feat(compaction): configurable threshold and model (global + per-model)

[tobiastornros]

Ever noticed how Opus gets kinda dumb around 50-60% context? Yeah, me too. Built this to escape the dumbzone.

Adds compaction settings at both global and per-model levels:

• threshold (1-100) - trigger compaction earlier, before models go stupid
• model - use a cheaper model for compaction (why waste Opus on summaries?)

Global config:

{
  "compaction": {
    "threshold": 80,
    "model": "anthropic/claude-haiku-5-20260205"
  }
}

Per-model config:

{
  "provider": {
    "anthropic": {
      "models": {
        "claude-opus-4-20250514": {
          "compaction_threshold": 50,
          "compaction_model": "zhipu/glm-4-plus"
        }
      }
    }
  }
}

Per-model overrides global. Default is 100 (current behavior).

Closes #11930
Relates to #11314, #8140, #10017, #8629

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Related PRs Found

Potential duplicates/related work:

1. feat(session): add custom compaction thresholds along with prevention of continuous compaction #10123 - feat(session): add custom compaction thresholds along with prevention of continuous compaction

   • Directly related: adds custom compaction thresholds (similar feature)

2. feat(opencode): add context overflow prevention with configurable thresholds #8810 - feat(opencode): add context overflow prevention with configurable thresholds

   • Overlaps on configurable thresholds for context management

3. feat(opencode): trigger compaction earlier and add multi-file read #9656 - feat(opencode): trigger compaction earlier and add multi-file read

   • Related: addresses triggering compaction earlier (part of feat(compaction): configurable threshold and model (global + per-model) #11931's motivation)

4. fix(opencode): accumulate step tokens for compaction #10215 - fix(opencode): accumulate step tokens for compaction

   • Related: compaction token handling

5. fix(opencode): prevent context overflow during compaction #11453 - fix(opencode): prevent context overflow during compaction

   • Related: context management during compaction

These PRs suggest there's been ongoing work on compaction and context threshold features. PR #10123 appears most similar, implementing custom compaction thresholds. You may want to check if #11931 supersedes or builds upon any of these, particularly #10123.

[tobiastornros]

Future idea: we could add compaction_max_tokens on model level too, an absolute token cap regardless of percentage. Useful for cost control or latency-sensitive setups. Kept it simple for now, can extend later.