[feat] Configure the context window for custom OpenAI-compatible models

[cenab]

[feat] Configure the context window for custom OpenAI-compatible models

What problem are you trying to solve?

At 4585c8e455ac5bdeb76b618343ce748ef08134f4, a custom OpenAI-compatible
endpoint can be selected with api_provider, base_url, and api_key_helper,
but crates/cli/src/agent.rs constructs every SessionConfig with
context_window: 200_000. A custom endpoint with a 32,768-token window therefore
cannot declare its actual context budget through the same settings.

The existing provider works: an external Rust test runner used
crab_api::create_backend with the ordinary configuration and streamed requests
through the native OpenAI serializer and SSE parser for tsubasa-fast and
tsubasa-pro. Four controlled requests covered text and HTTP 401, with a
4,096-token output request. Each outgoing body passed the real Tsubasa strict
schema and input/output budget check before a fixture response. The existing
38 OpenAI tests also pass. These checks do not establish a full CLI session or
live model qualification; the fixed session window is a separate source finding.

Proposed solution

Add an optional context-window setting for a selected custom model and pass it
to SessionConfig. Preserve the current default when no override is present.
The same value should govern compaction/admission, including output headroom.
Tests could load the override through normal settings and verify that session
construction uses 32,768, while an unconfigured model preserves existing behavior.

Alternatives considered

Reducing --max-tokens bounds output but does not correct a 200,000-token
session context assumption. A new provider or custom transport would duplicate
the already-working OpenAI path.

Prior art

The existing custom-provider settings establish the endpoint-selection route.
Related settings issue #6 and provider proposal #128 were reviewed; this request
specifically concerns the session context limit.

Scope

LLM provider, Settings, Session, CLI.

Are you willing to contribute?

• I'm willing to submit a PR.
• I can help test.

Prepared with AI assistance on macOS arm64 with Rust 1.98.1. No credentials or
private prompts are included.