v2.0.2

Overdrive v2.0.2

Focused skill-intake release on top of v2.0.1. No ovd-plan state semantics changed.

Added

• Added Addy Osmani's pinned engineering-method skills:
  interview-me, doubt-driven-development, source-driven-development,
  api-and-interface-design, code-simplification,
  documentation-and-adrs, performance-optimization,
  test-driven-development, and debugging-and-error-recovery.
• Added Matt Pocock's pinned grill-me and grilling skills.
• Updated Muratcan Koylan's context-engineering source and added
  harness-engineering and self-improvement-loops.
• Expanded the official Anthropic skills source with algorithmic-art,
  canvas-design, claude-api, skill-creator,
  web-artifacts-builder, webapp-testing, and the official
  theme-factory.
• Added four Overdrive-authored wrapper skills:
  brag-video, autoresearch-harness, clone-website-guide, and
  fact-checker.

Safety Notes

• Raw latent-spaces/brag is not copied because no clear redistribution
  license was visible during review; Overdrive ships an original
  brag-video wrapper instead.
• karpathy/autoresearch is treated as inspiration for
  autoresearch-harness; Overdrive does not install dependencies, start
  long loops, or spend GPU/API budget without explicit user approval.
• JCodesMore/ai-website-cloner-template is treated as a fresh-template
  workflow, not a global skill. clone-website-guide requires
  public/authorized targets and rejects phishing, impersonation, credential
  harvesting, and trademark misuse.
• hesreallyhim/awesome-claude-code remains research/link-only because it is
  a curated list rather than a portable skill source and its license is not a
  fit for redistribution.
• The Reddit-linked fact-checker path was unavailable, so Overdrive ships an
  original fact-checker skill.

Verification

• Manifest target: 160 unique skills: 23 local, 136 pinned upstream GitHub,
  and 1 installer-backed skill.
• Router smoke and benchmark coverage now includes the new skill families and
  conflict rules.
• Package and archive checks continue to exclude .overdrive/, .DS_Store,
  tarballs, old zips, secrets, auth state, local runtime state, and
  unsupported vendored third-party material.

v2.0.1

Overdrive v2.0.1

Docs and branding patch. No installer, CLI, or ovd-plan behavior changed from v2.0.0.

Why This Patch Exists

v2.0.0 shipped correct but unpolished: an early README draft with placeholder-feeling copy, diagrams with a blue-tinted background and gradient panels instead of the logo's flat black-and-silver palette, and a couple of stale links. This patch brings the published package in line with what the project actually looks and reads like today.

Highlights

• Rebuilt all three README diagrams to match the Overdrive logo exactly: flat black backgrounds and content boxes (no gradients), a flat silver border on every box, and a real cinematic lens flare (bright core, crisp streak, soft glow) on the headline wordmarks instead of a faint line.
• New hero tagline and rewritten intro paragraph describing the actual pipeline: global instructions built on Karpathy's system prompt, the 137-skill catalog, the skill router, and the project state management system. Dropped inaccurate "operating layer" phrasing.
• Added a paragraph on who Overdrive is for: simple to install, runs quietly in the background, ovd commands kept deliberately small, and designed to build on top of what modern models already do well rather than override it.
• Added small icon links (X, LinkedIn, GitHub) near the top of the README for easy access to Stefan's profiles.
• Named ovd-workflow and ovd-plan consistently as the project state management system across the README and docs, replacing inconsistent "project-state layer" / "project-management layer" phrasing.
• Added a "Why leaf?" explanation of ovd-plan node naming to the README and docs/ovd-plan-v2.md: a leaf is the smallest implementable unit, deliberately not called a task, step, phase, or action, so the plan tree can take whatever shape a project needs.
• Added a skills-provenance section: 22 pinned upstream open-source sources plus the 19 skills built specifically for Overdrive, counted directly from manifest.json. Added a docs/README.md index grouping the rest of the documentation by what a reader is trying to do.
• Removed em dashes and other AI-tell phrasing from public-facing prose.

Credits

• ovd-workflow, the project state management system, was designed by Eugen Bulboaca (GitHub).
• The installed global operating guide is based on Andrej Karpathy's Claude system prompt, adapted to fit Overdrive, via multica-ai/andrej-karpathy-skills.

Fixes

• Two documentation links pointed at an internal handoff/ directory that was never shipped in the public repo; both now point at the real design record or a plain-text note.
• The README's Safety/Updates section had dropped its uninstall command block during the v2.0.0 README rework; restored.
• The Claude Code plugin manifests (.claude-plugin/marketplace.json, plugins/overdrive/.claude-plugin/plugin.json) still referenced a retired X handle; updated to the current one.

Verification

• npm run consistency (1176 checks), npm run check, npm run test:smoke, npm run test:workflow, npm run eval:router
• npm pack --dry-run
• Diagram renders visually compared against the logo asset side by side; every gradient reference in the diagram SVGs confirmed to resolve (no missing defs).

Overdrive v2.0.0

Overdrive v2.0.0

Overdrive v2 adds a project state management system on top of the existing skill installer and router.

Highlights

• Added ovd-plan: a structured project tree in OVERDRIVE.md with executable leaves, success criteria, dependencies, verification, and recursive closure.
• Added the main v2 workflow commands:
  • /ovd-workflow for setup, codebase mapping, requirements, preferences, and decisions.
  • /ovd-plan for deliberate planning, research, editing, and plan-quality checks.
  • /ovd-go for orienting and executing active work.
  • /ovd-log for progress capture, concerns, handoffs, and closure.
• Added overdrive verify --plan to audit project-state layout and plan health.
• Reworked the README around a first-time-reader flow: what Overdrive is, install, what changes after install, the v2 mental model, common workflows, safety/privacy, and advanced internals.
• Moved design-extract to a local compatibility skill so no upstream repository is required at install time. The previously pinned upstream Git source was unreachable during v2 pin verification; upstream distribution has since moved to the designlang npm package.
• Added js-yaml as a runtime dependency for ovd-plan YAML parsing and hardened runtime dependency copying for npm/npx layouts.
• Expanded CI and local tests for the v2 project-management pipeline.
• Rebuilt the README and all three diagrams around the v2 architecture, and reorganized secondary docs under docs/ for a cleaner repository root.

Audit Fixes

An independent pre-release audit (Claude Fable 5) reviewed installer safety, v2 state semantics, CLI surfaces, packaging, licensing, and release readiness. Fixes landed in this release:

• Hook setup no longer rewrites a malformed existing agent settings file (~/.claude/settings.json, ~/.gemini/settings.json, ~/.codex/hooks.json). Install and uninstall now abort with a clear message and leave the file byte-identical, covered by a new regression test.
• ./update.sh no longer aborts on stock macOS bash 3.2 (set -u + empty argument arrays).
• Uninstall removes the empty ~/.overdrive/ scaffolding directories it created.
• Stale "upstream unavailable" wording for design-extract was corrected across docs and notices.

Credits

• ovd-workflow, the project state management system, was designed by Eugen Bulboaca (GitHub).
• The installed global operating guide is based on Andrej Karpathy's Claude system prompt, adapted to fit Overdrive, via multica-ai/andrej-karpathy-skills.

Safety Notes

• No telemetry.
• No API keys, OAuth state, browser profiles, cookies, MCP secrets, .env files, or account sessions are copied.
• Installs remain non-destructive by default and use Overdrive-managed markers.
• A malformed existing settings file aborts hook setup instead of being replaced.
• Existing v1 .overdrive/ layouts are detected and require explicit migration choices.
• The release does not claim measured output-quality gains; router/scorecard tooling is included for validation, not marketing proof.

Verification

Release checks passed on the final release candidate:

• npm run test:full (syntax, consistency, router benchmark, workflow, ovd-plan, smoke)
• ./verify.sh
• npm audit --omit=dev (0 vulnerabilities)
• npm pack --dry-run + package-content scans
• local/global install dry-runs and real install → use → update → uninstall round-trips in a clean environment
• tracked artifact, secret-pattern, and personal-path scans (hidden files included)

Overdrive v1.0.2

v1.0.2 - 2026-06-06

• Removed pre-launch compatibility surfaces so the public package uses only canonical Overdrive names: overdrive, ovd, overdrive-cli, .overdrive/, .overdrive.json, OVERDRIVE_WORKFLOW, /ovd-*, and ovd-workflow.
• Moved Stefan's README note to a top quote, added collapsible README sections for the installed global guide and skill-router, and removed redundant public draft files.
• Strengthened package and repository checks for tracked junk, local runtime state, obsolete brand surfaces, archive exclusions, and canonical install/uninstall behavior.

Overdrive v1.0.1

Overdrive v1.0.1

Small launch hotfix for the thin Claude Code plugin wrapper.

Fixed

• Fixed Claude Code marketplace/plugin manifests so claude plugin validate, marketplace add, and plugin install pass with Claude's current plugin schema.
• Added frontmatter metadata to the /ovd-* helper commands included in the Claude plugin wrapper.

Notes

• Core Overdrive skill catalog and installer behavior are unchanged from v1.0.0.
• npm package: overdrive-cli@1.0.1.

Overdrive v1.0.0

Overdrive v1.0.0 — an open-source, Mac-first, plug-and-play system that supercharges AI coding agents.

Install once across Claude Code, Codex, Gemini CLI, Antigravity, Cursor, and shared .agents. You get:

• 137 curated specialist skills for design, animation, web/app development, SEO, research, security, launch, browser QA, and more.
• A selective skill-router that loads the minimum useful skills per task and logs its choices.
• ovd-workflow — local, gitignored .overdrive/ project memory for state, decisions, constraints, routes, knowledge vaults, and handoffs.
• A Karpathy-based operating guide that keeps the agent thinking-first, objective, and verification-oriented.

Pinned, non-destructive, reversible. Apache-2.0 for Overdrive's original code; bundled third-party skills retain their own licenses and notices.

Upgrading from AgenticSupercharge: legacy markers, /as-* commands, and .agenticsupercharge/ state still work and migrate automatically to .overdrive/.

No performance benchmarks are claimed yet. A measured scorecard is planned for a v1.x follow-up.

v0.11.0 — Safe optional tool setup

• Changed the optional dependency policy from "document only" to "attempt safe setup during install, then fail open": Graphify, Claude Video/media helpers, and Design Extract browser support now get non-privileged installer attempts unless --no-tool-install is set.
• Added installer support for graphifyy==0.1.14 through pipx or a managed user-space virtualenv, never global pip, sudo, or --break-system-packages.
• Added safe helper setup paths for yt-dlp, ffmpeg, and Design Extract browser support while preserving runtime preflights, manual fallback commands, and no API-key collection.
• Updated Graphify router/instruction guidance to prefer an existing project graph for codebase relationship questions, recommend Graphify's own watch/git-hook freshness flow when stale, and avoid any AgenticSupercharge background graph daemon.
• Strengthened global objectivity guidance against sycophancy and weak-premise feedback loops, including proactive challenge for consequential, ambiguous, or irreversible decisions.
• Updated docs, third-party notices, readiness checks, router catalog, and consistency tests for the v0.11 plug-and-play dependency policy.

AgenticSupercharge v0.10.0

v0.10.0 - 2026-06-02

• Added four pinned MIT upstream skills: prompt-master, humanizer, design-extract, and claude-video, raising the manifest to 136 unique skills.
• Added safety transforms for design-extract and claude-video so agents check availability, ask before setup, avoid writing secrets, and never auto-install packages, browsers, Homebrew/apt/winget/pip tooling, or MCPs.
• Added an ethics note to humanizer so humanizing preserves meaning and facts without deceptive authorship, fake lived experience, or removing required AI disclosure.
• Added agentic-supercharge usage, /as-usage, and workflow tests for local, read-only token usage reporting from Claude Code JSONL logs with best-effort AS route attribution and no prompt/message content printing.
• Updated skill-router, smoke checks, router benchmark cases, docs, third-party notices, verified sources, skill readiness, and consistency checks for the v0.10 manifest and new routing boundaries.

AgenticSupercharge v0.9.0 - Graphify Code Intelligence

v0.9.0 - 2026-05-31

• Added Safi Shamsi's MIT Graphify source pinned at 91f4d120b630ee35c79bf3c75ccd186870a808f9, importing one optional graphify code/corpus intelligence skill.
• Added installer support for upstream skills whose source file is lowercase skill.md, normalizing to installed SKILL.md.
• Added a Graphify-specific safety transform that removes upstream Python auto-install behavior, preserves optional setup guidance for pip install graphifyy / pip install 'graphifyy[pdf]', and teaches agents to fall back to normal repo exploration when Graphify is unavailable.
• Updated skill-router, smoke checks, router benchmark cases, docs, third-party notices, verified sources, skill readiness, and consistency checks for the 132-skill v0.9 manifest.

AgenticSupercharge v0.8.0 - Context efficiency and prompt caching

v0.8.0 - 2026-05-31

• Added prompt-cache-friendly guidance and docs/prompt-caching.md while keeping caching controlled by the host runtime/harness.
• Added AS-Workflow preferences.md for lightweight per-project do-not rules and durable user preferences, with init/doctor/test coverage.
• Strengthened global instructions for lean context, escalating ~60% / ~75% / ~85-90% context-budget heads-ups, native plan-mode versus clarify-and-plan, and native orchestration instead of custom orchestrators.
• Kept the catalog at 131 unique skills: v0.8 adds no new skills and keeps context skills router-selectable rather than always-on.
• Updated README examples, router/catalog guidance, attribution, and consistency checks for the context-efficiency release.