fix: forward resolved context limit to Ollama

[zerone0x]

Description

This PR fixes the missing runtime half of #485 for Ollama-backed models.

Nanocoder already let users configure context limits via:

• /context-max / --context-max
• provider-level contextWindow
• per-model contextWindows
• subagent contextWindow

But those limits were only used inside Nanocoder for context accounting and auto-compact decisions. They were not being forwarded to Ollama itself, so the actual model runtime could still use a different context size.

This change now forwards the resolved Nanocoder context limit to Ollama as options.num_ctx on each request, so the configured limit affects the real runtime behavior too.

It also updates the docs to make that behavior explicit.

Type of Change

• New feature
• Bug fix
• Breaking change
• Documentation update

Testing

Automated Tests

• New features include passing tests in .spec.ts/tsx files
• All existing tests pass (pnpm test:all completes successfully)
• Tests cover both success and error scenarios

Ran:

• npm exec --yes pnpm -- test:ava source/ai-sdk-client/chat/provider-options.spec.ts
• npm exec --yes pnpm -- test:types
• npm exec --yes pnpm -- exec biome check source/ai-sdk-client/chat/provider-options.ts source/ai-sdk-client/chat/provider-options.spec.ts source/ai-sdk-client/chat/chat-handler.ts docs/configuration/providers/ollama.md docs/features/subagents.md

Manual Testing

• Tested with Ollama
• Tested with OpenRouter
• Tested with OpenAI-compatible API
• Tested MCP integration (if applicable)

Manual testing was not run in this environment.

Checklist

• Code follows project style guidelines
• Self-review completed
• Documentation updated (if needed)
• No breaking changes (or clearly documented)
• Appropriate logging added using structured logging

Closes #485

[akramcodez]

Thanks for the contribution - I think the overall direction is correct, and forwarding the resolved context limit to Ollama is absolutely the right approach.

However, after validating the issue and tracing the context resolution flow, I don't think we can confidently close this issue yet.

The main blocker is the subagent precedence issue. The original feature request explicitly calls out subagents using different context limits, but the current resolution order means a subagent override is ignored whenever a per-model contextWindows entry exists. I was able to reproduce and verify this behavior locally.

Additionally, I wasn't able to verify whether Ollama's OpenAI-compatible endpoint actually honors the injected options.num_ctx field at runtime. Since the effectiveness of this fix depends on that behavior, I'd like to see either manual validation or automated coverage demonstrating that the runtime context limit changes when using the /v1/chat/completions endpoint.

Once the subagent override issue is addressed and the Ollama runtime behavior is verified, I'd be happy to take another look.