What

The four compaction strategies (TruncationStrategy, SlidingWindowStrategy, SummarizationStrategy, ToolResultStrategy) computed their minimum-preserved floor with:

minimumPreserved := cmp.Or(max(strategy.MinimumPreserved*, 0), default)

Because max(0, 0) == 0 and cmp.Or(0, default) == default, this folds three distinct inputs into the type default (truncation 32, sliding window 1, summarization 8, tool result 16):

• unset (the intended default case) — correct
• an explicit 0 — silently overridden by the default
• a negative value — silently overridden by the default

So a caller who set MinimumPreservedGroups: 0 to preserve nothing got the default instead, and the aggressive preserve-none configuration was inexpressible. Only positive values worked.

This changes each MinimumPreserved* field to *int: a nil config uses the default, while a non-nil value is honored as-is with a negative clamped to 0:

minimumPreserved := default
if strategy.MinimumPreserved* != nil {
    minimumPreserved = max(*strategy.MinimumPreserved*, 0)
}

Why

The .NET port treats this as a nullable int and applies EnsureNonNegative(value) => Math.Max(0, value), so an explicit 0 disables the floor and negatives clamp to 0. The Go port silently discarded the value instead, leaving no equivalent for the preserve-none case. This restores cross-SDK parity.

Tests

• Added TestTruncationStrategy_ExplicitZeroPreservesNone: an index with 5 non-system groups, nil Trigger/Target, and MinimumPreservedGroups: ptr(0) now compacts and excludes all removable non-system groups rather than preserving the default 32.
• Added TestSlidingWindowStrategy_NegativeMinimumClampsToZero: a negative floor clamps to 0 (excluding all turns) instead of using the default 1.
• Updated existing call sites (contextwindow.go, the step18 example) and existing tests to the pointer field type.

go build ./..., go vet ./agent/compaction/..., and go test ./agent/compaction/... all pass.