DavidZapataOh/overdrive

Overdrive is a Claude Code plugin that scales the amount of process to the risk of each change. A deterministic classifier assigns every task to one of three lanes: direct for trivial edits, standard for ordinary work and critical for changes that touch data, auth, money, secrets or public contracts. Each lane brings only the planning, review and evidence it needs, so small changes stay fast and risky ones are never merged under-reviewed.
The plugin is 14 skills, 5 subagents, three hooks (session router, tool-call guard, stop check) and a small shell CLI, overdrive. Nothing runs outside Claude Code, and the runtime needs only Bash 3.2, git and the POSIX tools that ship with them.
## The three lanes
| Lane | Planning | Review | Evidence |
| direct | none; test first if testable | none; reread the diff and run the change's own check | the change's own test or command, run in the turn |
| standard | short decision plan with ### Task N: headings | one overdrive:reviewer per slice, before any pull request | test-driven where it applies; overdrive evidence green before the slice package |
| critical | full test-first plan that the user approves before the critical part is touched | overdrive:critical-reviewer after every task and after every change to it, then one overdrive:final-reviewer on the slice | pinned tests, red and green evidence, hard-case list, non-regression check; each task built by a separate overdrive:builder |
The classifier (overdrive risk) picks the lane from facts, not from size. It reads the paths a change touches, the lines it adds and removes, and, before any edit, the words of the request. Built-in floors put a change in critical for irreversible data and migrations, auth, tenant isolation, money, secrets, CI and delivery, public contracts, concurrency, a changed risk map, a removed or skipped test, a text file git shows as binary and a source file that no test covers. Dependency and test-configuration changes, deleted files and files made executable put it in standard. A change that raises nothing is direct.
Claude may raise a lane when it sees risk the classifier cannot, and it never lowers one. A failed classification is critical. Only the user can lower a lane, by answering the stated risk, and the answer is recorded as a decision. Signals, floors and the steps of each lane are in docs/lanes.md.
## Install
Installation is the same on macOS, Linux and Windows, through Claude Code's plugin commands. The repository is private for now, so authenticate with GitHub first:
sh
gh auth login && gh auth setup-git
export CLAUDE_CODE_PLUGIN_PREFER_HTTPS=1
claude plugin marketplace add DavidZapataOh/overdrive
claude plugin install overdrive@overdrive
### One-command setup on a new Mac
Paste this block once into Terminal. It uses the GitHub CLI (gh) if it is installed and installs it with Homebrew if it is not. Homebrew may or may not be on the machine: without it the block stops and points to for gh. If any step fails, the block stops there; fix that step and paste the block again. It assumes the claude command is already installed.
sh
{ command -v gh >/dev/null 2>&1 || { command -v brew >/dev/null 2>&1 && brew install gh; }; } \