localctx

Release Notes

Changelog

Versioned milestones across all 14 Apple Silicon releases (April 10 – November 4, 2026) for the localctx CLI, native Metal runtime, and MCP server.

4.9.0 & 4.8.0 · 18-Repo Held-Out PR Search, 5-Language Symbol Audit & Warm Search Cache (2026-11-04)

  • 18-Repo Post-Sept-2026 Held-Out PR Localization (v4.9.0 & v4.8.0, 2026-11-04): Day-0 cold evaluation (0 s setup wait) across 18 unseen open-source repositories never used during tuning: Set 3 (n = 90 across 9 post-Sept-2026 repos) reached 53.3% first-try file match on localctx ask and 37.8% on zero-GPU localctx search vs. 23.3% ripgrep; Set 5 (n = 90 across 9 newer repositories across Rust, TypeScript, Python, Ruby, and Go) reached 38.9% on zero-GPU localctx search vs. 28.9% ripgrep (65.6% vs. 53.3% top-3 file match), while disclosing per-repository losses where ripgrep leads on lobsters/lobsters (20.0% vs. 10.0%), mastodon/mastodon (40.0% vs. 30.0%), and Set 3 rubocop/rubocop (30.0% vs. 20.0% zero-GPU search, while localctx ask reaches 50.0%) (2026-11-04 benchmark report).
  • 5-Language Universal Ctags Symbol Audit & 99.5 ms Warm Search Cache (v4.9.0): Independent symbol audit against Universal Ctags across 10 repositories in Rust, Go, Python, TypeScript, and Ruby recalled 300/300 definitions (100.0%) and 200/200 references (100.0%) at 99.6% precision (249/250 definitions and 249/250 references verified). External per-commit SSD search cache cut warm Set 5 search median latency to 99.5 ms (down from 183.7 ms cold, 0 output divergences across 25-run steady-state sweeps) and CLI flag queries to 65.5 ms warm p50, alongside Agent-Audit 50 (85.0% first-try file match on localctx ask and 77.5% on localctx search vs. 72.5% ripgrep across 40 positive questions, 50/50 verified citations, and 4/10 vs. 0/10 negative trap abstentions).

4.7.0-beta.1 & 4.6.0 · Hardware Beta & Active-Control Benchmark (2026-11-01)

  • 180-PR Active-Control Day-0 Benchmark (v4.7.0-beta.1, 2026-11-01): Expanded pre-registered Day-0 cold bug localization to 180 real merged upstream GitHub bug-fix PRs (120 frozen Set 1 PRs + 60 unseen Set 2 PRs, seed = 20261101, 0 s setup wait after git clone) against a zero-install ripgrep active control—reaching 46.7% right file on first try and 23.8% top-3 line-hunk overlap on localctx ask (E4B, 881 tokens/query = 7.8× fewer tokens sent to Claude than ripgrep's 6,876) and 44.4% on zero-GPU localctx search (0.42 s p50, 67.0% top-5 multi-file bug coverage vs. 55.9% ripgrep), while reporting where ripgrep leads on raw speed (0.07 s) and where sense 1.14.2 leads after its 477.2 s index build on zero-identifier symptom PRs (42.2% vs. 37.8%) and on rails/rails (46.7% vs. 43.3%).
  • Disk-Safe Index Guardrails & Non-Destructive Agent Config Merge (v4.7.0-beta.1): Added pre-flight APFS free-space enforcement (15 GiB reserve floor, single-partition default cap, and --dry-run size estimator on --build-kv), separated three-section localctx stats diagnostics, and enforced .bak-localctx backup-before-merge protection on localctx install --write.
  • 120-PR Blind Unseen Repo Baseline (v4.6.0, 2026-10-31): Initial pre-registered Day-0 cold bug localization right after git clone across six unseen open-source repositories—40.8% right file on first try with 0 s setup wait (E4B, 438 tok/query) and 26.4× fewer tokens sent to Claude than ripgrep (2026-10-31 benchmark report).
  • Active CLAUDE.md, AGENTS.md & .mdc Rule Enforcement (2026-10-22): Automatic in-place indexing of root and nested AGENTS.md, CLAUDE.md, and .cursor/rules/*.mdc files—attaching matching file- and symbol-scoped rules (≤ 120 tokens, 22/22 adherence in 0.84 ms, 0 MB GPU) right when an agent touches a file and flagging broken rule paths in localctx review (2026-10-22 benchmark report).
  • Zero-Python Native Apple Silicon Runtime & Uncommitted Edit Sync (2026-10-12): Native Apple Metal arm64 execution for Gemma 4 E4B (4.24 GB RAM, 42/43 = 97.7% on basecamp/fizzy) and 26B-A4B (14.47 GB RAM, 43/43 = 100.0%) with 128/128 exact output match vs. reference (0.00000 max difference), 4.30 s p50 (26B-A4B) warm answers across 50 uncommitted file edits (16.28× faster than v4.5; E4B not measured on Gate 7), 1.73 s/question batch mode, 20 sub-65 ms zero-GPU code tools, and Mac unified-memory guards (2026-10-12 benchmark report).
  • Offline License Tokens, Boundary Checks & Free Symbol-Graph Cap: Cryptographically signed offline license tokens (7-day validity + 14-day grace) in macOS Keychain, canonicalized repository root and secret exclusion (.env*, keys, credentials), and deterministic 65,536-token cross-file symbol graph selection on Free.

4.5.0 · 12-Language Syntax Parser & Warm 26B Index (2026-09-25)

  • Compiled 12-language syntax parser, cross-file symbol graph ranking, and Rails view-to-Stimulus-to-controller linking—lifting multi-file trace accuracy from 35.0% to 100.0% (20/20), 10-language Day-0 onboarding to 21/21 (100.0%), and warm basecamp/fizzy Q&A to 42/43 (97.7%) on 26B-A4B and 40/43 (93.0%, −79.80% cloud tokens) on E4B (2026-09-25 benchmark report).

4.4.0 · Zero-Setup 10-Language Onboarding & Cold vs. Warm Test (2026-09-10)

  • Day-0 cold onboarding across 10 languages with zero setup wait (20/21 cross-file facts in 11.34 s mean vs. a 38-minute cold index build), sub-0.45 s refusal on deleted uncommitted code (5/5), and controlled verification showing vague bug reports with no symbol names drop from 80.0% (warm E4B) to 30.0% without a warm index (2026-09-10 benchmark report).

4.3.0 · 1,017-File Monorepo Scale & Laptop Thermal Soak (2026-08-26)

  • Indexed all 1,017 files (567,912 tokens) of basecamp/fizzy across 17 partitions with 4-agent parallel load testing, 16 GB RAM caps, 90.0% exact ±3 line hits in 60.6K slices, and +0.1% unplugged battery latency drift across 50 back-to-back queries (2026-08-26 benchmark report).

4.2.0 · Multi-Partition Routing & E4B vs. 26B-A4B Comparison (2026-08-11)

  • Added 17.75 ms 32-partition metadata scanning and top-1..5 slice routing—doubling multi-file Q&A accuracy from 44.4% to 88.9% across 1.81M tokens—alongside six deterministic line-citation guards and head-to-head testing of compact E4B vs. large 26B-A4B (2026-08-11 benchmark report).

4.0.0 · 64K Disk Partitions Across 1.81M Repo Tokens (2026-07-24)

  • Replaced single giant 256K buffers with 64K-token on-disk partitions—delivering 3.29× faster index builds (572.2 tok/s), 25.52 ms zero-copy disk attach, and 5.41× faster warm answers (2.04 s vs. 11.02 s) across 1.81M tokens in 4.2 GB active GPU RAM (2026-07-24 benchmark report).

3.0.0 · Sub-2-Second Always-Warm Background Daemon (2026-07-09)

  • Persistent Unix-socket background daemon with 18 µs in-RAM cache snapshot resets and 15-step Metal speed tuning—cutting warm response time to 1.39 s p50 (1,612.9 ms mean, 16/16 under 2.0 s) and reducing tokens sent to Claude by 27.7× across Ruby, Rust, Go, and TypeScript repos (2026-07-09 benchmark report).

2.0.0 · 20-Task Live Claude Coding Loop & 25.5 ms Disk Reload (2026-06-24)

  • Paired 20-task live coding agent run on basecamp/fizzy cutting billed cloud input tokens by 76.7% (113,173 vs. 484,681) and median task completion time by 2.83× (6.02 s vs. 17.01 s) at 20/20 accuracy, paired with 25.5 ms direct NVMe cache loading and 38.5% lower indexing GPU RAM (2026-06-24 benchmark report).

1.1.0 · Multi-Step Q&A vs. Vector RAG & 2.5 ms Git Edit Sync (2026-06-09)

  • Measured +14.0 to +15.4 percentage points higher multi-step Q&A accuracy over top-6 vector embedding RAG (n = 150), +15.0 points on LongBench-v2, 30/30 accuracy on uncommitted git file edits in 2.5 ms (0.0% stale), and 49/50 localization on basecamp/fizzy (2026-06-09 benchmark report).

1.0.0 · 4-Bit Apple Metal Context Engine & Persistent Caches (2026-05-26)

  • High-precision non-evicting 4-bit Apple Metal attention engine delivering 3.16× faster 256K generation (17.42 vs. 5.51 tok/s), 4×–64× lower numerical drift than stock MLX quantized attention, 16/16 parity across compact (E4B) and workstation (26B-A4B) Gemma 4 checkpoints, and 654.1 ms 256K disk cache reload (2026-05-26 benchmark report).

0.2.0 · Cache-Eviction Engineering Erratum — Retracted (2026-05-11)

  • Post-run audit retracted experimental 4-bit/2-bit cache-block eviction after uncovering chunk-starvation Defect F1 (768/1,024 positions starved) and log discrepancies F2–F3 across 2K–128K sweeps, motivating v1.0's non-evicting 4-bit cache (2026-05-11 erratum).

0.1.0 · Compressed GPU Span Index vs. CPU Keyword Search — Falsified (2026-04-24)

  • Falsified compressed GPU span scoring on live Gemma 4 26B-A4B code prompts after 5/5 synthetic vector recall fell to 3/5 (60.0%) under common-word noise and zero-GPU keyword search (20/20) beat weighted GPU spans (19/20 at 16.1 GB RAM) (2026-04-24 benchmark report).

0.0.0 · Constant-Memory Rolling Buffer Prototype — Falsified (2026-04-10)

  • Falsified fixed-size rolling GPU state compression on planted-fact retrieval across 64 to 256 steps after older tokens faded within 3 updates (0.0148 recall at 64 steps vs. 1.0000 with an exact cache), closing lossy rolling buffers (2026-04-10 benchmark report).