( model: string, betas?: string[], )
| 97 | } |
| 98 | |
| 99 | export function getContextWindowForModel( |
| 100 | model: string, |
| 101 | betas?: string[], |
| 102 | ): number { |
| 103 | // Allow override via environment variable (ant-only) |
| 104 | // This takes precedence over all other context window resolution, including 1M detection, |
| 105 | // so users can cap the effective context window for local decisions (auto-compact, etc.) |
| 106 | // while still using a 1M-capable endpoint. |
| 107 | if ( |
| 108 | process.env.USER_TYPE === 'ant' && |
| 109 | process.env.CLAUDE_CODE_MAX_CONTEXT_TOKENS |
| 110 | ) { |
| 111 | const override = parseInt(process.env.CLAUDE_CODE_MAX_CONTEXT_TOKENS, 10) |
| 112 | if (!isNaN(override) && override > 0) { |
| 113 | return override |
| 114 | } |
| 115 | } |
| 116 | |
| 117 | const explicitContextSuffix = parseModelContextSuffix(model) |
| 118 | if (explicitContextSuffix) { |
| 119 | return explicitContextSuffix.contextTokens |
| 120 | } |
| 121 | |
| 122 | const cap = getModelCapability(model) |
| 123 | if (cap?.max_input_tokens && cap.max_input_tokens >= 100_000) { |
| 124 | if ( |
| 125 | cap.max_input_tokens > MODEL_CONTEXT_WINDOW_DEFAULT && |
| 126 | is1mContextDisabled() |
| 127 | ) { |
| 128 | return MODEL_CONTEXT_WINDOW_DEFAULT |
| 129 | } |
| 130 | return cap.max_input_tokens |
| 131 | } |
| 132 | |
| 133 | if (betas?.includes(CONTEXT_1M_BETA_HEADER) && modelSupports1M(model)) { |
| 134 | return 1_000_000 |
| 135 | } |
| 136 | if (getSonnet1mExpTreatmentEnabled(model)) { |
| 137 | return 1_000_000 |
| 138 | } |
| 139 | if (process.env.USER_TYPE === 'ant') { |
| 140 | const antModel = resolveAntModel(model) |
| 141 | if (antModel?.contextWindow) { |
| 142 | return antModel.contextWindow |
| 143 | } |
| 144 | } |
| 145 | return MODEL_CONTEXT_WINDOW_DEFAULT |
| 146 | } |
| 147 | |
| 148 | export function getSonnet1mExpTreatmentEnabled(model: string): boolean { |
| 149 | if (is1mContextDisabled()) { |
no test coverage detected