bug: auto-compact fires every turn for 3P models not in context window table (8k fallback causes infinite loop)

[Vasanthdev2004]

Summary

Auto-compact fires on every single message when using a 3P provider (OpenAI, Gemini, GitHub, Mistral) with a model not in the context window lookup table, making the CLI unusable.

Root Cause

In src/utils/context.ts, when a 3P provider model isn't found in openaiContextWindows.ts, getContextWindowForModel() falls back to 8,000 tokens:

// Unknown models get a conservative 8k default so auto-compact triggers
// before hitting a hard context_window_exceeded error.
return 8_000

The auto-compact threshold math in src/services/compact/autoCompact.ts then produces a negative effective context window:

contextWindow = 8,000
reservedTokensForSummary = min(maxOutput, 20,000) = 8,000–20,000
effectiveContext = 8,000 - 8,000 = 0 (or negative)
autoCompactThreshold = 0 - 13,000 = -13,000

Since any message (even 1 token) exceeds -13,000, auto-compact triggers every turn.

Affected Models

Any model not in src/utils/model/openaiContextWindows.ts that doesn't prefix-match an existing entry. Confirmed examples:

• MiniMax-M2.5, MiniMax-M2.5-highspeed, MiniMax-M2.1, MiniMax-M2.1-highspeed — all have 204,800 context per MiniMax docs, but only MiniMax-M2.7 is in the table, and prefix matching can't help (MiniMax-M2.5 doesn't start with MiniMax-M2.7)
• gemini-3-flash, gemini-3.1-pro (before PR feat(models): update Gemini model context windows and output limits #602 merges)
• Any custom/Ollama/LiteLLM model name that doesn't match the table

Fix Options

Quick fix

Add missing model entries to openaiContextWindows.ts (e.g., all MiniMax variants at 204,800 context / 131,072 max output).

Real fix

The 8,000 fallback in context.ts is too aggressive. It was intended to prevent context_window_exceeded errors, but it creates a worse problem — infinite compaction loops. Options:

1. Raise the fallback to 128,000 (reasonable default for modern 3P models)
2. Add a floor in getEffectiveContextWindowSize() so the result is always ≥ a usable minimum (e.g., Math.max(result, 50_000))
3. Guard in autoCompact — skip auto-compact when the effective context is ≤ 0 (treat as "unknown, don't auto-compact")

Option 3 is the safest — it prevents the infinite loop without guessing the right context window for unknown models. The error-level and warning-level thresholds can still fire to inform the user.

Reproduction

1. Set CLAUDE_CODE_USE_OPENAI=1 (or CLAUDE_CODE_USE_GEMINI=1, etc.)
2. Configure a model not in the table, e.g., MiniMax-M2.5
3. Start a conversation
4. Observe: auto-compact triggers after every message

Environment

• OpenClaude version: 0.1.8+ (current main)
• Provider: Any 3P provider with a model not in the context window table