[austinchennn]

Description:

Fixes #xxx

No linked issue. Found while reviewing vRT's context management: CascadeSession.history only grew, so long voice sessions would eventually exceed the text LLM's context window and memory grew with session length.

Behaviours Completed

• Requests to the llm send only a recent window of the history, bounded by message count (default 200) and an approximate token budget (default 32 000).
• session.history itself is capped (default twice maxMessages), so memory no longer grows with session length.
• Trimming keeps the conversation valid: a tool result is never kept without the assistant message that called it, the window never starts with a tool result, and it starts with a user message when one fits. Instructions are always sent.
• history.process lets apps replace the default trim; trimHistory and estimateTokens are exported to build on.

Files Changed

• packages/cascade/src/history.ts (new): HistoryLimits, estimateTokens, trimHistory.
• packages/cascade/src/index.ts: CascadeOptions.history; all history writes go through #remember, requests through #window.
• packages/cascade/test/cascade.test.ts: tests for the budget, tool call/result pairing, the stored cap, custom process, and unchanged behaviour for short sessions.

Brief Explanation

trimHistory marks every index that would orphan a tool result if the window started there, then walks back from the newest message accumulating the token estimate (about 4 characters per token) and picks the earliest valid start within budget, preferring a user message. If nothing fits, it falls back to the latest valid start. #remember applies the same function with the stored cap after each push.

Testing

npx vitest run packages: 551 tests pass across 34 files, including the new cascade tests. Typecheck in the worktree was not meaningful because it had no node_modules and resolved @vrealtime/* from an older checkout; relying on CI typecheck.

Unsure About

The token estimate is a character heuristic, not a real tokenizer, so the budget is approximate. Older history is dropped rather than summarized; summarization/compaction could be a follow-up.