Commit Graph

5 Commits

Author SHA1 Message Date
kami 047e2a4070 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
2026-07-01 14:29:56 +04:00
kami e0c222392c feat(context,kernel): compression pipeline stage 3 — replay-safe tier summarizer
ContextSummarizedEvent (registered) keyed on stable session sequence, not the
ephemeral pack ordinal, so replay reuses the recorded summary artifact and never
re-calls the LLM (invariants #6/#8). TierContextSummarizer mirrors
JournalCompactionService for freeform context (tool outputs/docs/retrieved text)
the journal path doesn't cover. ContextState + DefaultContextReducer fold the
event; reducer replay-safety unit-tested.
2026-07-01 14:03:47 +04:00
kami b50688df1d feat(context): compression pipeline stage 2 — fact sheet + strict allocation
FactSheet derives load-bearing spans (newest-first, capped) and pins them as an
L0 SYSTEM entry never dropped by the compressor (level 7). Event-derived each
assembly, not a persisted file, per invariant #1. Strict budget allocation
(level 9) compresses structured groups before freeform so conversation can't
starve tool logs / artifacts.
2026-07-01 13:56:09 +04:00
kami 61be90070f feat(context): compression pipeline stage 1 — level policy + deterministic stages
Additive CompressionPolicy (levels 1-9), ProtectedSpanTagger (regex
load-bearing spans), FormatCompressor (json→dotted lines), ContextClassifier
(static/structured/freeform). Stage 1 (FORMAT_COMPRESS) wired into
DefaultContextPackBuilder behind the policy; structured entries only, lossless.
PromptRenderer repetition-anchoring of steering directives.
2026-07-01 13:53:19 +04:00
kami bd4dd91bf1 feat(context): pin decisionJournal source type through compression 2026-06-04 00:36:12 +04:00