feat(context,infra): compression pipeline stages 4-5 — token pruning + relevance + ToMe

- build() is now suspend: pipeline runs all stages in fixed order, each gated by level
- TOKEN_PRUNE (level 3): TokenPruner interface + LLMLingua-2 sidecar (sidecars/llmlingua)
  + HttpTokenPruner adapter (fails open if sidecar down); prunes freeform, preserves
  protected spans, skips tier-0 turns when TIER_SPLIT on
- TOME_MERGE (level 8): ToMeMerger collapses near-duplicate freeform turns (Jaccard)
- Stage 5 selection: RelevanceScorer + EmbeddingRelevanceScorer (cosine over Embedder);
  query-conditioned reorder so least-relevant freeform drops first under budget
- [orchestration] compression_level + token_pruner_url config, wired in Main
- suspend ripple fixed across builder callers/stubs
This commit is contained in:
2026-07-01 14:29:56 +04:00
parent e0c222392c
commit 047e2a4070
25 changed files with 509 additions and 65 deletions
@@ -29,7 +29,7 @@ class PromptRendererOrderingTest {
)
@Test
fun `stage tool loop renders chronologically with the task before tool turns`() {
fun `stage tool loop renders chronologically with the task before tool turns`() = kotlinx.coroutines.runBlocking {
// The builder stamps ordinals from input order; the renderer must honour them so the
// model sees: system, task, assistant call, tool result, assistant call, tool result.
val builder = DefaultContextPackBuilder(DefaultContextCompressor())
@@ -49,6 +49,7 @@ class PromptRendererOrderingTest {
"tool" to "result1", "assistant" to "call2", "tool" to "result2"),
messages.map { it.role to it.content },
)
Unit
}
@Test