[Bug] Context usage indicator always shows <1% with BYOK models via LiteLLM proxy

[Rubyj]

Description

When using BYOK custom models routed through a LiteLLM proxy (v1.81.15), the context usage indicator in the CLI status bar always shows <1% regardless of actual context consumption. The /cost command also shows Input: 0 tokens (related to #157).

Environment

• Droid CLI: v0.81.0
• macOS (arm64)
• BYOK config: provider: "anthropic" pointing to a LiteLLM proxy backed by Azure/Bedrock
• LiteLLM version: 1.81.15
• Models affected: All custom BYOK models (both Anthropic and OpenAI provider types)
• showTokenUsageIndicator: true in settings.json

Steps to Reproduce

1. Configure a BYOK model with provider: "anthropic" pointing to a LiteLLM proxy
2. Start a droid session using the custom model
3. Run several prompts that consume meaningful context
4. Observe the context usage indicator -- always shows <1%
5. Run /cost -- input tokens show as 0

Investigation: The proxy IS returning correct usage data

I tested the LiteLLM proxy directly with curl and confirmed all token usage fields are correctly populated in every response path:

Non-streaming:

"usage": {"input_tokens": 16, "cache_creation_input_tokens": 0,
          "cache_read_input_tokens": 0, "output_tokens": 8, "total_tokens": 24}

Streaming (message_start event):

"usage": {"input_tokens": 16, "cache_creation_input_tokens": 0,
          "cache_read_input_tokens": 0, "output_tokens": 1}

Streaming (message_delta event):

"usage": {"output_tokens": 8}

Streaming (message_stop event):

"usage": {"input_tokens": 16, "output_tokens": 8}

Streaming with extended thinking: Also returns correct input_tokens and output_tokens.

All Anthropic SSE event types (message_start, message_delta, message_stop) contain properly structured usage data. The proxy is fully compliant with the Anthropic streaming format.

Likely Root Causes

1. Context window size unknown for custom models: The proxy's /v1/models endpoint returns minimal metadata (no max_tokens or context_window field). Droid likely computes context % as input_tokens / context_window_size, and without knowing the context window for a custom model, it may default to 0 or an extremely large value, yielding <1%.

2. Token accumulation not working for BYOK streaming: Even though usage data is present in streaming events, Droid may not be parsing/accumulating input_tokens from the message_start event for BYOK provider models (per [Bug] /cost only shows partial token counts when using BYOK models #157).

Expected Behavior

• Context usage indicator should reflect actual token consumption as a percentage of the model's context window
• For BYOK models where context window is unknown, Droid could either:
  • Allow users to specify context_window in the BYOK config
  • Use a sensible default based on the model name (e.g., claude-* → 200K)
  • Fall back to showing raw token count instead of a percentage

Related

• [Bug] /cost only shows partial token counts when using BYOK models #157 (/cost shows partial token counts for BYOK models)