feat(opencode): support per-model compaction config

[joserdf]

Issue for this PR

Closes#43703

Type of change

• New feature

What does this PR do?

Adds per-model compaction configuration so models with different context
windows can use different compaction thresholds.

Before: the top-level compaction block (auto, prune,
tail_turns, preserve_recent_tokens, reserved) was global — one
value applied to every model. The only per-model lever was
limit.context, which shifts the trigger point but cannot set an
independent buffer per model.

After: each model can override any compaction field via
provider.<id>.models.<model>.compaction. Model-level fields are merged
over the global config (model wins per-field), so a 32k local model can
compact early while a 128k frontier model keeps a larger usable window.

Example:

{
"provider": {
"my-provider": {
"models": {
"small-model": {
"limit": { "context": 32000, "output": 4096 },
"compaction": { "reserved": 4000 }
},
"big-model": {
"limit": { "context": 128000, "output": 8192 },
"compaction": { "reserved": 24000 }
}
}
}
}
}

Implementation:

• Extracted the compaction schema into ConfigProviderV1.Compaction and
  reused it for both the global config and the per-model Model schema
  (single source of truth, no behavior change for existing configs).
• Carried the resolved compaction config onto the runtime
  Provider.Model so the session layer can read it.
• Added compactionConfig() in session/overflow.ts that merges
  model-level fields over the global block; usable(), isOverflow(),
  select() (tail_turns / preserve_recent_tokens) and the auto-compact
  call sites in session/prompt.ts now resolve through it.
• prune intentionally stays on the global config: it is a background
  cleanup task keyed only by sessionID, and the model is not readily
  available at that call site.

How did you verify your code works?

• tsgo --noEmit passes in packages/core and packages/opencode.
• bun test test/session/compaction.test.ts — 59 pass, 1 skip, 0 fail
  (added 4 tests: per-model auto, per-model reserved, per-model
  tail_turns, and a regression check that models without an override
  keep using the global config).
• bun test test/provider/ — 557 pass, 0 fail.
• bun test test/config/ (core) — 50 pass, 0 fail.
• The 5 failures in the full test/session/ run are pre-existing
  inotify_add_watch: No space left on device environment failures —
  they fail identically on the base dev branch without these changes.

Screenshots / recordings

N/A (not a UI change)

Checklist

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

[Enough1122]

AI code review — automated review for reference; please use your judgment.

• packages/opencode/src/provider/provider.ts:1057-1063 — the inline Compaction struct duplicates ConfigProviderV1.Compaction but with weaker validation: here fields are plain Schema.Number (negatives accepted), while the core config schema uses NonNegativeInt. A models.json/provider-metadata entry with e.g. reserved: -1000 would validate here and then inflate usable() (input.limit.input - (-1000)), silently raising overflow thresholds. Reuse the core schema (export its type too) so both layers agree.
• packages/opencode/src/provider/provider.ts:1542 — compaction: model.compaction ?? existingModel?.compaction replaces any previously-known per-model block wholesale on provider refresh, inconsistent with the sibling headers: mergeDeep(existingModel?.headers ?? {}, ...). A partial upstream update like {auto: false} will drop an earlier {reserved: 60000} instead of merging. Consider mergeDeep for parity, or document the replace semantics.
• packages/opencode/src/session/prompt.ts:1166-1171 and 1329-1334 — auto: model.compaction?.auto ?? true conflates "the model opted out of auto-compaction" with the create() flag describing who initiated this run. Both call sites are automatic triggers, so with auto: false they now record a user-style compaction; enforcement arguably belongs in isOverflow (which already short-circuits) plus an early return here. Worth double-checking the intended semantics of the auto argument.
• packages/core/src/v1/config/config.ts:146-150 — clean dedup of the inline struct into ConfigProviderV1.Compaction; field-for-field identical annotations, no behavior change. This makes the provider-layer duplication in the first bullet more avoidable.
• packages/opencode/test/session/compaction.test.ts:566-632 — good override matrix (auto, reserved, tail_turns, unaffected-model isolation). Two gaps worth closing: provider-schema rejection of negative values, and refresh-time merge/partial-update behavior for an existing model's compaction block.