@luckydraw/cumulus 1.0.31 → 1.0.33

package/CHANGELOG.md

@@ -1,5 +1,26 @@
 # Changelog
 
+## v1.0.33
+
+Patch release. A third of the past-conversation context in every prompt was the same messages printed twice. Fixing it cuts retrieved-context size by a quarter on a real thread with no loss of recall.
+
+- **32.6% of all rendered history lines were duplicates.** Measured over 3,084 prompt captures on the gateway (1,078 with a `[Relevant Past Messages]` block): 8,327 of 25,554 lines were re-renderings of a message index already present in that same prompt — exact identity, not a similarity heuristic, since every line carries its own `#index`. That cost 6.51M of 18.51M history tokens (35.2%), affected 59% of captures, and ran as high as 42.8% on individual threads. Multiplicity was always `2x` and never `3x` anywhere on disk — the bounded distribution that pins a single emit-once-more code path.
+- **Cause: `expandWithSegments` emitted each hit twice.** Expanding a conversation segment emits every scored result inside it; the outer loop then reached those same results and emitted each again. The guard that looked like it prevented this tested a _message_ index against `processedSegments`, which holds _segment start_ indices — two number spaces in one Set, so it only fired when a hit happened to sit exactly on a processed boundary. `searchHistory` is clean (Set-merged); `formatRetrievedContext` does not dedupe, so every duplicate was billed twice against the RAG budget and displaced a different message that would otherwise have fit.
+- **Fix: one `emit()` gate owns uniqueness** (Rule #8 — no second, defensive dedupe downstream, which would be a bridge between two notions of "unique"). The surviving copy keeps the real search score, never the segment decay.
+- On the `fable-config-amnesia` bench case (a real 3,809-message thread): history messages 81 → **49**, retrieved context 8,011 → **5,991** tokens (−25%), full prompt −15.7%, **identical recall** (3/3 gold facts), stored-content items unchanged. Where the budget is not binding the prompt simply gets smaller; on the budget-capped turns — whose history blocks were 41.6% duplicates — the freed space is instead spent on additional distinct messages.
+- Recorded but deliberately **not** changed: a purely-expanded neighbour scores `segmentBest × 0.6` while task 089's relevance floor drops anything under `max(0.15, topScore × 0.7)`, so since 089 segment context has been unreachable and the duplicate was the whole of segment expansion's observable output. Reviving it (raise the decay) or removing it (Rule #8) is a benchable dilution decision, not a bug fix.
+- RLM-touching: bench pass — 2/2 real cases, 12/12 ladder, static prompt 1791/1800 untouched, speed line held against a same-session control run on unmodified HEAD. 1657/1658 tests green (+4; the one failure is a pre-existing `migrate.test.ts` network flake that fails identically on unmodified HEAD under full-suite load), tsc clean, build clean, lint delta 0, negative control proven.
+
+## v1.0.32
+
+Patch release. The adaptive context budget's feedback loop can run in both directions for the first time — its "how full was this prompt?" sensor was pointed at the model's response instead of the prompt. Also picks up `@luckydraw/blex@0.1.19`, so mermaid diagrams render without text spilling outside their boxes.
+
+- **The adaptive budget has never once grown, on any thread.** Measured across all 91 `*.adaptive.json` sidecars on the gateway: 0 GROW, 107 SHRINK, 813 FLAT — 49 threads sitting on the 100,000 floor, including every high-traffic one. `recordTurn`'s `tokensUsed` is documented as "tokens actually packed into the prompt", but the gateway fed it `estimateTokens(fullResponse) + userQueryTokens` — the assistant's **output** plus one prompt component, excluding the system prompt, the recent-conversation window and **all retrieved context**, the largest part and the one the budget exists to size. Growth needs `tokensUsed ≥ 55,250` at the floor; the largest value that shape ever produced anywhere is 7,615. Shrink needs TTFT alone, so it worked — hence a one-way ratchet.
+- **One helper, every caller.** New exported `packedPromptTokens()` sums the five real components (instructions, always-include, recent window, RAG, user message). The gateway feeds it from the figures it already computes for `debug.systemPromptBreakdown`; the agentic path overwrites the instructions figure with its own text, since that branch sends a different instruction block. The TUI, which fed a _third_ definition (the RAG slice alone), now uses the same helper — a source-level test fails if either call site diverges again (Rule #8).
+- **Capacity is measured against the budget the retriever actually gets.** `atCapacity` compared against a private `effectiveBudget()` (`0.75×CL − 10,000`) while packing used `getTotalContextBudget()` (`CL × 2/3`) — two notions of "the budget" in one class, ~55k apart at a 775k pin. The private one is deleted, not kept alongside.
+- Policy constants are untouched, but this is **not** a no-op: the grow branch is now reachable, so a thread that packs ≥85% of its budget and answers in under 4s can grow by 1.2×. Whether larger budgets help is a separate, benchable question (task 088 P3 measured dilution as the real failure class) — this restores a mechanism that only ever ran one way, and makes the sidecars record real packed totals.
+- RLM-touching: bench pass — 2/2 real cases, 12/12 ladder, static prompt 1791/1800, speed line held against a same-session control run on unmodified HEAD (the differ's flag against the older `after-162.json` baseline was machine drift, not the change). 1654/1654 tests green (+9), tsc clean, build clean, lint delta 0, three negative controls proven.
+
 ## v1.0.31
 
 Patch release. The stall classifier now gets an output budget a reasoning model can actually finish in — closing the last case where it failed closed and never nudged.