Mastra Adds a Guardrail Before Tool Output Hits Context

Mastra's core package hit version 1.57.0 on Aug. 10 with a new hook, processToolResult, that inspects tool output before it reaches the model, per the release notes.

The hook runs after a tool executes and before its output enters message history. That's the exact point where prompt injection or a stray secret slips into an agent's context unfiltered. Developers can scan the payload, transform it, or abort the turn before the model ever sees it.

The same release adds transient: true, a flag for per-turn reminder context that skips permanent storage. An agent gets a nudge without it piling up in memory. Tool calls now get full JSON schemas attached to MODEL_GENERATION spans, useful for debugging what a tool actually returned. Sandbox execution latency dropped from about 400 milliseconds to roughly 16 milliseconds at the median, per the changelog.

The interception point is the part worth copying, not the feature list. Most agent code still treats tool output as trusted the moment it comes back; processToolResult flips that, it's untrusted by default now. Anything wiring an LLM agent to fetch web pages, read files, or hit third-party APIs needs a checkpoint before it reaches a prompt.