orcs-oss/orcs-code
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
f6a4455ecfd06def107e3589c55208244d7c055e
orcs-code/src/utils/model/openaiContextWindows.ts
Vasanth T aeaa658f77 fix: prevent infinite auto-compact loop for unknown 3P models (#635) (#636) 
- Raise context window fallback from 8k to 128k for unknown OpenAI-compat models.
  The 8k fallback caused effective context (8k minus output reservation) to go
  negative, making auto-compact fire on every single message.
- Add safety floor in getEffectiveContextWindowSize(): effective context is
  always at least reservedTokensForSummary + 13k buffer, ensuring the
  auto-compact threshold stays positive.
- Add missing MiniMax model entries (M2.5, M2.5-highspeed, M2.1, M2.1-highspeed)
  all at 204,800 context / 131,072 max output per MiniMax docs.
- Add tests for MiniMax variants, 128k fallback, and autoCompact floor.

Fixes #635

Co-authored-by: root <root@vm7508.lumadock.com>
2026-04-13 02:03:02 +08:00

319 lines
13 KiB
TypeScript
Raw Blame History

/**
* openaiContextWindows.ts
* Context window sizes for OpenAI-compatible models used via the shim.
* Fixes: auto-compact and warnings using wrong 200k default for OpenAI models.
*
* When CLAUDE_CODE_USE_OPENAI=1, getContextWindowForModel() falls through to
* MODEL_CONTEXT_WINDOW_DEFAULT (200k). This causes the warning and blocking
* thresholds to be set at 200k even for models like gpt-4o (128k) or llama3 (8k),
* meaning users get no warning before hitting a hard API error.
*
* Prices in tokens as of April 2026 — update as needed.
*/
const OPENAI_CONTEXT_WINDOWS: Record<string, number> = {
// GitHub Copilot — values from https://api.githubcopilot.com/models (2026-04-09)
// Namespaced so they don't collide with bare model names from other providers.
'github:copilot': 128_000,
// Claude
'github:copilot:claude-sonnet-4': 216_000,
'github:copilot:claude-haiku-4': 200_000,
'github:copilot:claude-haiku-4.5': 144_000,
'github:copilot:claude-sonnet-4.5': 200_000,
'github:copilot:claude-sonnet-4.6': 200_000,
'github:copilot:claude-opus-4': 200_000,
'github:copilot:claude-opus-4.6': 200_000,
// GPT
'github:copilot:gpt-3.5-turbo': 16_384,
'github:copilot:gpt-4': 32_768,
'github:copilot:gpt-4-0125-preview': 128_000,
'github:copilot:gpt-4-o-preview': 128_000,
'github:copilot:gpt-4.1': 128_000,
'github:copilot:gpt-4o': 128_000,
'github:copilot:gpt-4o-2024-08-06': 128_000,
'github:copilot:gpt-4o-2024-11-20': 128_000,
'github:copilot:gpt-4o-mini': 128_000,
'github:copilot:gpt-5-mini': 264_000,
'github:copilot:gpt-5.1': 264_000,
'github:copilot:gpt-5.2': 400_000,
'github:copilot:gpt-5.2-codex': 400_000,
'github:copilot:gpt-5.3-codex': 400_000,
'github:copilot:gpt-5.4': 400_000,
'github:copilot:gpt-5.4-mini': 400_000,
// Gemini
'github:copilot:gemini-2.5-pro': 128_000,
'github:copilot:gemini-3-flash-preview': 128_000,
'github:copilot:gemini-3.1-pro-preview': 200_000,
// Grok
'github:copilot:grok-code-fast-1': 256_000,
// LiteLLM format — when OpenClaude talks to a LiteLLM proxy, Copilot models
// keep their "<provider>/<model>" naming convention (standard LiteLLM routing)
// instead of the "github:copilot:<model>" namespaced form used by /onboard-github.
// Entries below cover the aliases currently exposed by LiteLLM's github_copilot
// provider — this is a curated subset, not an exhaustive mirror of the
// namespaced entries above. Values are sourced from copilotModels.ts to stay
// consistent with the /onboard-github path.
'github_copilot/claude-sonnet-4.6': 200_000,
'github_copilot/claude-opus-4.6': 200_000,
'github_copilot/claude-haiku-4.5': 144_000,
'github_copilot/gpt-4.1': 128_000,
'github_copilot/gpt-4o': 128_000,
'github_copilot/gpt-5-mini': 264_000,
'github_copilot/gpt-5.4': 400_000,
'github_copilot/gpt-5.4-mini': 400_000,
'github_copilot/gemini-2.5-pro': 128_000,
'github_copilot/gemini-3-flash': 128_000,
'github_copilot/grok-code-fast-1': 256_000,
// NOTE: bare Claude model names (e.g. 'claude-sonnet-4') are intentionally
// omitted. Different OpenAI-compatible providers may impose different context
// limits for the same model name, so we cannot safely hardcode values here.
// OpenAI
'gpt-5.4': 1_050_000,
'gpt-5.4-mini': 400_000,
'gpt-5.4-nano': 400_000,
'gpt-4o': 128_000,
'gpt-4o-mini': 128_000,
'gpt-4.1': 1_047_576,
'gpt-4.1-mini': 1_047_576,
'gpt-4.1-nano': 1_047_576,
'gpt-4-turbo': 128_000,
'gpt-4': 8_192,
'o1': 200_000,
'o1-mini': 128_000,
'o1-preview': 128_000,
'o1-pro': 200_000,
'o3': 200_000,
'o3-mini': 200_000,
'o4-mini': 200_000,
// DeepSeek (V3: 128k context per official docs)
'deepseek-chat': 128_000,
'deepseek-reasoner': 128_000,
// Groq (fast inference)
'llama-3.3-70b-versatile': 128_000,
'llama-3.1-8b-instant': 128_000,
'mixtral-8x7b-32768': 32_768,
// Mistral
'mistral-large-latest': 256_000,
'mistral-small-latest': 256_000,
'devstral-latest': 256_000,
'ministral-3b-latest': 256_000,
// MiniMax (all M2.x variants share 204,800 context, 131,072 max output)
'MiniMax-M2.7': 204_800,
'MiniMax-M2.7-highspeed': 204_800,
'MiniMax-M2.5': 204_800,
'MiniMax-M2.5-highspeed': 204_800,
'MiniMax-M2.1': 204_800,
'MiniMax-M2.1-highspeed': 204_800,
'minimax-m2.7': 204_800,
'minimax-m2.7-highspeed': 204_800,
'minimax-m2.5': 204_800,
'minimax-m2.5-highspeed': 204_800,
'minimax-m2.1': 204_800,
'minimax-m2.1-highspeed': 204_800,
// Google (via OpenRouter)
'google/gemini-2.0-flash':1_048_576,
'google/gemini-2.5-pro': 1_048_576,
// Google (native via CLAUDE_CODE_USE_GEMINI)
'gemini-2.0-flash': 1_048_576,
'gemini-2.5-pro': 1_048_576,
'gemini-2.5-flash': 1_048_576,
// Ollama local models
// Llama 3.1+ models support 128k context natively (Meta official specs).
// Ollama defaults to num_ctx=8192 but users can configure higher values.
'llama3.3:70b': 128_000,
'llama3.1:8b': 128_000,
'llama3.2:3b': 128_000,
'qwen2.5-coder:32b': 32_768,
'qwen2.5-coder:7b': 32_768,
'deepseek-coder-v2:16b': 163_840,
'deepseek-r1:14b': 65_536,
'mistral:7b': 32_768,
'phi4:14b': 16_384,
'gemma2:27b': 8_192,
'codellama:13b': 16_384,
'llama3.2:1b': 128_000,
'qwen3:8b': 128_000,
'codestral': 32_768,
}
/**
* Max output (completion) tokens per model.
* This is separate from the context window (input limit).
* Fixes: 400 error "max_tokens is too large" when default 32k exceeds model limit.
*/
const OPENAI_MAX_OUTPUT_TOKENS: Record<string, number> = {
// GitHub Copilot — values from https://api.githubcopilot.com/models (2026-04-09)
'github:copilot': 16_384,
// Claude
'github:copilot:claude-sonnet-4': 16_000,
'github:copilot:claude-haiku-4': 64_000,
'github:copilot:claude-haiku-4.5': 32_768,
'github:copilot:claude-sonnet-4.5': 32_000,
'github:copilot:claude-sonnet-4.6': 32_000,
'github:copilot:claude-opus-4': 32_000,
'github:copilot:claude-opus-4.6': 32_000,
// GPT
'github:copilot:gpt-3.5-turbo': 4_096,
'github:copilot:gpt-4': 4_096,
'github:copilot:gpt-4-0125-preview': 4_096,
'github:copilot:gpt-4-o-preview': 4_096,
'github:copilot:gpt-4.1': 16_384,
'github:copilot:gpt-4o': 4_096,
'github:copilot:gpt-4o-2024-08-06': 16_384,
'github:copilot:gpt-4o-2024-11-20': 16_384,
'github:copilot:gpt-4o-mini': 4_096,
'github:copilot:gpt-5-mini': 64_000,
'github:copilot:gpt-5.1': 64_000,
'github:copilot:gpt-5.2': 128_000,
'github:copilot:gpt-5.2-codex': 128_000,
'github:copilot:gpt-5.3-codex': 128_000,
'github:copilot:gpt-5.4': 128_000,
'github:copilot:gpt-5.4-mini': 128_000,
// Gemini
'github:copilot:gemini-2.5-pro': 64_000,
'github:copilot:gemini-3-flash-preview': 64_000,
'github:copilot:gemini-3.1-pro-preview': 64_000,
// Grok
'github:copilot:grok-code-fast-1': 64_000,
// LiteLLM format — see note on context windows above.
'github_copilot/claude-sonnet-4.6': 32_000,
'github_copilot/claude-opus-4.6': 32_000,
'github_copilot/claude-haiku-4.5': 32_768,
'github_copilot/gpt-4.1': 16_384,
'github_copilot/gpt-4o': 4_096,
'github_copilot/gpt-5-mini': 64_000,
'github_copilot/gpt-5.4': 128_000,
'github_copilot/gpt-5.4-mini': 128_000,
'github_copilot/gemini-2.5-pro': 64_000,
'github_copilot/gemini-3-flash': 64_000,
'github_copilot/grok-code-fast-1': 64_000,
// NOTE: bare Claude model names omitted — see context windows comment above.
// OpenAI
'gpt-5.4': 128_000,
'gpt-5.4-mini': 128_000,
'gpt-5.4-nano': 128_000,
'gpt-4o': 16_384,
'gpt-4o-mini': 16_384,
'gpt-4.1': 32_768,
'gpt-4.1-mini': 32_768,
'gpt-4.1-nano': 32_768,
'gpt-4-turbo': 4_096,
'gpt-4': 4_096,
'o1': 100_000,
'o1-mini': 65_536,
'o1-preview': 32_768,
'o1-pro': 100_000,
'o3': 100_000,
'o3-mini': 100_000,
'o4-mini': 100_000,
// DeepSeek
'deepseek-chat': 8_192,
'deepseek-reasoner': 32_768,
// Groq
'llama-3.3-70b-versatile': 32_768,
'llama-3.1-8b-instant': 8_192,
'mixtral-8x7b-32768': 32_768,
// Mistral
'mistral-large-latest': 32_768,
'mistral-small-latest': 32_768,
// MiniMax (all M2.x variants share 131,072 max output)
'MiniMax-M2.7': 131_072,
'MiniMax-M2.7-highspeed': 131_072,
'MiniMax-M2.5': 131_072,
'MiniMax-M2.5-highspeed': 131_072,
'MiniMax-M2.1': 131_072,
'MiniMax-M2.1-highspeed': 131_072,
'minimax-m2.7': 131_072,
'minimax-m2.7-highspeed': 131_072,
'minimax-m2.5': 131_072,
'minimax-m2.5-highspeed': 131_072,
'minimax-m2.1': 131_072,
'minimax-m2.1-highspeed': 131_072,
// Google (via OpenRouter)
'google/gemini-2.0-flash': 8_192,
'google/gemini-2.5-pro': 65_536,
// Google (native via CLAUDE_CODE_USE_GEMINI)
'gemini-2.0-flash': 8_192,
'gemini-2.5-pro': 65_536,
'gemini-2.5-flash': 65_536,
// Ollama local models (conservative safe defaults)
'llama3.3:70b': 4_096,
'llama3.1:8b': 4_096,
'llama3.2:3b': 4_096,
'qwen2.5-coder:32b': 8_192,
'qwen2.5-coder:7b': 8_192,
'deepseek-coder-v2:16b': 8_192,
'deepseek-r1:14b': 8_192,
'mistral:7b': 4_096,
'phi4:14b': 4_096,
'gemma2:27b': 4_096,
'codellama:13b': 4_096,
'llama3.2:1b': 4_096,
'qwen3:8b': 8_192,
'codestral': 8_192,
}
function lookupByModel<T>(table: Record<string, T>, model: string): T | undefined {
// Try provider-qualified key first: "{OPENAI_MODEL}:{model}" so that
// e.g. "github:copilot:claude-haiku-4.5" can have different limits than
// a bare "claude-haiku-4.5" served by another provider.
const providerModel = process.env.OPENAI_MODEL?.trim()
if (providerModel && providerModel !== model) {
const qualified = `${providerModel}:${model}`
const qualifiedResult = lookupByKey(table, qualified)
if (qualifiedResult !== undefined) return qualifiedResult
}
return lookupByKey(table, model)
}
function lookupByKey<T>(table: Record<string, T>, model: string): T | undefined {
if (table[model] !== undefined) return table[model]
// Sort keys by length descending so the most specific prefix wins.
// Without this, 'gpt-4-turbo-preview' could match 'gpt-4' (8k) instead
// of 'gpt-4-turbo' (128k) depending on V8's key iteration order.
const sortedKeys = Object.keys(table).sort((a, b) => b.length - a.length)
for (const key of sortedKeys) {
if (model.startsWith(key)) return table[key]
}
return undefined
}
/**
* Look up the context window for an OpenAI-compatible model.
* Returns undefined if the model is not in the table.
*
* Falls back to prefix matching so dated variants like
* "gpt-4o-2024-11-20" resolve to the base "gpt-4o" entry.
*/
export function getOpenAIContextWindow(model: string): number | undefined {
return lookupByModel(OPENAI_CONTEXT_WINDOWS, model)
}
/**
* Look up the max output tokens for an OpenAI-compatible model.
* Returns undefined if the model is not in the table.
*/
export function getOpenAIMaxOutputTokens(model: string): number | undefined {
return lookupByModel(OPENAI_MAX_OUTPUT_TOKENS, model)
}
Reference in New Issue View Git Blame Copy Permalink