Claude Code Context Window Handling

Current implementation note, reviewed on 2026-09-05. See LLM compatibility for usage reporting and general proxy limits.

Agent Maestro bridges three different context-window models when Claude Code uses VS Code language models through the Anthropic-compatible proxy:

• GitHub Copilot advertises each model's prompt budget through LanguageModelChat.maxInputTokens.
• VS Code/Copilot accepts provider-specific configuration.contextSize to choose the prompt/input budget used for the actual request.
• Claude Code has its own model and compaction logic, including special handling for 1M-context models.

The goal is to make all three use the same source of truth: the selected VS Code model's advertised maxInputTokens.

Source Of Truth

Use maxInputTokens from the selected VS Code language model as the prompt/input budget. Do not infer the context window from model names except where Claude Code requires a marker for its own client-side behavior.

The prompt budget is separate from the response budget:

• configuration.contextSize is the prompt/input budget. In Copilot debug output, this appears as maxPromptTokens.
• Anthropic/OpenAI/Gemini response limits remain controlled by request options such as max_tokens, max_output_tokens, or equivalent fields.

Configurator Behavior

Agent Maestro: Configure Claude Code Settings writes Claude Code environment variables using the selected model metadata.

For models whose maxInputTokens looks like a 1M tier, Agent Maestro appends [1m] to ANTHROPIC_MODEL so Claude Code enables its extended-context path. The current band is:

• greater than 800_000
• less than 1_500_000

This intentionally targets the current 1M tier. Future larger tiers should get a separate rule instead of widening this band.

The configurator also writes:

• CLAUDE_CODE_AUTO_COMPACT_WINDOW=<selected maxInputTokens> when the selected model has a positive maxInputTokens value.
• CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=85 only when the user has not already set a value.

Users must rerun Agent Maestro: Configure Claude Code Settings for these env vars to be written into an existing Claude Code settings file.

Fallback token estimates and /v1/messages/count_tokens responses report VS Code's raw token count unscaled. Earlier compaction is driven by the two env vars above — lower CLAUDE_AUTOCOMPACT_PCT_OVERRIDE for more headroom — rather than by inflating the reported counts.

Proxy Behavior

Before calling LanguageModelChat.sendRequest, Agent Maestro applies withCopilotConfiguration to generation requests. The helper returns updated options for Copilot-provided models; other providers are left unchanged. For a model advertising maxInputTokens: 1000000, this adds the numeric budget:

{
  "configuration": {
    "contextSize": 1000000
  }
}

This avoids VS Code/Copilot falling back to a smaller default context size for a model that advertises a larger prompt budget.

Model Routing

Claude Code uses the [1m] marker and context-1m beta header as client-side signals for its 1M context path. Agent Maestro does not translate those signals into synthetic Copilot model IDs; it passes the requested model ID into normal exact/fuzzy matching against Copilot's advertised model list.

This keeps routing aligned with Copilot's current model catalogue, where the real model IDs and their maxInputTokens values are the source of truth. For example, if Copilot advertises claude-opus-4.8 as a 1M model, Claude Code requests should continue to route to claude-opus-4.8 instead of a fabricated internal variant.

Known Limits

• The [1m] threshold is a 1M-tier heuristic based on currently advertised Copilot model windows.
• Claude Code's status-line percentage can differ from auto-compaction math when CLAUDE_CODE_AUTO_COMPACT_WINDOW is set; this is expected Claude Code behavior.
• If Copilot changes the meaning of maxInputTokens or configuration.contextSize, update this design and the tests together.