v3.4.0 — Compound Learning Playbook & 18 Upstream Open-Source Benchmarks

BoostX v3.4.0 Release Notes

Highlights

• Compound Learning & Bottleneck Playbook (docs/COMPOUND_LEARNING_PLAYBOOK.md): Public-ready engineering playbook cataloging 17 Defect & Bottleneck Archetypes (PAT-VEC-01..07, PAT-COR-01..05, PAT-CON-01..05), 7 Maintainer Governance Fingerprints (GOV-01..07), 4 False-Positive Mirage Guardrails (FALS-01..04), Two-Tier Autonomy (Tier A Autonomous vs. Tier B Report-First), and Bounded Up-Tree Federation (/xbrief <-> /xsync --federated).
• Classic Academic Engineering Workflow Diagram: Plain monochrome white-mode Mermaid lowchart TD visualizing the 6-phase closed-loop execution and ratification architecture.
• 18 Upstream Open-Source Contributions (Batch 0 + Group A + Group B): Full empirical benchmark receipts (4.0x to 1,735.6x speedups) across libpysal, obler, momepy, osmnx, opengeos/geoai, weecology/DeepForest, ydataai/ydata-profiling, os-climate/physrisk, ladybug-tools/uwg, Deltares/hydromt, CLIMADA-project/climada_python, and
  atcap/invest.
• Up-Tree Federated Doc Sync (xsync/scripts/audit_workspace_docs.py, 462 LOC): Added --federated parent-umbrella boundary scanning (5,458 total CLI LOC across 8 deterministic Python CLI engines).

BoostX v3.3.0 - Context-Conditioned Universal Laws, Mode C 5-Gate Falsification & 30-Point Audit Fixes

BoostX v3.3.0 Highlights

• Context-Conditioned Universal Laws (UL-01..UL-08) and 5-Gate Universal Law Falsification Test (Mode C) in /xcreate-skill.
• Wave 2.5 Graveyard and Anti-Pattern Check across /xdeep-research, /xboost, and /xplan.
• 30-Point Full-Fleet Rule and Script Audit Remediation (verify_evidence.py, ocr_runner.py, laya_triage.py, audit_engine.py, audit_workspace_docs.py).
• YAML Trigger Budget: 371.29 +/- 14.84 chars across all 14 /x* skills (100% within <= 400-char target).
• 8 Bundled Python CLI Engines (5,058 LOC) and 600-Node Obsidian Second Brain parity.

v3.2.0 — Composable /x* Agent Skills Suite (13 Skills) & Unified /xdeep-research

BoostX v3.2.0 — Prefix /x* Namespace Migration & Unified /xdeep-research (13 Composable Skills)

• Standardized /x Prefix Namespace (14 → 13 Composable Skills)*:
  • Migrated every skill in the BoostX ecosystem from the -x suffix to the x- prefix so typing /x in any IDE or CLI slash-command menu immediately surfaces all 13 BoostX skills in a single contiguous block without memorizing base names:
    • /xboost, /xbrief, /xcontext, /xcreate-skill, /xdeep-research, /xkickoff, /xlaya, /xobsidian, /xplan, /xpublish-audit, /xpublish-qgis, /xsync, /xultra
  • All 13 SKILL.md YAML frontmatter descriptions preserve backward-compatible legacy triggers (/*x and unprefixed aliases) while maintaining 100% compliance (393.08 ± 8.18 chars) within the ≤ 400-char �udit_engine.py --platform skill context budget guard.
• Unified /xdeep-research (Retiring Standalone
  esearchx):
  • Merged
    esearchx's 3-Tier Adaptive Scope Routing (Tier 1: Fast Inline Probe ≤ 2 sources, Tier 2: Single Isolated Scout, Tier 3: 5-Wave Adversarial Swarm), 4-Tier Primary-Source Authority Ladder, and Capability-Aware Subagent Write-Permission Contract directly into /xdeep-research (and /xboost Track B).
• Automated Legacy Directory Cleanup & Empirical Re-Benchmarking:
  • Updated scripts/install.ps1 and scripts/install.sh to automatically purge deprecated *x folders from ~/.gemini/config/skills, ~/.claude/skills, and ~/.grok/skills during upgrades.
  • Re-ran the full multi-trial benchmark harness across 109 installed skills, 752 markdown files, 25 red-team defects (100% catch rate in 6.45 ms), and 214 real IDE sessions (987 prompt turns).

v3.1.0 — 14 Composable Agent Skills (/ultrax Swarm Multiplier & /contextx Cross-Platform Telemetry)

• /ultrax (13 → 14 Composable Skills) — Dynamic Topology-Matched Multi-Agent Swarm & Blind Verification Multiplier:
  • Order-Independent Canonical Stage DAG: Normalizes any slash-command combination regardless of typing order into Stage 0: RETRIEVE (/briefx, /obsidianx, /layax) → Stage 1: RESEARCH (/researchx, /deep-researchx) → Stage 2: PLAN (/planx) → Stage 3: BUILD (/boostx) → Stage 4: AUDIT & SYNC (/syncx, /publishx-audit) → Stage 5: RATCHET (/obsidianx) → Stage 6: VERIFY (/goal).
  • Dynamic N × K Topology Sizer (Zero Fixed-Count Bloat):
    • Complexity Floor (2-Agent Fast Pair): Automatically bypasses swarm fan-out on localized 1-file tasks (1 Maker + 1 Blind Verifier).
    • Archetype 1 (3 + 2 Best-of-3 Tournament): Spawns 3 orthogonal Makers (Vector A: Stdlib/Minimal, Vector B: Performance/Algorithmic, Vector C: Deep Module/UX) + 2 Blind Verifiers (Adversarial Spec Critic + Red-Team Edge-Case Breaker). Supports Objective Auto-Promote vs. Aesthetic Visual QA Side-by-Side (_vA/_vB/_vC) modes.
    • Archetype 2 (N + 2 Partitioned Batch Swarm): Enforces a strict ≤ 3 items/files per subagent cap (N = ceil(M / 3), max 6 concurrent per wave) and locks a Pre-Locked Shared Seam Header in the main thread before fan-out to prevent cross-partition interface drift.
    • Archetype 3 (Multi-Wave DAG): Chains Scout Wave → Builder Wave → Blind Audit Wave for full-stack cross-layer features.
  • Universal 5-Domain Adaptation, Multi-Host Dispatch & Build-Lock Serialization:
    • Adapts seamlessly across Software/Web/Cloud, Data/ML/SQL, CAD/GIS/Physics, Presentation/UI/Docs, and General Research/Strategy across Antigravity (invoke_subagent), Claude Code (Task), and Sequential Personas.
    • Enforces Wave 1 Isolation (write-only to partitioned paths or Workspace="share"/"branch", zero concurrent dotnet build/pytest/cargo lock contention) followed by Wave 2 Serialized Build & Red-Team Verification and ≤ 150-Word Caveman Receipt returns.
• /contextx Cross-Platform Token Telemetry Upgrade (scripts/check_context.py):
  • Added pure-stdlib Protobuf wire decoding for SQLite conversations/<id>.db step metadata alongside transcript.jsonl parsing, surfacing active-session token counts (input_tokens, cache_read_tokens, output_tokens, thinking_tokens, cache_hit_pct) and cross-platform aggregation flags (--tokens, --days <N>, --all-time) across Antigravity 2.0, Antigravity IDE, and Antigravity CLI.
• Zero-Persona Neutrality & Bidirectional Ecosystem Wiring:
  • Added opt-in /ultrax composable hooks across /boostx, /planx, /briefx, /researchx, /deep-researchx, /syncx, /obsidianx, and /kickoffx.
  • Sanitized personal hardware/persona references across /boostx, /create-skillx, and /kickoffx, and added automated persona-neutrality checks to /create-skillx and publishx-audit/scripts/audit_engine.py.

Install

See README. Tag: v3.1.0.

git clone --branch v3.1.0 https://github.com/iscvlab/boostx.git

BoostX v3.0.0 — 13-Skill Composable Agent Ecosystem & 7 CLI Engines

• Unified BoostX (-x) Composable Ecosystem & 13-Skill Suite Expansion:
  • Standardized -x Command Namespace: Unified all core workflow skills under the -x suffix while preserving backward-compatible legacy triggers in YAML descriptions:
    • /brief → /briefx
    • /obsidian-vault → /obsidianx
    • /update-doc → /syncx
    • /plan → /planx
    • /context → /contextx
    • /kickoff → /kickoffx
    • /laya → /layax
    • /create-skill → /create-skillx
    • /publish-audit → /publishx-audit
    • /publish-qgis → /publishx-qgis
    • Retained /boostx, /researchx, and /deep-researchx.
  • 5 Newly Added Skills in Public Distribution (8 → 13 Skills):
    • /obsidianx (NEW): Portable 2-Hop Obsidian Second Brain & Cross-Domain Knowledge Graph Architect. Features zero-hardcode dynamic vault resolution via OS obsidian.json (open: true) or ancestor .obsidian/ discovery, strict opt-in execution (zero vault I/O when /obsidianx is omitted from the prompt), 2-Hop navigation (00 Index.md → Hub - <Domain>.md → Atomic Note), Cross-Domain Concept Bridges, and a 4-check Post-Ratchet Graph Verifier.
    • /planx (NEW — BoostX v2.0 Architecture Planner): Locks a 2D Plan Contract (Scope × Risk), runs blast-radius & invariant discovery, evaluates Ponytail × Deep Module Vectors A/B/C, embeds an Inline Grilling Decision Matrix (DEC-01..N) and Red → Green Verification Table inside implementation_plan.md with zero chat Q&A spam, and halts for user approval.
    • /layax (NEW — System 1 Semantic Decision & Triage Sidecar): Non-autoregressive System 1 decision layer (choice, score, noul) powered by convaiinnovations/laya (~7–33 ms local forward pass). Accelerates /syncx /layax contradiction hunting, /briefx /layax large-corpus section shortlisting, /deep-researchx /layax candidate triage, and /publishx-audit /layax public claim verification, with 100% graceful fallback to native tools when laya is not installed.
    • /contextx (NEW): IDE conversation transcript compaction analyzer (scripts/check_context.py) reporting compaction round, total session steps, and 4-tier memory drift risk (HEALTHY, NOMINAL, CAUTION, CRITICAL).
    • /kickoffx (NEW): End-of-session handover protocol enforcing pre-handover /syncx verification, task-smart command routing, strategic next-step options, and a self-contained copy-pasteable kickoff prompt for fresh conversations.
  • 6th Bundled Zero-Dependency CLI Engine (syncx/scripts/audit_workspace_docs.py):
    • Added repo-agnostic workspace documentation auditor verifying relative links, Zero-LaTeX compliance, SHA-256 mirror parity, and --deprecated sweeps.
  • Composable Pipeline Wiring & Visual Architecture Diagrams:
    • Wired left-to-right composable handoffs (Context → Research → Plan/Execute → Sync + Modifiers/Sidecars) across /briefx, /researchx, /deep-researchx, /planx, /boostx, /syncx, /obsidianx, and /layax.
    • Added vector architecture diagrams (docs/assets/boostx-master-ecosystem.svg and docs/assets/boostx-composable-routing.svg).

Install

See README. Tag: v3.0.0.

git clone --branch v3.0.0 https://github.com/iscvlab/boostx.git

v2.0.0 - Core 8 Portable Agent Skills Suite

• Core 8 Portable Skill Suite Expansion:
  • /deep-researchx (NEW): Universal, domain-agnostic 5-Wave Multi-Agent Deep Research & 100%-Proof Synthesis Engine:
    • Wave 1 (Intent × Domain 2D Contract): Dynamically synthesizes evaluation axes, quotas, and disambiguation rules across Built Environment/Case Studies, Physical Sciences/Simulations, Software/Systems, Natural Sciences, and Policy/Strategy.
    • Wave 2 (Parallel Harvester Swarm): Concurrent subagent fan-out with mandatory primary deep-link scraping (banning bare homepage URLs and snippet-only claims).
    • Wave 3 & 3.5 (Blind Adversarial Auditor + scripts/verify_evidence.py + Prune & Replace Loop): Independent red-team falsification of completion vs policy years, attributions, units, and metrics, paired with a deterministic Python CLI verifier (verify_evidence.py) and zero-forgery replacement harvesting.
    • Wave 4 & 5 (100%-Proof Synthesis & Composable Handoff): Strict [VERIFIED FACT] vs [RECOMMENDED SYNTHESIS] binding, dual Markdown + JSON/SQLite persistence, and direct spec-to-code handoff into /boostx and /goal.
  • /researchx (NEW): Fast tactical primary-source research protocol:
    • 3-Tier Execution Routing: Inline probe (≤ 2 sources) vs single isolated scout vs parallel multi-angle swarm.
    • 4-Tier Primary-Source Authority Ladder: Local pinned manifests (package.json, pyproject.toml, *.csproj, metadata.txt) and local SQLite/XML/AST probes first, version-matched official specs second, upstream source/tests third, standards/papers fourth.
    • Capability-Aware Subagent Routing: Prevents read-only subagent write deadlocks and enforces ≤ 150-word compressed main-thread return digests.
  • /boostx (v2.0 Upgrade):
    • 2D Task × Project Goal Contract Synthesis: Crosses 6 Task Archetypes (DEBUG/RCA, BUILD/FEATURE, SIMULATE/SOLVER, REFACTOR/SIMPLIFY, OPTIMIZE/SCALE, AUDIT/PLAN/SHIP) with 5 Project Verification Harnesses (CAD/Rhino/GH, GIS/QGIS, Web UI/Figma, Data/GCP/ML, Agent Skills/CLI).
    • Simulation Physics Triad: Enforces explicit CAD-to-SI unit bridges, CFL/timestep caps, zero-load identity checks (zero NaN/Inf), and 1D analytical toy benchmarks.
    • Composable Escalation: Native hooks for /goal marathon state checkpointing, /researchx tactical lookups, and /deep-researchx 5-wave formulation.
  • /create-skill & /publish-audit (--platform skill Globalization & Context Budget Guard):
    • Added --platform skill profile to audit_engine.py verifying frontmatter schema, ≤ 350-char description budget (preventing system-prompt truncation of sibling skills), Universal Globalization placement, POSIX link integrity, and sys.dont_write_bytecode hygiene.
  • /brief & /update-doc Upgrades:
    • /brief: Added 4-tier authority hierarchy, Two-Pass reading (Pass A structural skeleton/cardinality before Pass B section deep dive), and deferred/frozen zone gates.
    • /update-doc: Added 4-level Conflict Arbitration Hierarchy (Code/Data Receipts > Canonical SSOT > Chat Lock > Drafts), ghost reference sweeps, quantitative cardinality sweeps, and Sync Ledger output.

Install

See README. Tag: v2.0.0.

git clone --branch v2.0.0 https://github.com/napattan/agent-skills.git

v1.1.0 add /brief

/brief is now in the suite. Read related workspace markdown (cap 8), then do /brief <task> in the same turn. HTML only when the task is a specific page.

Also: /create-skill machine-path ban uses C:\Users\<user> so /publish-audit --platform github does not treat ellipsis placeholders as a real home path.

Pin: git clone --branch v1.1.0 https://github.com/napattan/agent-skills.git

v1.0.0 Portable skills

v1.0.0 Portable skills

First tagged public line of this suite. Same SKILL.md files on Claude Code, Grok, Gemini/Antigravity, Cursor, Codex, and similar loaders. IDE or CLI. Windows, macOS, and Linux.

Sister repo: mcp-bridges (QGIS, Rhino, Grasshopper, Illustrator, Figma).

---

Why this release exists

Earlier untagged main mixed host-specific APIs, a thesis-only doc auditor, and a /create-skill flow that wrote stubs or Gemini-only paths. v1.0.0 is the portable, publishable line: no one-machine paths, no baked-in project filenames, no required vendor tool names.

---

Skills in this tag

Skill	Command	What it does	What it stops
BoostX	/boostx	Red gate, Ponytail short-diff ladder, dual-OS edits, green verification. Task classes: code (failing command), docs (stale-claim grep), audit (engine exit code). Uses host subagents/background tools if they exist; never requires a vendor API name.	Speculative patches; fake unit tests on doc work; Antigravity-only instructions on other hosts
Create Skill	/create-skill	Discovers skill roots (project .agents / .grok or user ~/.claude / ~/.grok / ~/.gemini/config). Drafts YAML description, waits for approval, writes a full SKILL.md. Dual-OS directory create. No machine-user paths.	Thin drafts; wrong folder; Gemini-only or Grok-only write paths
Update Doc	/update-doc	(1) Ground-truth delta from this chat. (2) Scan markdown and HTML in the current workspace. (3) Follow relations (symbols, metrics, status words, links). (4) 5-tier inner-to-outer sync. (5) Re-grep deprecated terms to zero. Optional: your own scripts/audit_docs.py.	Stale claims; hardcoded file lists from someone else's repo
Publish Audit	/publish-audit	python scripts/audit_engine.py <dir> --platform generic|qgis|food4rhino|github. Secrets (masked), machine paths, .env / .ps1 / .sh / .toml. GitHub missing LICENSE = blocking. Food4Rhino missing .gha/.dll = blocking. README Published / LIVE / 1-click = advisory.	Credential leaks; unescaped % in QGIS metadata; claiming PyPI/NPM without a profile
Publish QGIS	/publish-qgis	SemVer, metadata interpolation, single-root zip, Qt6 / QGIS 4 AST check, package slug ladder. Run after /publish-audit is green.	Plugin portal zip rejections; Qt6 deprecations

Public status words (owned by /update-doc): LIVE, REGISTERED, PREPARED, IN PIPELINE, DRAFT.

---

Layout

agent-skills/
├── README.md
├── CHANGELOG.md
├── LICENSE
├── skills/
│   ├── boostx/SKILL.md
│   ├── create-skill/SKILL.md
│   ├── update-doc/SKILL.md
│   ├── publish-audit/SKILL.md + scripts/audit_engine.py
│   └── publish-qgis/SKILL.md + scripts/check_qt6.py, package_qgis.py
└── scripts/install.sh, install.ps1

---

Install

Clone skills/ into your host skill root (see README table). Or:

git clone --branch v1.0.0 https://github.com/napattan/agent-skills.git ~/.claude/skills
git clone --branch v1.0.0 https://github.com/napattan/agent-skills.git ~/.grok/skills
git clone --branch v1.0.0 https://github.com/napattan/agent-skills.git .agents/skills

Windows: scripts/install.ps1. POSIX: scripts/install.sh.

---

Not in this package

The old thesis HTML auditor (audit_workspace_docs.py with fixed layer counts) is not shipped. If your own repo needs extra checks, add scripts/audit_docs.py there. /update-doc will run it when present.

---

Related

• mcp-bridges
• Profile: github.com/napattan