|
2 | 2 |
|
3 | 3 | All notable changes to CKB will be documented in this file. |
4 | 4 |
|
| 5 | +## [Unreleased] |
| 6 | + |
| 7 | +## [9.1.0] - 2026-04-16 |
| 8 | + |
| 9 | +### Added |
| 10 | + |
| 11 | +- **LIP v2.1 utilisation** — three high-ROI LIP RPCs wired into the query |
| 12 | + engine, gated on the handshake's `supported_messages`: |
| 13 | + - `stream_context` (v2.1) → `explainFile` attaches up to 10 |
| 14 | + semantically-related symbols (2048-token budget) in `facts.related`. |
| 15 | + New streaming transport reads N `symbol_info` frames + `end_stream`. |
| 16 | + - `query_expansion` (v1.6) → `searchSymbols` expands ≤ 2-token queries |
| 17 | + with up to 5 related terms before FTS5, recovering vocabulary-mismatch |
| 18 | + recall without touching precision on compound queries. |
| 19 | + - `explain_match` (v2.0) → semantic search hits carry up to two ranked |
| 20 | + evidence chunks with line ranges, text, and per-chunk scores (top-5 |
| 21 | + hits, bounded round-trip cost). |
| 22 | +- **`lip.Handshake` runs on engine startup** and the daemon's |
| 23 | + `supported_messages` list is stashed for feature gating |
| 24 | + (`Engine.lipSupports`). Daemon version and supported-count logged. |
| 25 | +- **LIP index status probing** — `probeHandshake` now follows up with |
| 26 | + `IndexStatus` and caches the result. New `Engine.LIPStatus()` returns |
| 27 | + `{Reachable, IndexedFiles}` so consumers can distinguish "daemon down" |
| 28 | + from "daemon up, nothing indexed." |
| 29 | +- **`ckb review` warns when LIP index is empty** — stderr advisory with |
| 30 | + `lip index <repo>` command when daemon is reachable but has no content. |
| 31 | + Suppressed in `--ci` to keep CI logs clean. |
| 32 | +- `NoAutoFetch` option on `SummarizePROptions` and `SummarizeDiffOptions` |
| 33 | + for parity with `ReviewPROptions`. |
| 34 | +- Troubleshooting section in `docs/plans/review-cicd.md` covering shallow |
| 35 | + CI clones, auth-failure remediation, air-gapped pipelines, and depth-0 |
| 36 | + checkout alternatives. |
| 37 | +- Auth-error detection on auto-fetch with clear remediation guidance. |
| 38 | +- `ckb review --no-auto-fetch` flag for air-gapped pipelines. |
| 39 | +- Test coverage for `GitAdapter.EnsureRef` — happy path, missing-ref |
| 40 | + auto-fetch, unreachable origin, and empty-input guard. |
| 41 | +### Changed |
| 42 | + |
| 43 | +- **LIP health: push-driven, not polled** — Engine opens a long-lived |
| 44 | + connection to the daemon at startup (`internal/lip/subscribe.go`) with |
| 45 | + `index_changed` frames and per-ping `index_status` snapshots instead of |
| 46 | + 60 s TTL polling. Worst-case staleness drops from 60 s to ~3 s. |
| 47 | +- **`lipFileURI` path normalisation** — handles absolute paths and |
| 48 | + already-prefixed `file://` URIs without producing malformed results. |
| 49 | + |
| 50 | +### Fixed |
| 51 | + |
| 52 | +- **Bug-pattern false positive on `sync.Mutex.Lock()`** — removed `"Lock"` |
| 53 | + from `LikelyReturnsError` heuristic patterns; `sync.Mutex.Lock` returns |
| 54 | + nothing and dominated real-world matches with false positives. |
| 55 | +- **`err` shadowing in `subscribe.go`** — four shadow sites eliminated by |
| 56 | + reusing outer `err` or renaming to `pingErr`/`readErr` where scope |
| 57 | + isolation requires it. |
| 58 | + |
| 59 | +- **LIP rerank: coherence gate + position-weighted seeding** (#209) — the |
| 60 | + Fast-tier semantic rerank (`internal/query/lip_ranker.go`) used to average the |
| 61 | + top-5 seed embeddings with uniform weight and always apply the result. When |
| 62 | + the top-5 pointed in different directions the centroid collapsed toward zero |
| 63 | + and amplified noise; when the top seed was strong the blend still diluted it. |
| 64 | + Seeds are now L2-normalised and position-weighted (`1/(rank+1)`), the |
| 65 | + resulting centroid norm is read as a coherence score in `[0, 1]`, and the |
| 66 | + rerank falls back to pure lexical order when coherence is below |
| 67 | + `MinCoherence` (default `0.35`). Blend weights, seed count, and threshold are |
| 68 | + surfaced as `RerankConfig` so future tuning does not need to touch call |
| 69 | + sites. Injected `embedBatchFn` makes the ranker unit-testable without a |
| 70 | + running daemon. |
| 71 | +- **LIP rerank: gate on `!MixedModels`** (#208) — when the LIP index contains |
| 72 | + vectors from more than one embedding model (e.g. partial re-index during a |
| 73 | + model upgrade), cosine similarity across those vectors is mathematically |
| 74 | + meaningless. `RerankWithLIP` and `SemanticSearchWithLIP` now consult a |
| 75 | + cached `Engine.lipSemanticAvailable()` check (60 s TTL, single `IndexStatus` |
| 76 | + RPC) and fall back to lexical ranking when the daemon is down or reports |
| 77 | + `mixed_models`. A new `lip_mixed_models` degradation warning (70% capability) |
| 78 | + surfaces in response metadata so users learn *why* results look weaker |
| 79 | + instead of silently ranking on garbage. |
| 80 | + |
5 | 81 | ## [9.0.1] - 2026-04-15 |
6 | 82 |
|
7 | 83 | ### Fixed |
|
0 commit comments