Context window token count display broken for Claude models

[joe-hireable]

Issue Description

The context window token count display in Roo Code is not working correctly when using Claude models (via any provider). This affects all Claude models, including Claude 4 Sonnet.

Specific Problems

1. Incorrect Token Limit Display: Claude 4 Sonnet has a 1 million token context window capacity, but Roo Code only allows/displays 200k tokens maximum.

2. Context Condensing Functionality Breaking: The broken token count affects the context condensing functionality in two ways:

   • When condensing context is enabled automatically
   • When running condensing manually using the "condense context" button next to the token counter display

Impact

This issue prevents users from utilizing the full context window capabilities of Claude models, particularly Claude 4 Sonnet which has a significantly larger context window than what's being allowed. It also makes the condensing feature unreliable as it's operating based on incorrect token counts.

Environment

• Models affected: All Claude models via any provider
• Specific example: Claude 4 Sonnet (1M token capacity shows as 200k)

Expected Behavior

• Context window token count should accurately reflect the actual capacity of each Claude model
• Claude 4 Sonnet should show 1M token capacity instead of 200k
• Context condensing should work properly based on accurate token counts

Thank you for looking into this issue!

[dosubot]

Hi @joe-hireable! I'm Dosu and I’m helping the Roo-Code team.

Roo Code currently defaults the context window for Claude models to 200,000 tokens, even though Claude 4 Sonnet supports up to 1 million tokens. This default is used unless a higher value is provided by the model info endpoint from your provider, such as LiteLLM. As a result, the UI and context condensing features are limited by this 200k cap, which doesn't match Claude 4 Sonnet's actual capacity. The documentation and release notes do not mention explicit support for 1M token windows or a way to override this for Claude 4 Sonnet, so this appears to be a limitation in the current implementation. Context condensing relies on these token limits, so inaccurate values can break or limit its effectiveness. For now, check if your provider can supply the correct context window size via its model info endpoint, or monitor for future Roo Code updates that address this gap.
Sources:

• Anthropic provider docs
• LiteLLM provider docs
• Intelligent Context Condensing

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[roomote]

I am working on fixing the context window token count display for Claude models. I will investigate the codebase to identify where the 200k token limit is hardcoded and update it to properly reflect the actual context window capacities of Claude models, particularly Claude 4 Sonnet with its 1M token capacity.

[roomote]

Fixed in PR #5727. Updated Claude 4 models across all providers to show correct 1M token context window instead of 200k.

Changes made:

• Updated claude-sonnet-4-20250514 and claude-opus-4-20250514 context window from 200k to 1M tokens in Anthropic provider
• Updated corresponding models in Bedrock and Vertex providers
• Updated OpenRouter default model info fallback from 200k to 1M tokens
• Enhanced LiteLLM fetcher to detect Claude 4 models and use 1M token default
• Updated test expectations to reflect the new 1M token capacity

Testing:

• All critical checks passing (compile, integration tests, security analysis)
• Context window changes validated in unit tests
• UI token counter and context condensing will now show/use correct 1M token capacity

The fix addresses both the UI display issue and the underlying context condensing functionality that was incorrectly using the 200k limit.

[mrubens]

The 200k context window is correct.