/mastra-0.10