Issue for this PR

Closes anomalyco#37551
Closes anomalyco#41358
Related: anomalyco#43703, anomalyco#37629

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Compaction sizing is currently absolute-token based (preserve_recent_tokens) and the auto-trigger uses a fixed reserved headroom. Two problems follow from that:

1. When a session is smaller than preserve_recent_tokens, select() walks every turn without exceeding the budget, ends with keep.start === 0, and returns "head = everything, no tail" — so the whole conversation gets summarized and nothing is kept verbatim.
2. A fixed reserved buffer makes the effective trigger point scale with the model's context window, so the same config compacts a small-context model very differently from a large one (the concern raised in [FEATURE]: per-model compaction threshold (reserved/buffer) anomalyco/opencode#43703).

This adds three opt-in, token-based knobs. When they are unset, selection and trigger behavior are unchanged.

• extract_ratio / recent_ratio: size the verbatim tail as (1 - extract_ratio) of the current scoped tokens, and size the recent-context slice as recent_ratio of it. Because they scale with the conversation, a small session is never fully summarized.
• trigger_ratio: compact when usage reaches trigger_ratio x context — a percentage of the window — instead of the fixed-headroom threshold. This is inherently context-window-relative, which is what multi-model setups need.

It also improves the summary itself: the compaction prompt is a context-restoration document, and the newest slice of the conversation is passed to the summarizer as a relevance signal (<recent_context>) so the summary is weighted toward the session's current direction. This helps the agent keep the task goal across the compaction boundary (anomalyco#41358).

All three ratios are computed on demand from the scoped conversation (history minus hidden prior-compaction turns).

Recommended settings ("infinite session")

These are the settings I've been using for about 6-8 months (even before this new compaction system). You can raise extract_ratio to compact more of the session to have more context remaining between compactions, but in my testing anything near the 0.6 range loses too much fidelity for the agent to continue seamlessly, as if nothing happened. With these settings my long-running sessions keep going without the agent losing track of what it was doing — or how to do something it was doing well before the compaction.

"compaction": {
  "trigger_ratio": 0.85,
  "extract_ratio": 0.4,
  "recent_ratio": 0.15
}

Because the ratios scale with the conversation, the same config behaves sensibly on both small and large sessions.

How did you verify your code works?

• bun typecheck passes in packages/core and packages/opencode, rebased on current dev.
• Ran a real ~815k-token session through a manual compaction: with extract_ratio: 0.4 / recent_ratio: 0.15, the summary was produced as the restoration document, the verbatim tail was preserved via tail_start_id, and follow-up questions were answered from the preserved tail. Verified the split is token-based (via the model token estimate), not message-count based.
• Verified the small-session case: with a ratio set, a session smaller than the old absolute budget keeps its proportional tail instead of summarizing everything.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR