fix: use non-interactive flag for Nova Sonic history and system promp…

[prettyprettyprettygood]

…t events

Description

When a BidiAgent session is renewed (stopped and recreated with conversation history), Nova Sonic repeats or re-summarizes prior assistant responses. This happens because conversation history messages are sent with interactive: True, causing Nova Sonic to treat them as new live inputs requiring a response rather than prior context.

Additionally, there are no size guards on replayed history, so unbounded message accumulation can cause failures or degraded performance on session renewal.

Public API Changes

_get_text_content_start_event now accepts an optional interactive parameter (defaults to True, fully backward compatible):

# Before
def _get_text_content_start_event(self, content_name: str, role: str = "USER") -> str:

# After
def _get_text_content_start_event(self, content_name: str, role: str = "USER", interactive: bool = True) -> str:

No changes to any public-facing API. All changes are internal to BidiNovaSonicModel.

Changes

• History messages are now sent with interactive: False so Nova Sonic treats them as prior context, not new inputs
• System prompt events are also sent with interactive: False to match Nova Sonic best practices
• Individual history messages are truncated to 50KB, total history capped at 200KB
• When the size limit is reached, oldest messages are dropped first to prioritize recent context
• History replay always starts with a user-role message (leading assistant messages are dropped)

Related Issues

Documentation PR

Type of Change

Bug fix

Testing

• Tested with a local BidiAgent harness that refreshes sessions every 30 seconds

• Verified via debug logging that history messages are sent with interactive: false

• Confirmed Nova Sonic no longer repeats prior responses after session renewal

• Validated that oldest messages are correctly dropped when history exceeds size limits

• I ran hatch run prepare

Checklist

• I have read the CONTRIBUTING document
• I have added any necessary tests that prove my fix is effective or my feature works
• I have updated the documentation accordingly
• I have added an appropriate example to the documentation to outline the feature, or no new docs are needed
• My changes generate no new warnings
• Any dependent changes have been merged and published

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[github-actions]

Assessment: Request Changes

The core fix (setting interactive=False for history and system prompt events) is well-motivated and the implementation logic is sound. The two-pass approach for size-limited history with newest-first priority is a good design choice.

Review Categories

• Testing: This is the primary blocker — the PR introduces significant new behavior (interactive flag, truncation, size caps, leading-message stripping) with zero new or updated tests. The existing tests don't assert on the interactive field, so the central bug fix has no automated verification.
• Maintainability: Size limit constants are inlined as magic numbers; extracting them as module/class constants would improve discoverability and testability.
• Minor efficiency: A small redundant UTF-8 re-encoding in the truncation path.

The approach and reasoning are solid — just needs test coverage to land safely.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!