# GitHub Issue: SessionCompaction.compactIfNeeded never triggers for @ai-sdk/openai-compatible providers #53935

Description

Bug

Automatic context compaction (compaction.auto: true) never triggers for providers using @ai-sdk/openai-compatible. The SessionCompaction.compactIfNeeded check is not called in the provider adapter's iteration loop, so context grows unbounded until the user manually runs /compact.

Environment

• OpenCode v1.18.0 (Homebrew)
• Provider: alibaba-token-plan (uses @ai-sdk/openai-compatible)
• Models: qwen3.8-max, deepseek-v4.1-flash, glm-5.2 (custom model limits configured via provider.alibaba-token-plan.models.*.limit)

Config

{
  "compaction": {
    "auto": true,
    "tail_turns": 15
  },
  "provider": {
    "alibaba-token-plan": {
      "models": {
        "qwen3.8-max": {
          "limit": { "context": 200000, "output": 16384 }
        }
      }
    }
  }
}

What happens

1. Session starts, context grows normally.
2. Context crosses the model's limit.context (200K tokens).
3. compactIfNeeded is never called — the @ai-sdk/openai-compatible provider adapter's loop does not include the compaction check.
4. Context keeps growing: 200K -> 400K -> 600K -> 800K+ tokens per message.
5. time_compacting stays NULL for all sessions.
6. Manual /compact works fine — the compaction mechanism itself is functional, just the automatic trigger is missing.

Evidence

• 405 sessions over 12 weeks, all with time_compacting = NULL.
• Only 1 session has compaction parts — created by manual /compact (part has auto: false).
• A test session crossed 200K at message 53 and grew to 491K without triggering compaction. After manual /compact, context dropped to 77K.
• The Anthropic SDK's BetaToolRunner has its own compactionControl check in its asyncIterator, but this code path is Anthropic-specific (this.client.beta.messages.create). The @ai-sdk/openai-compatible adapter uses a different loop.

Expected behaviour

SessionCompaction.compactIfNeeded should be called after each LLM response in all provider adapters, not just the Anthropic one. When compaction.auto: true and the model has limit.context configured, compaction should trigger when context approaches the limit.

Impact

This is causing significant token waste. Over 12 weeks:

• 14.6B tokens consumed across 50K LLM calls.
• 80.9% of tokens (11.8B) are "excess" from unbounded context growth.
• Average tokens per message: 292K (should be <200K with compaction).
• Output/input ratio: 1.7% (should be 5-15%).

Workaround

Manual /compact when a session crosses ~180K tokens. A monitoring script helps track this.

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response