fix: throw MaxTokensError when contentBlockStop carries truncated tool_use JSON

[serhiizghama]

Problem

When a Bedrock model hits its token limit mid-tool-call, the stream emits contentBlockStop before messageStop. At that point JSON.parse fails on the incomplete input, throwing a SyntaxError — before the subsequent messageStop { stopReason: 'maxTokens' } is processed. The SyntaxError propagates to the caller as a generic ModelError, hiding the MaxTokensError that callers need to recover from:

try {
  await agent.invoke(...)
} catch (e) {
  if (e instanceof MaxTokensError) { /* never reached */ }
  // Caller sees opaque SyntaxError wrapped in ModelError
}

Reported in issue strands-agents/harness-sdk#2451.

Solution

In streamAggregated, defer the SyntaxError instead of throwing it immediately:

1. Save the error in pendingParseError and let the loop continue to consume the remaining stream events.
2. After the loop, if pendingParseError is set:
   • If finalStopReason === 'maxTokens' → throw MaxTokensError (truncation by token limit).
   • Otherwise → re-throw the original SyntaxError (genuinely malformed JSON from the model).

Testing

• Updated the existing test that was asserting the old (incorrect) behavior: now it expects MaxTokensError when contentBlockStop arrives with truncated JSON and stopReason === 'maxTokens'.
• Added a new test for the non-maxTokens path: malformed JSON with stopReason === 'endTurn' still surfaces ModelError(cause: SyntaxError).

[strands-agent]

This repository has been merged into the strands-agents/harness-sdk monorepo and will be archived shortly. All new development happens there.

If this PR is still relevant, please recreate it against the monorepo. The code now lives under strands-ts/. Full commit history was preserved, so your base should be findable.

Apologies for the disruption, and thank you for contributing!