feat: make context length configurable instead of hardcoded

[Coder-Wangyankun]

Summary

This PR addresses issue #228 — making context length a configurable parameter instead of keeping it hardcoded.

Problem

Previously, context trimming parameters were hardcoded in two locations:

• session_actor.rs:272: BudgetAwareContextReducer::new(5, 0.8)
• client/mod.rs:224-225: keep_last_n_assistant_messages: Some(5), context_budget_threshold: Some(0.8)

Users had no way to override the model's context window size for budget calculations, nor could they tune the trimming threshold or how many messages to preserve.

Solution

Three new user-configurable parameters, exposed through every code path:

Parameter	Type	Default	Description
context_window	Option<u64>	model's built-in	Override context window size for budget calculations
context_budget_threshold	Option<f32>	0.8 (80%)	Fraction of window at which trimming triggers
keep_last_n_assistant_messages	Option<usize>	5	Recent assistant messages preserved during trimming

Usage

REST API:

{
  "message": {"role": "user", "content": "hello"},
  "overrides": {
    "context_window": 200000,
    "context_budget_threshold": 0.95,
    "keep_last_n_assistant_messages": 20
  }
}

TOML profile:

[profiles.myprofile]
context_window = 200000
context_budget_threshold = 0.95
keep_last_n_assistant_messages = 20

SDK:

let config = AgentClientConfig::new()
    .with_context_window(200_000)
    .with_context_budget_threshold(0.95)
    .with_keep_last_n_assistant_messages(20);

Changes

14 files changed, 248 insertions, 9 deletions.

Backward Compatibility

Fully backward compatible — all new fields are Option with the original hardcoded values as defaults when not set.