Changelog

All notable changes to Forgewright.

[Unreleased]

Features targeting v8.8.0 and beyond. Not yet released.

Added

• Forgewright Rule Compliance Loop — Telemetry, Rule Ledger, Execution Policy, Context Manager, and Rule Validator to self-correct and enforce rules.
• AI Reasoning Research Integration — Deep NotebookLM research across 14 sources (OpenAI o1/o3, Anthropic extended thinking, Claude Code best practices) producing 15 actionable lessons (reconstructed from docs).
• Reasoning Checkpoint — SOLVE Step 6.4 mandates a 1–2 sentence reasoning pause after every CHECK result (reconstructed from docs).
• Adversarial Review — SOLVE Step 6.7 spawns a fresh-context reviewer for FEATURE/DEBUG tasks touching ≥3 files (reconstructed from docs).
• Anti-Narrative Verification — VERIFY Rules 6-7 mark narrative claims as automatically FALSE without command output (reconstructed from docs).
• Tests-First Ordering — Self-check protocol enforces test stubs BEFORE implementation for complex tasks (reconstructed from docs).
• Context Reset — STUCK rule expanded with Step 4 "Reset context" (reconstructed from docs).
• Guardrail Hardening — 13 rule categories integrated into Middleware ④ with 17 regression tests (reconstructed from docs).
• UI Design Gate — Mandatory design contract required before any frontend edit in SOLVE Step 3.D (reconstructed from docs).
• 58 Regression Tests — 19 AI reasoning + 17 guardrail + 13 UI design gate + 9 audit tests (reconstructed from docs).
• Client-Server Sequence Flow Generator — Automated Mermaid sequence diagram generation powered by GitNexus static call-graphs (reconstructed from docs).

Changed

• Workflow Consolidation — Archived overlapping GitHub Action workflows into .github/workflows/archive/ and simplified ci.yml.
• Game Extraction — Removed the game/ directory from the repository structure to decouple game assets from orchestrator CI.
• NPM Workspaces — Configured mcp and src/cli as NPM workspaces in the root package.json to deduplicate dependencies and share the lockfile.
• Pre-commit Hardening — Replaced bash -n with shellcheck in lint-staged for stronger shell script validation.
• Script Reorganization — Reorganized scripts/ into subdirectories (bootstrap, ci, docs, hooks, mcp, memory, release, runtime, skills, telemetry, testing, utilities) and generated shims for backward compatibility.

---

[8.7.0] — 2026-07-01

v8.7 — Evidence-Gated Kernel & Parallel Skill Distillation

Added

• Forgewright Lite — Evidence-Gated Kernel (Upgraded v3): Lightweight reasoning kernel for fast models (Gemini Flash), featuring turn-level script verification via .forgewright/verify/<turn>.json, turn-blocking platform hooks (Claude Code, Gemini CLI, Cursor, Codex CLI), ≤7k tokens boot budget, and objective escalations to Sonnet/Opus models.
• Automated Skill Distillation & Batch Upgrader: upgrade-skills.py queries NotebookLM CLI in parallel via multi-agent subagents to distill all 83 skills to Lite overlays (LITE.md).
• Self-Healing Skill Indexing: Dynamic generation of kernel/INDEX.md by scanning all LITE.md overlays, automatically mapping triggers and paths.
• Agent Benchmark Command (forge agent benchmark): CLI command for evaluating agent performance.
• Expert CLI Mode: Optional premium model escalation for high-stakes planning, architecture, security, and gate decisions via local Claude CLI or Codex CLI (reconstructed from docs).
• Token Tracking by Default: token_tracking enabled by default for all pipelines.
• Webhook State & Telemetry Protocol: Replaces OSC/MCP event mechanism with HTTP Webhook for token and state reporting.
• Momentum Breaker Protocol: Detects and documents LLM psychological traps during execution.
• 99% Empirical Confidence & UI Visual Gate: Protocol for high-confidence assertions and visual verification.
• GraphRAG Memory V4 (FluxMem): SQLite relational cognitive graph (flux_nodes & flux_edges) replacing JSON files. Includes Procedural Circuits caching, ASIP edge decay (0.5) / reinforcement (1.2), and passive 10m idle checkpointing (reconstructed from docs).

Changed

• Gemini 3.x Native Optimization: Direct support for Gemini 3.5 Flash (thinking_level: MINIMAL) and Gemini 3.1 Pro (thinking_level: HIGH, temperature 1.0). Scoped temperature policy by task.
• Hexagonal & CQRS Migration: PipelineState refactored to Hexagonal architecture with CQRS pattern.
• WebSocket → HTTP Webhook: MCP state reporting migrated from WebSocket to HTTP Webhook.
• Marketing & Conversion Skills: Upgraded with 2026 AI/CRO trends and Kelly Criterion.
• Orchestration: Complexity-scaled evaluation, escalation ladder, normalized step numbering.
• Skill count — 83 skills (from 57).

Fixed

• Gemini temperature policy scoped by task type to prevent blanket temperature 1.0 enforcement.
• MCP workspace detection handles unresolved IDE template variables.
• RPC event queuing during WebSocket connection.
• Compliance enforcement for GitNexus analyze and sequence update in post-commit.

[8.6.0] — 2026-06-10

v8.6 — GraphRAG Memory V3 & IDE MCP Unification

Added

• GraphRAG Memory V3: NetworkX-based graph memory with clustering, auto-pruning, and pattern learning (reconstructed from docs).
• Unified IDE MCP Setup (fw-mcp.sh): Single script to configure MCP for all editors (reconstructed from docs).

Changed

• ForgeNexus deprecated: Added DEPRECATED notice; all new projects should use GitNexus (reconstructed from docs).
• Post-commit hook: Auto-reindexes GitNexus on commit (reconstructed from docs).

[8.5.0] — 2026-05-29

v8.5 — GitNexus Migration

Added

• GitNexus Integration: Migrated code intelligence from ForgeNexus to GitNexus (38K+ stars, npm install, auto-setup for all editors, 16 MCP tools) (reconstructed from docs).
• forgewright-mcp-setup.sh v3.0.0: Unified setup script configuring both forgewright and gitnexus MCP servers.
• Multi-repo Support: New gitnexus group for cross-repo analysis.

Breaking

• npx forgenexus analyze → gitnexus analyze
• forgenexus_* MCP tools → gitnexus_* MCP tools
• fw-mcp.sh forgenexus → gitnexus setup (GitNexus) + forgewright-mcp-setup.sh (ForgeWright)

[8.4.0] — 2026-05-10

v8.4 — Mandatory Pipeline Rule & Enhanced ASIP

Added

• Mandatory Forgewright Rule: EVERY user request MUST go through the Forgewright pipeline after installation. Added enforcement box to README and AGENTS.md (reconstructed from docs).
• Enhanced Research Gate: ASIP Research Gate now checks NotebookLM availability first, falls back to Web Search, synthesizes insights, and updates session tracker (reconstructed from docs).
• Session Tracker (scripts/forgewright-session-tracker.sh): Tracks consecutive plan failures; ≥2 consecutive failures trigger mandatory Research Gate (reconstructed from docs).
• Plan Quality Tracking: Mandatory plan → score → meta-evaluate → check ≥9 → execute loop with max 3 iterations.

Changed

• ForgeNexus community detection threshold increased to 20K.
• MCP setup flow simplified for multi-project use.
• Memory optimization (round 2) for ForgeNexus indexer.

[8.3.0] — 2026-04-24

v8.3 — Adaptive Self-Improving Protocol (ASIP)

Added

• Adaptive Self-Improving Protocol (ASIP) (skills/_shared/protocols/asip.md): Combines Plan Quality Loop + Execution Blocker Loop. Mandatory 2-failure-then-research with NotebookLM. Skills improve over time based on real failures (reconstructed from docs).
• ASIP Middleware (skills/production-grade/middleware/10-asip.md): Enforcement middleware for ASIP protocol.
• ASIP Metrics (.forgewright/asip-metrics.json): Tracking file for ASIP protocol performance.
• Lessons file (.forgewright/lessons.md): Project-specific learnings from ASIP failures.
• Multica Hub Token Stats: Token statistics component added to status dashboard (reconstructed from docs).
• Token Tracking & Cost Analytics: forge token on, budget management, real-time tracking, and cost dashboard (reconstructed from docs).

Changed

• Middleware chain updated with ASIP integration.
• Production-grade SKILL.md updated with execution learnings.

[8.2.0] — 2026-06-25

v8.2 — Test Coverage Reporting & State Isolation

Fixed

• Vitest Coverage Reporter: Removed minimatch version override in mcp/package.json to resolve conflicts with test-exclude and enable Vitest coverage generation and reporting.
• Pipeline Test State Isolation: Refactored mcp/src/state/pipeline-manager.test.ts to execute inside isolated temporary directories, preventing test runs from writing to and polluting the workspace's .forgewright/ state directory.

Added

• Sequence Diagram Generator Update: Automated sequence diagram generator to trace routes, extract parameters, and map client-to-server call trees using GitNexus.
• Submodule Auto-Update Check: Added scripts/forgewright-submodule-check.sh to allow projects using Forgewright as a submodule to automatically check, fetch, and pull Forgewright updates in their git hooks (pre-commit or post-merge).

[8.1.0] — 2026-04-29

v8.1 — Autonomous Pipeline Hardening (Clarify-to-Deploy)

Added

• Strict Cognitive Rules (rules/.tieu-mo-deployment-flow.md) — Enforced ground truth for the AI orchestrator to prevent DevOps-related hallucinations during Clarify phase.
• Automated Vercel Project Binding (scripts/task-runner.sh) — Automatically generates vercel.json with a sanitized project name before deployment to resolve Vercel directory name collision bugs.
• HTTP Validation Loop (scripts/task-runner.sh) — Added a self-verifying HTTP ping check after assigning the custom Vercel subdomain to ensure DNS propagation before confirming task completion.
• Global Memory Injection — Integrated memory-wrapper injection to ensure the Orchestrator inherently understands the CI/CD boundaries.

[8.0.0] — 2026-04-13

v8.0 — Stability, Resilience, Audit, Quality, Timeout

Added

• Circuit Breaker Protocol (skills/_shared/protocols/circuit-breaker.md) — State machine pattern (CLOSED → OPEN → HALF_OPEN) to prevent cascading failures in parallel dispatch workers.
• Bulkhead Isolation Protocol (skills/_shared/protocols/bulkhead.md) — Resource limits per worker type (memory, CPU, duration) with failure containment.
• Verification Protocol (skills/_shared/protocols/verification.md) — 4-level verification framework (Contract Compliance → Acceptance Criteria → Integration → Quality Score).
• Quality History Schema (.forgewright/quality-history.schema.json) — JSON Schema for tracking quality scores across sessions.
• Production Config Template (.production-grade.yaml.example) — Example config with circuit breaker, bulkhead, quality gate, middleware chain, and timeout settings.

Changed

• Documentation Count Fix — Mode count corrected to 23 (from 22) in README.md, README.vi.md, CLAUDE.md, AGENTS.md. Protocol count corrected to 27 (from 15) in README.md and README.vi.md badges.
• Skills Count Fix — AGENTS.md corrected to "55 AI skills" (from 54).
• AI Build Mode Simplification — Removed "scrape", "crawl website" triggers and Web Scraper skill from AI Build mode. Web Scraper has its own trigger and skill.
• DEFINE Phase — Added investigation timeout management (T0 no limit, T0.5 10min, T1 15min, T1.5 15min, T2 20min).
• TASK Contract JSON Schema (skills/_shared/protocols/task-contract.md) — Added JSON Schema block for contract validation.
• READ-ONLY Enforcement (skills/code-reviewer/SKILL.md) — Phase 0 enforcement ensuring code-reviewer never modifies source code.
• Authority Boundary Checks (skills/_shared/protocols/guardrail.md) — SOLE OWASP and SOLE SLO authority enforcement via guardrail custom rules.
• Enhanced Secret Detection (skills/security-engineer/SKILL.md) — JWT algorithm: none, AWS v4 signatures, base64-encoded secrets, env var patterns, hardcoded credentials, connection strings with embedded secrets.
• Self-Healing Loop (skills/production-grade/phases/build.md) — Error classification → retry protocol (max 3) → rollback → escrow report.

Security

• Authority boundary enforcement via guardrail custom rules (SOLE OWASP for security-engineer, SOLE SLO for SRE).
• Secret detection patterns expanded with 20+ regex patterns for API keys, JWTs, credentials, and connection strings.

Internal

• 15 phase and cross-cutting audits completed with 150+ findings. All P0 issues resolved. Breaking changes: 0.

[7.8.2] — 2026-04-09

Quality Overhaul — Stability, CI/CD, Architecture, Maintainability

Added

• Middleware Chain Extraction (skills/production-grade/middleware/) — 10 middleware files extracted from SKILL.md into separate files for maintainability: 01-session-data.md through 10-graceful-failure.md. SKILL.md reduced from 1700+ lines to ~400 lines.
• Modes Reference (skills/production-grade/modes/README.md) — Quick-reference index for all 19 modes, gate counts, skill counts, and shared behaviors.
• Architecture Decision Records (docs/adr/) — 4 ADRs documenting key decisions: orchestrator separation, KuzuDB read-only MCP, parallel pre-commit, skills count oscillation prevention.
• Health Monitoring Script (scripts/forgeNexus-health.sh) — Automated index health check, staleness detection, and auto-reindex trigger.
• CLI Integration Tests (scripts/test-cli.sh) — 16-test suite covering CLI commands, skills system, CI/CD, and version consistency.

Changed

• KuzuDB readOnly mode (forgenexus/src/data/db.ts) — MCP server now opens database in read-only mode, eliminating lock conflicts with concurrent analyze CLI runs. Multiple MCP servers across projects can run safely alongside indexing.
• Pre-commit parallelization (.husky/pre-commit) — Restructured into 2-wave architecture: ESLint + Prettier run in parallel (wave 1), TypeScript + Vitest run in parallel (wave 2). ESLint uses --cache for incremental checks. ~2x speedup.
• GitHub Actions CI parallelization (.github/workflows/ci.yml) — All jobs (lint-and-format, typecheck, test, commitlint) now run in parallel. test-coverage waits only for test. Total CI time reduced significantly.
• SKILL.md Middleware Table — Middleware chain now documented with explicit file references, eliminating duplicate inline protocol reading.
• build.md — Self-healing loop (up to 5 attempts, search-augmented, escrow report) integrated per self-healing-execution.md.
• DryRun Interceptor — Added as middleware 03b in the middleware chain table.
• ANTIGRAVITY.md references removed — Replaced with CLAUDE.md in SKILL.md and phases/sustain.md.

Fixed

• husky version (package.json) — Corrected husky: "^10.0.0" (non-existent) to husky: "^9.1.7".
• typo: forgwrightVersion (scripts/mcp-generate.sh, mcp-generator/templates/) — Fixed to forgewrightVersion in both script and Handlebars template.
• Version consistency — package.json version field bumped to 7.8.1, matching VERSION file.
• mcp-config.json version — forgewright_version updated from 7.0.0 to 7.8.1.
• MCP server config — Added forgewright-forgenexus entry to .cursor/mcp.json pointing to local forgewright repo.

Dependencies

• No changes — All dependency versions unchanged.

Impact Assessment

• Breaking changes: 0
• Backward compatibility: 100%
• Migration required: None
• User affected: Developers using pre-commit hooks (faster), MCP server users (no lock conflicts)

[7.8.1] — 2026-04-08

ForgeNexus Enterprise — Phase 1-3

Added

• ForgeNexus Enterprise Features (Phase 1-3)
  • PR Review with blast radius analysis (forgenexus pr-review <base> [head])
  • Symbol impact analysis (forgenexus impact <symbol>)
  • OpenAPI contract checking (oasdiff integration)
  • Auto Wiki generation with multi-provider LLM support
  • Auto Reindex workflow (incremental/full)
  • Multi-repo group management (group contracts, group status, group query)
  • Contract verification with oasdiff
  • Cross-repo impact analysis
• GitHub Actions Workflows (.github/workflows/)
  • pr-review.yml — PR blast radius + OpenAPI
  • auto-wiki.yml — Auto wiki generation
  • auto-reindex.yml — Incremental/full reindex
  • contract-verification.yml — OpenAPI breaking changes
  • multi-repo.yml — Multi-repo sync
  • multi-repo-impact.yml — Cross-repo impact
• Composite Actions (.github/actions/)
  • pr-review/action.yml — PR review action with dry-run
  • auto-wiki/action.yml — Wiki generation action
  • auto-reindex/action.yml — Reindex action
• Reusable Workflows (.github/workflows/reusable/)
  • pr-review.yml — Reusable PR review workflow
• Documentation
  • .github/README.md — Tổng hợp tất cả workflows
  • .github/actions/*/README.md — Hướng dẫn chi tiết
• Test Suite (scripts/test-cli.sh)
  • Test tất cả CLI commands

Impact Assessment

• Breaking changes: 0
• Backward compatibility: 100%
• Migration required: None
• User affected: Only users wanting to use Enterprise features

[7.8.0] — 2026-04-06

⚠️ Breaking Change — ForgeNexus Database Migration

ForgeNexus has migrated from better-sqlite3 to KuzuDB (graph database). Existing SQLite indexes must be migrated. Run the migration script:

node forgenexus/scripts/migrate-sqlite-to-kuzu.js [--dry-run]

Added

• ForgeNexus Groups (forgenexus/src/data/groups.ts) — Multi-repo contract tracking. Create named groups, add repos, sync contracts, query cross-repo execution flows. 8 new MCP tools: group_list, group_create, group_add, group_sync, group_contracts, group_query, group_status, group_remove. CLI: forgenexus group <list|create|add|remove|sync>.
• Claude Code Hooks (forgenexus/.claude/hooks/) — Auto-install on forgenexus setup: pre-tool-use.ts enriches grep/search context with graph data; post-tool-use.ts auto-reindexes after git commits.
• SQLite → KuzuDB Migration Script (forgenexus/scripts/migrate-sqlite-to-kuzu.js) — Automated migration for existing ForgeNexus SQLite indexes. Run with --dry-run to preview.
• README Mermaid Diagrams — 7 diagrams replacing ASCII art: Architecture Overview, Middleware Chain, Session Lifecycle, ForgeNexus Analyze Pipeline, Multi-Repo Group Management, Claude Code Hooks Flow, Request → Mode → Skills Routing.

Changed

• ForgeNexus Database Backend (forgenexus/src/data/db.ts, schema.ts, registry.ts) — Replaced better-sqlite3 with kuzu ^0.11.3. Schema migrated from single nodes/edges tables to per-type node tables (CodeNode, Community, Process) and per-edge-type rel tables (CALLS, IMPORTS, EXTENDS, etc.). KuzuDB handles FTS and vector extensions natively.
• Husky Git Hooks (.husky/) — Upgraded from v9 to v10. All hook scripts updated to use #!/usr/bin/env sh + source husky.sh. husky.sh rewritten with v10-compatible bootstrap logic.
• ESLint + Security Fixes (forgenexus/src/analysis/detect-changes.ts, prompts) — Fixed 34 ESLint errors (unused vars, prefer-const, no-useless-escape, no-extra-semi, dead code). Added security overrides for esbuild ^0.25.0 and minimatch ^9.0.5. Fixed 2 critical prompt bugs: checkStaleness variable interpolation and detect_impact template literal.
• Prettier Format — Reformatted all 32 source files.
• Root-Level Scripts (package.json) — Added build:forgenexus, test:forgenexus, lint:forgenexus, format:forgenexus, build:cli, dev:cli, typecheck:cli.

Dependencies

• Removed: better-sqlite3, @types/better-sqlite3
• Added: kuzu ^0.11.3 (graph DB with FTS + vector extensions)
• Added overrides: esbuild ^0.25.0, minimatch ^9.0.5

[7.7.1] — 2026-04-06

Added

• ForgeNexus 2.1.0 (forgenexus/) — Parallel worker threads, Leiden algorithm, binding propagation, suffix trie, incremental FTS, embedding cache, early-exit optimization, 15+ framework detection

Changed

• ForgeNexus README — Full 9-phase pipeline architecture diagram, expanded features table, performance comparison vs original pipeline, GitNexus integration docs
• VERSION — Bumped to 7.7.1
• Version badge in README, AGENTS.md

[7.7.0] — 2026-04-02

Added

• Forge-Nexus MCP Server (mcp/) — Global MCP server v1.0.0 with PipelineManager, SkillParser, PromptEngine. Enables Claude/Cursor AI assistants to interact with Forgewright pipeline across all projects. Listens on stdio via Model Context Protocol SDK.
• GitHub Actions CI/CD (.github/workflows/ci.yml) — Full pipeline: ESLint, Prettier format-check, TypeScript build, Vitest unit tests with v8 coverage, coverage threshold gate, commitlint on PRs.
• Dev tooling (mcp/) — ESLint with TypeScript ESLint plugin, Prettier formatting, Vitest test runner, commitlint for Conventional Commits, .eslintrc.json, .prettierrc, vitest.config.ts with pool: forks.
• Husky Git hooks (.husky/) — Pre-commit hook (ESLint + Prettier + TypeScript + Vitest), commit-msg hook for commitlint. Root package.json with prepare: husky install for auto-initialization on clone.
• Global Setup Script (scripts/setup-project.sh) — Links Forgewright to any project without git submodule. Detects tech stack, runs ForgeNexus analyze, prints Cursor MCP config snippet.
• GitHub Templates (.github/) — Issue templates (bug report, feature request), PR template with Forgewright pipeline checklist.
• Cursor Subagent Review Workflow (.cursor/agents/) — 5 specialized subagents: quality-reviewer, security-auditor, spec-reviewer, verifier, chat-interpreter for structured code review.
• ForgeNexus MCP Tools (via MCP server) — forgenexus_query, forgenexus_context, forgenexus_impact, forgenexus_detect_changes, forgenexus_rename, forgenexus_cypher for code intelligence.
• Auto-initialization Check — MCP server and ForgeNexus index auto-initialize on session start if .forgewright/mcp-server/mcp-config.json is missing.
• Antigravity Plugin (.antigravity/) — Production-grade plugin system with 52 skills, 15 shared protocols, preset templates, game dev workflows.

Changed

• Skill count — 48 skills (corrected after oscillation 52 → 53 → 52)
• README — Added Mermaid flow diagrams for pipeline, subagent review, parallel dispatch; version badge bumped to 7.7.0
• VERSION — Bumped to 7.7.0
• AGENTS.md — Updated skill count to 52, skill directory table
• CI pipeline — Skills updated with activation rules, parallel-dispatch with Mermaid diagrams

Fixed

• .vite build cache ignored in .gitignore
• mcp/node_modules/ tracked artifacts removed (added to .gitignore)
• ForgeNexus code intelligence block updated with npx forgenexus analyze refresh note

[7.0.0] — 2026-03-14

Added — New Protocols (5)

• Project Onboarding (project-onboarding.md) — 5-phase deep project analysis: fingerprint → health check → pattern analysis → risk assessment → profile generation. Produces .forgewright/project-profile.json and .forgewright/code-conventions.md.
• Session Lifecycle (session-lifecycle.md) — Cross-session continuity with start/save/end hooks. Resume interrupted sessions, detect drift, memory integration.
• Quality Gate (quality-gate.md) — Universal per-skill validation: 4 levels (Build, Regression, Standards, Traceability), 0-100 quality scoring, configurable thresholds. Works in sequential AND parallel modes.
• Brownfield Safety (brownfield-safety.md) — Safety net: auto git branching, baseline snapshots, protected paths, change manifest, regression checks, rollback.
• Quality Dashboard (quality-dashboard.md) — Real-time tracking, final dashboard, machine-readable JSON reports, cross-session trending, early warning system.

Added — Workflow

• /onboard — Run deep project analysis without starting pipeline.

Added — Project State

• .forgewright/ directory for persistent project state (profile, conventions, session logs, quality reports).

Changed — Orchestrator

• Session lifecycle pre-flight (Step 0.5) — Loads project profile, session state, memory context, quality trends before work begins.
• Enhanced codebase discovery — Deep onboarding replaces shallow scanning for brownfield. Learns patterns, conventions, risk.
• Context-aware routing — Suggests addressing health issues before building (failing tests, CVEs, tech debt).
• Quality gate integration — Runs after EVERY skill. Mini-scorecard + aggregate at gates.
• Brownfield safety net — Auto-activates for all brownfield projects in any mode.
• Session lifecycle hooks — PHASE_COMPLETE, TASK_COMPLETE, GATE_DECISION, ERROR hooks throughout.
• Quality Dashboard final summary — Replaces legacy text banner with comprehensive report.
• 5 new common mistakes added.

Changed — Phase Dispatchers

• BUILD phase — Quality gate & regression checks after each task. Change manifest. Brownfield regression at completion.
• HARDEN phase — Quality gate per task, quality scoring in summary, brownfield merge readiness check.

Changed — Memory Manager

• Active lifecycle hooks — Orchestrator now calls memory at 6 lifecycle points (was spec-only).
• Context integration — Memory + project profile provide full context without re-scanning.

Changed — Metadata

• Skill count — 46 skills unchanged
• Protocol count — 7 → 12 (added 5 new protocols)
• Version — 6.1.0 → 7.0.0

[5.2.0] — 2026-03-10

Added

• Parallel dispatch — Git worktree-based parallel task execution. Independent tasks (BUILD: T3a/T3b/T3c, HARDEN: T5/T6a/T6b) run in isolated worktrees simultaneously.
• Task Contract protocol — JSON-based contract defining exact input/output/constraints for each parallel worker. Workers can only read listed files and write to allowed directories.
• Task Validator protocol — 7-step post-execution validation: contract compliance, boundary violations, forbidden patterns, build check, test check, import verification, API/schema conformance.
• Merge Arbiter protocol — Ordered merge strategy with auto-resolution for configs (package.json, docker-compose.yml), integration testing, and per-branch rollback.
• Worktree Manager script — Shell script managing git worktree lifecycle: create, status, validate, merge, cleanup, resume.
• Scope Analysis Engine — Analyzes project complexity (5-factor weighted score), estimates execution time for both modes, assesses 6 risk categories, and generates data-driven recommendation.
• Anti-hallucination pipeline — 6-layer defense: contract boundaries → forbidden pattern grep → build verification → test execution → import path resolution → API/schema conformance.
• init-parallel setup command — Convenience command that symlinks worktree-manager and updates .gitignore.
• Checkpoint & resume — Failed parallel workers preserve state in worktree for retry from checkpoint.

Changed

• Skill count — 17 → 18 (added parallel-dispatch orchestrator skill).
• Protocol count — 4 → 7 (added task-contract, task-validator, merge-arbiter).
• Execution strategy — Orchestrator now presents scope analysis with complexity scoring, time estimates, and risk predictions before asking user to choose parallel or sequential.
• BUILD phase — Supports parallel mode via parallel-dispatch skill dispatch.
• HARDEN phase — Supports parallel mode with strict authority boundary enforcement in contracts.
• AGENTS.md — Updated with parallel-dispatch skill and new protocol references.

[5.0.0] — 2026-03-06

Changed

• Migrated to Antigravity — complete platform migration from Claude Code to Antigravity.
• All Claude Code-specific APIs replaced with Antigravity equivalents:
  • AskUserQuestion → notify_user with markdown numbered options
  • Agent() parallel sub-agents → sequential skill execution
  • Skill() invocation → direct SKILL.md reading and instruction following
  • TeamCreate/TaskCreate/TaskUpdate/TeamDelete → task.md tracking + task_boundary
  • WebSearch/WebFetch → search_web/read_url_content
  • Read/Write/Edit/Glob/Grep → view_file/write_to_file/replace_file_content/find_by_name/grep_search
  • smart_outline/smart_unfold/smart_search → view_file_outline/view_code_item/grep_search
  • Bash → run_command
• Workspace directory renamed: Claude-Production-Grade-Suite/ → Antigravity-Production-Grade-Suite/
• Config file reference: CLAUDE.md → ANTIGRAVITY.md
• Skill installation simplified: skills load directly from skills/ directory, no marketplace needed
• Skill Maker simplified: removed plugin packaging (Phase 3) and marketplace registration (Phase 5), skills install directly to skills/ directory
• Parallel execution removed — Antigravity executes skills sequentially. Architecture quality (14 specialized skills, authority hierarchies, approval gates) is fully preserved.
• Author updated to antigravity-code
• All 50+ markdown files across 14 skills updated with new tool references and branding

[4.2.0] — 2026-03-06

Added

• Adaptive routing — orchestrator now analyzes the user's request and routes to the right skills automatically. No longer requires full pipeline for every task.
• 10 execution modes: Full Build, Feature, Harden, Ship, Test, Review, Architect, Document, Explore, Optimize, Custom. Each with appropriate skill composition, gates, and parallelism.
• Request classification — automatic intent detection maps user requests to modes. "Add auth to my API" → Feature mode (PM + Architect + Backend + QA). "Review my code" → Review mode (Code Reviewer only).
• Execution plan presentation — user sees which skills will run and can adjust, escalate to full pipeline, or proceed.
• Custom mode — multi-select skill menu for requests that don't fit standard patterns.
• Lightweight mode execution — non-Full-Build modes skip unnecessary overhead (engagement/parallelism prompts only for 3+ skill modes).

Changed

• Plugin description broadened from "build a complete production-ready system" to "any software engineering work that benefits from structured, production-quality execution."
• "When to Use" expanded to cover: adding features, hardening, deploying, testing, reviewing, documenting, optimizing, exploring — not just greenfield builds.
• Full Build pipeline preserved unchanged as one mode within the adaptive orchestrator.

[4.1.0] — 2026-03-05

Added

• Engagement modes — 4-level interaction depth (Express, Standard, Thorough, Meticulous) chosen at pipeline start.
• Architecture Fitness Function — Solution Architect now DERIVES architecture from constraints instead of picking templates.
• Scale & Fitness Interview — Adaptive 1-4 round interview (depth scales with engagement mode).
• Adaptive PM interview — Express: 2-3 questions. Standard: 3-5. Thorough: 5-8. Meticulous: 8-12.

Changed

• Engagement mode propagated to ALL 14 skills — every skill reads settings.md and adapts decision surfacing.
• Solution Architect Phase 1 rewritten from 5 shallow questions to a comprehensive adaptive discovery process.
• Product Manager Phase 1 rewritten with 4 interview depth profiles matching engagement modes.
• Software Engineer parallelism revised — shared foundations established SEQUENTIALLY before parallel service agents.
• Frontend Engineer parallelism revised — UI Primitives built SEQUENTIALLY first, then Layout + Features parallel.

[4.0.0] — 2026-03-05

Changed

• Two-wave parallel execution — orchestrator splits work into Wave A (build + analysis) and Wave B (execution against code).
• Internal skill parallelism — 8 skills now spawn parallel Agents for independent work units.
• Dynamic task generation — orchestrator reads architecture output and creates tasks accordingly.

Added

• Parallelism preference — user selects performance mode at pipeline start.
• Token economics — parallel execution is both faster AND cheaper.

[3.3.0] — 2026-03-05

Added

• Brownfield awareness — orchestrator detects greenfield vs existing codebase at startup.
• Codebase discovery — parallel scan of project root for package.json, go.mod, etc.

Changed

• MECE intent-based skill routing — all 14 skill descriptions rewritten from keyword triggers to intent descriptions.

Fixed

• Protocol loading crash — added || true fallback.
• Polymath priority — uncertainty expressions now correctly route to polymath.

[3.2.0] — 2026-03-05

Added

• Auto-update with consent — orchestrator checks for new versions on pipeline start.

Fixed

• Protocol loading crash — added || true fallback to all cat commands.
• MECE intent-based skill routing — replaced keyword trigger matching with intent descriptions.

[3.1.0] — 2026-03-05

Added

• Polymath co-pilot — the 14th skill. Thinks with you before, during, and after the pipeline.
• 6 Polymath modes: onboard, research, ideate, advise, translate, synthesize.

[3.0.0] — 2026-03-04

Changed

• Full rewrite — shared protocols, 4 protocols, sole-authority conflict resolution.
• Large skills split into router + on-demand phases for 65% token savings.

Added

• Phase-based skill splitting for 7 skills.
• Conditional task execution: frontend auto-skip, data-scientist auto-detect.
• Partial execution: "just define", "just build", "just harden", "just ship", "just document".

[2.0.0] — 2026-03-04

Changed

• Bundle all 13 skills into a single plugin install.
• Unified workspace architecture.
• Prescriptive UX Protocol enforced across all skills.

Added

• Skill Maker as pipeline phase.
• VISION.md: ten principles governing the ecosystem.

[1.0.0] — 2026-03-03

Added

• Initial release: production-grade orchestrator plugin.
• 12 specialized agent skills coordinated through dependency graph.
• 3 approval gates, autonomous execution between gates.
• DEFINE > BUILD > HARDEN > SHIP > SUSTAIN pipeline.