Tags

v0.17.0

v0.17.0: Explicit harness declaration, MCP server, codebase cleanup

Breaking change: --wide and --strict flags removed. he now requires
explicit path declaration via .he.yml config or CLI arguments.
Reference following (@path, markdown links) always runs.

New:
- MCP server with stdio and HTTP transports (he_analyze, he_lint, he_recommend)
- Explicit harness declaration via .he.yml or CLI args
- Phase 12 harness simulation specification

Cleanup:
- Consolidated 4x truncate, 3x estimateTokens, 2x tokenize into internal/textutil
- Replaced 180-line classifier with 3-line IsSkippableDoc
- Removed dead code across 6 packages
- Fixed ONNX runtime crash on repeated init/destroy (sync.Once + ref counting)
- Fixed ONNX shared library extraction race (atomic write + rename)
- Fixed export token display bug (DisplayName key mismatch)
- Fixed swallowed errors in writeJSON, analysis.Run, MCP server.Serve

Testing:
- ~160 Go tests across all packages (was ~40)
- 32 frontend tests via vitest
- JUnit XML reporting for Go and frontend in GitLab CI
- Bumped Go to 1.26

Thanks to Chad Woolley (@cwoolley-gitlab) for providing gitlab-ai-harness
as a real-world test case. His feedback and harness structure drove several
bug fixes and directly informed the explicit-declaration design.

v0.16.0

CI: Automated macOS builds via hosted M2 Pro runner

Both linux-amd64 and darwin-arm64 binaries are now built and released
automatically in CI. No more manual macOS binary uploads.

- Added build-macos job using saas-macos-large-m2pro runner
- Release job creates asset links for both platforms

v0.15.0

Wide mode default, auto-detection fix, docs cleanup

- Wide mode is now the default discovery behavior (--strict to disable)
- Auto-detect bare skills/ and agents/ directories
- Methodology component updated: sentence-BERT, Phase 8/9 metrics
- New dashboard screenshot
- AGENTS.md and ROADMAP.md trimmed
- MUST use make build convention documented

v0.14.0

Phase 9: Knowledge graph analysis + recommendations engine

- Wide mode discovery: scan docs/, ARCHITECTURE.md, follow @path and markdown link references
- Cross-level redundancy, contradiction, and shadow detection
- Progressive disclosure scoring and token budget by level
- Recommendations engine with ranked leverage points (CLI + API + frontend)
- README rewritten around the leverage pitch
- CI release job 409 fix

v0.13.0

Phase 8: Information Per Token metrics

New metrics:
- MDL contribution (leave-one-out compression, bytes/token)
- Corpus cross-entropy surprisal (bigram, bits/token)
- IDF density (mean IDF/token)
- Composite information score (normalized 0-1)

Also includes:
- Instruction type extraction refactored to analysis package
- XML wrapper tag similarity inflation fix
- Contradiction detector threshold calibration (polarity 0.6, scope 0.7)

v0.12.0

v0.11.0

v0.11.0: Semantic Embedding Engine

Replace TF-IDF with sentence-BERT (all-MiniLM-L6-v2) for similarity
detection via ONNX Runtime. Platform-specific builds (darwin-arm64,
linux-amd64). CI updated for CGo + Git LFS.

v0.10.0

v0.10.0: architecture refactor, historical metrics, CI improvements

v0.9.0

feat: test strategy classification for instructions

- Classify rules by testability: pattern, llm-judge, manual
- Add ID, TestStrategy, Testable, Keywords, Pattern fields to Instruction
- Unicode emoji detection via regex ranges
- strategy_counts in /api/instructions response
- ROADMAP.md for evaluation loop implementation

v0.8.0

feat: deontic_strength metric for normative language analysis

v0.7.0

v0.5.1

v0.5.1 - Documentation improvements

Add references and clickable links to all metrics in Methodology section:

- TF-IDF/Cosine Similarity: Salton & Buckley (1988)
- Redundancy Score: DRY principle, Hunt & Thomas (1999)
- Inheritance Violation: Liskov Substitution Principle
- Lateral Duplication: Roy & Cordy (2007) clone detection survey
- Compression Ratio: Benedetto et al. (2002)
- Distinct-N: Li et al. (2016) NAACL-HLT
- Instruction Density: RFC 2119
- Verb Density: Strunk & White (1959)
- Cross-File Vocabulary Overlap: Manning et al. (2008)
- Specificity Ratio: Spärck Jones (1972)
- Position Importance Score: Liu et al. (2023) arXiv:2307.03172
- Negative Constraint Ratio: Anthropic Claude's Character
- Hedging Density: Singhal et al. (2023) arXiv:2310.03716
- Boilerplate Lead Score: Xiao et al. (2024) arXiv:2309.17453

Updated dashboard screenshot.

v0.5.0

Harness Explorer v0.5.0

Major release renaming the project from 'Instruction Hierarchy Explorer' (ihe)
to 'Harness Explorer' (he), reflecting industry terminology convergence
(Microsoft VS Code, promptfoo, PrimeIntellect).

BREAKING CHANGES:
- Binary: ihe → he
- Go module: gitlab.com/andunn/ihe → gitlab.com/andunn/he
- Container: ENTRYPOINT changed to 'he'
- Update any scripts or aliases referencing the old name

NEW METRICS (research-backed, all local/offline):

1. Position Importance Score
   - Measures instruction density in first/last 10% vs middle 80%
   - Based on 'Lost in the Middle' (arXiv:2307.03172)
   - Score >2.0 indicates good structure; ~1.0 means flat distribution

2. Negative Constraint Ratio
   - Ratio of NEVER/AVOID/PROHIBITED vs MUST/ALWAYS/SHOULD
   - Based on Claude's Character research
   - Balanced ratio (~0.3-0.5) suggests healthy constraint mix

3. Hedging Density
   - Frequency of hedging words (might, could, perhaps, generally)
   - Based on RLHF Length Bias research (arXiv:2310.03716)
   - Lower = more direct, actionable instructions

4. Specificity Gradient
   - Positive reframe of inheritance_violation (1 - score)
   - Higher = children differentiate well from parents

5. Boilerplate Lead Score
   - Detects files starting with generic intros vs substance
   - Based on Attention Sinks research (arXiv:2410.13835)
   - Higher = more files lead with important content

UI IMPROVEMENTS:
- Three-column metrics dashboard with logical groupings
- Color-coded quality indicators (green/amber/red dots)
- Progress bars showing relative metric values
- Hover tooltips explaining each metric's meaning

OTHER CHANGES:
- Auto-detect .gitlab/duo/skills directory (GitLab Duo)
- Auto-detect .cursor/rules directory (Cursor IDE)
- CI creates proper GitLab Release with binary download links
- Project logo added (assets/he.png, LFS tracked)
- Unit tests for new metric functions

INSTALLATION:
  curl -L https://gitlab.com/andunn/he/-/releases/v0.5.0/downloads/he-darwin-arm64 -o he
  chmod +x he

CONTAINER:
  docker run --rm -v $(pwd):/work -p 8765:8765 registry.gitlab.com/andunn/he:v0.5.0

v0.4.0

v0.4.0 - Public release preparation

Major changes:
- Moved to gitlab.com/andunn/ihe
- Added MIT license (GitLab copyright)
- Added Dockerfile and container CI pipeline
- Converted all visualizations to canvas for performance
- Added Methodology section with LaTeX formulas and pseudocode

New features:
- Section Coverage now works across all granularities
- Redundancy Candidates shows content excerpts
- Git history resolves symlinks correctly
- Radial compression chart and radar diversity chart

Cleanup:
- Removed legacy Observable Framework files
- Simplified README and AGENTS.md
- Updated module path and all Go imports

Screenshot demo uses Anthropic's official Claude Code plugins.

v0.3.0

v0.3.0 - React Frontend & Enhanced Metrics

Major Changes:
- Replaced Observable Framework with Vite + React + shadcn/ui
- Single-page dashboard (all visualizations on one page)
- New backend metrics and API endpoints

Frontend (frontend/):
- Vite + React + TypeScript + Tailwind CSS v4
- shadcn/ui components (Card, Select, Slider, Table)
- D3 visualizations with React useRef pattern
- Custom hooks for API data fetching (useApi.ts)
- TypeScript interfaces for type safety (types.ts)

New Visualizations:
- SectionMatrix: Two-column layout showing which sections appear in which files, shared sections highlighted
- InstructionList: All MUST/NEVER/ALWAYS/SHOULD statements with filtering by type and file
- HistoryTimeline: Git commit history for analyzed files

New Backend Metrics:
- max_hierarchy_depth: Deepest file in hierarchy
- avg_hierarchy_depth: Average depth across files
- total_instruction_count: Count of imperative statements
- avg_verb_density: Action verbs per token

New API Endpoints:
- /api/sankey: Nodes (files + sections) and links for hierarchy visualization
- /api/history: Git log for each analyzed file
- /api/instructions: Extracted imperative statements with type, file, line, context

Existing Visualizations (ported to React):
- Similarity Heatmap
- File Network (circular static layout)
- Chord Diagram
- Treemap
- Circle Packing
- Compression Chart
- Diversity Chart
- Redundancy Candidates List

Build:
- cd frontend && npm run build
- cp -r frontend/dist web/dist
- go build -o bin/ihe ./cmd/ihe

Development:
- Terminal 1: ./bin/ihe <path> AGENTS.md .opencode/skills
- Terminal 2: cd frontend && npm run dev -- --port 3000

v0.2.0

Rewrite: Python → Go + Observable Framework

- Single binary distribution (Go with embedded static assets)
- Observable Framework dashboard replacing vanilla HTML/JS
- D3 visualizations: chord diagram, treemap, circle packing, heatmap, file network
- New metrics: compression ratio, distinct-n, instruction density, vocab overlap, specificity
- Hand-rolled TF-IDF vectorizer (no external dependencies)
- Multi-granularity chunking (document, section, paragraph, rule, sentence)
- Cross-platform builds via GitLab CI

v0.1.0