# Changelog

All notable changes to this project are documented in this file. The format is
based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this
project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

## [2.0.0] - 2026-08-16

### Added
- **Rust detector pack** (`references/language-packs/rust.md`) — ownership-aware
  detection: `unsafe` escape hatches, ignored `Result`s, needless clones,
  lock-guard-across-`.await`, `#[allow(...)]` suppression hiding.
- **Language packs** for TypeScript/JavaScript, Python, and Go, consolidated from
  the inline detectors into a single `references/language-packs/` catalogue.
- **Harness-agnostic design** — the swarm pattern, three agent definitions, and
  language packs run on any harness that can spawn parallel read-only workers and
  collect structured JSON; dispatch maps to the harness's own worker mechanism.
- **`scripts/validate.py`** — deterministic validator (frontmatter, reference-link
  resolution, one-level-deep reference check, no machine-local paths,
  language-pack registration, directive-injection scan).
- **Eval harness** — `evals/` fixtures (TS/Python/Go/Rust) with seeded issues and
  a manifest, plus `scripts/score.py` for precision/recall scoring.
- **`BENCHMARK.md`** — evaluation methodology and known-detection coverage.
- **`CONTRIBUTING.md`** and this changelog.
- **GitHub Actions CI** (`.github/workflows/validate.yml`) — runs `validate.py`
  and the eval harness on every push and PR.
- **`SECURITY.md`** threat model and a directive-injection scan in `validate.py`;
  untrusted-code rule and trajectory-slop detection in the agent prompts;
  **`references/decisions.md`** architecture decision records.
- **Performance-finding discipline** — a `performance` finding must state a
  measurable basis (algorithmic argument or reproducible measurement); raw
  speedup claims are unsafe evidence.
- **Silent-failure coverage escalation** — a swallowed-error finding with no test
  coverage is flagged `RISKY` (try/catch is the most under-tested construct in
  agent-generated PRs).
- **Negative eval fixtures** — clean-code and red-herring fixtures that must
  produce zero findings; `scripts/score.py` now measures precision, not just
  recall.
- **Verify-before-apply + correction-rate** — the consolidator re-reads
  high-confidence / RISKY findings against the tree before accepting them and
  reports a correction rate.
- **Correction feedback loop** — `references/corrections.md` + ADR-008/009 turn
  real-world corrections into gated detection-pattern improvements.

### Changed
- `SKILL.md` restructured for cross-platform portability (Agent Skills standard).
- Language detectors moved from inline tables to `references/language-packs/`.
- Config path renamed `.kensei/simplify.yaml` → `.simplify-swarm.yaml`.
- Internal tool/skill references sanitized; the automated-fix-pass trigger no
  longer names a specific internal tool.
- Verification ladder + named terminal states (loop-specification pattern).
- README overhauled: badges, table of contents, architecture diagram,
  simplify-code comparison, approval-gate behaviour, cross-platform install.

### Fixed
- Pitfall numbering (was 1–9 then 12–13) and a stray duplicated fragment in the
  approval-gate pitfall.
- README contradiction: it claimed "SAFE changes applied automatically" while the
  skill enforces a hard approval gate. README now documents the gate correctly.

## [1.4.0] - 2026-06-26 (internal — not previously released publicly)

### Added
- `references/concurrency-fix-patterns.md` — extract-under-lock, read-copy-update,
  and degradation-fallback patterns.
- `references/swarm-triage-pitfalls.md` — severity-inflation patterns and a user
  triage workflow.
- automated-fix-pass trigger, untracked-files exception, and multi-agent
  consensus elevation merge rule.
- Pitfalls: `find -path` vs `Path.glob` depth mismatch, approval-gate violation,
  severity inflation.

## [1.3.0] - 2026-06-09

### Added
- Hard approval gate: present the consolidation report and stop before applying
  any tier, including SAFE.
- Dispatch progress visibility: a status block before the parallel dispatch so
  the user is not left staring at silence.

## [1.1.0] - 2026-06-08

### Changed
- Hygiene agent optimized with SkillOpt against 8 annotated messy files.

## [1.0.0] - 2026-06-08

### Added
- Initial release: multi-agent code simplification (Hygiene, Clarity, Correctness)
  with risk-tiered application, language-specific detectors for TS/JS, Python,
  and Go, and structured JSON output per agent.