Autonomous AI Engineering
with Verification Gates
Supervise coding agents with absolute verification gates. Prevent broken commits and unauthorized edits with worktree isolation, automated rollbacks, and an open audit log.
Get up and running in minutes
One-line installers for macOS, Linux, and Windows. The installer clones PolyLever, installs bun if needed, and prints next steps. Safe to re-run.
curl -fsSL https://raw.githubusercontent.com/lerugray/generalstaff/master/install.sh | bash
After install
Run gs welcome for guided setup
Register your first project
Run one verified cycle end-to-end
Requires git, bash (Git Bash on Windows), bun 1.1+, and an LLM provider (Anthropic, Claude Code, OpenRouter, or Ollama).
AI orchestration
with safety built in
PolyLever treats agentic AI as adversarial input to your codebase. Structure beats better prompts — verification gates, worktree isolation, and open audit logs enforce discipline on every cycle.
Every cycle must pass tests, produce a non-empty diff, and survive independent review before a commit lands.
Verification Gates
Every cycle must pass tests, produce a non-empty diff, and survive independent review before a commit lands.
From bootstrap to verified commit
Every autonomous cycle flows through the same enforced pipeline. The dispatcher orchestrates agents, the verification gate decides outcomes, and failures roll back automatically.
Bootstrap Project
Initialize a project configuration with hands-off rules and verification commands via gs welcome.
Assign Task
Queue a work order via session, cycle, or heartbeat inbox — the dispatcher picks up the task.
Initialising agent workspace... done
Engineer Agent Executes
The engineer runs in an isolated worktree, making changes against declared scope only.
Verification Gate Runs
Tests must pass and the diff must be non-empty. Boolean gates enforced in code, not prompts.
Reviewer Agent Validates
Independent reviewer confirms scope match and checks every diff against hands-off patterns.
Commit or Rollback
Verified cycles commit to bot/work. Failures roll back with a full audit trail in PROGRESS.jsonl.
Everything you need to ship safely
From first-run wizard to advanced heartbeat dispatch — comprehensive guides for running autonomous AI workflows with verification discipline.
Traditional AI Coding vs PolyLever
Autonomous agents fail predictably when structure is replaced by instructions. PolyLever encodes discipline as code — gates the model cannot bypass.
Prompt-Based
- Agent says the task is done
- Tests may fail after completion
- Empty diffs can still look successful
- Protected files rely on instructions
- Review happens after the damage
- Logs are partial or hidden
Gate-Enforced
- Task is not done until checks pass
- Tests, diff, and scope are verified every cycle
- Hands-off files are blocked by rule
- Work runs in .bot-worktree on bot/work
- Failed cycles roll back automatically
- Prompts, responses, tool calls, and diffs land in PROGRESS.jsonl
See the gate in action
Watch how PolyLever orchestrates agents, enforces verification, and maintains a complete audit trail — cinematic presentation of real workflow patterns.
Cycle Execution
Dispatcher orchestrates engineer → verification → reviewer with session budgets and structured SITREPs.
Orchestrate with absolute control
Start with gs welcome, register a project, and run your first verified cycle. Your codebase deserves gates, not just prompts.









