What happened?

Assistant turns on amazon-bedrock with adaptive-thinking Claude models (Opus 4.6, Opus 4.7, Sonnet 4.6) are silently truncated at exactly 4096 output tokens with stopReason: "length", regardless of the model's actual maxTokens (128000 in pi-ai's registry) or any user-configured modelOverrides.maxTokens.

Calling the same model directly via the AWS CLI with explicit maxTokens: 6000 returns cleanly with outputTokens: 6000. AWS will produce well past 4096 when the field is set. Cap is on Pi's side.

Steps to reproduce

1. Run pi against a Bedrock adaptive-thinking model:

pi --provider amazon-bedrock --model us.anthropic.claude-opus-4-7

2. Send a single prompt that elicits >4096 output tokens of plain text, no tool use
3. Inspect the session JSONL — assistant turn ends with stopReason: "length", usage.output: 4096.

Expected behavior

Adaptive anthropic models on Bedrock should not terminate at 4096 tokens

Version

0.75.4