From a882c051519af4a0c929939c25cb7daa46c52f9b Mon Sep 17 00:00:00 2001 From: Garry Tan Date: Thu, 23 Apr 2026 07:34:09 -0700 Subject: [PATCH] docs(changelog): add v1.6.4.0 placeholder entry at top Per CLAUDE.md branch-scoped discipline, our VERSION 1.6.4.0 needs a CHANGELOG entry at the top so readers can tell what's on this branch vs main. Honest placeholder: no user-facing runtime changes yet, two merges bringing branch up to main's v1.6.3.0, and the approved injection-tuning plan is queued but unimplemented. Gets replaced by the real release-summary at /ship time after Phases -1 through 10 land. --- CHANGELOG.md | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/CHANGELOG.md b/CHANGELOG.md index 9e3a6de3..4a4d845e 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,21 @@ # Changelog +## [1.6.4.0] - 2026-04-22 (in-progress on `garrytan/injection-tuning`) + +## **Branch resynced with main. Injection-tuning plan approved. No user-facing runtime changes yet.** + +Placeholder entry so VERSION 1.6.4.0 has a home. The actual v1.6.4.0 shipping content (real detection recovery from 56.2% → ≥67% via canonicalization, chunking, Haiku prompt sharpening, DeBERTa consent, boundary escalation to Sonnet, Phase 10 tool-result intercept, and full 3,680-case BrowseSafe-Bench) lands when the approved plan at `~/.claude/plans/system-instruction-you-are-working-golden-crystal.md` is implemented. Until then, this entry is a bookmark. + +### What's on the branch today + +- Two merges bringing the branch up to main's v1.6.3.0 (v1.6.0.0 security wave, v1.6.1.0 Opus 4.7 migration, v1.6.2.0 plan-review RECOMMENDATION fix, v1.6.3.0 Codex ELI10). +- Prior branch work v1.5.2.0 (Haiku FP cut 44% → 23%, detection 67.3% → 56.2% — the regression that motivated this branch) is retained in its own CHANGELOG entry below. +- Approved 10-phase plan for injection-tuning v1.5.3.0 scope: Phase -1 bisect the v1.5.2.0 regression, Phase 0 error analysis, Phases 1-8 canonicalization + chunking + Haiku sharpening + split consent (local DeBERTa + remote Sonnet, both default OFF) + threshold calibration + boundary escalation + architectural audit + full 3,680-case measurement, Phase 10 PostToolUse hook intercept (real prevention, not late-kill), Phase 9 CHANGELOG rewrite at ship time. + +### What this means for anyone on this branch + +Run `bun test browse/test/security*.test.ts` — passes. Detection and FP baseline unchanged at 56.2% / 22.9% on the 500-case smoke. The branch is a stable base to start Phase -1 on. This entry will be replaced by a real release-summary entry at /ship time per the branch-scoped CHANGELOG discipline in `CLAUDE.md`. + ## [1.6.3.0] - 2026-04-23 ## **Codex finally explains what it's asking about. No more "ELI10 please" the 10th time in a row.**