stepgate 0.1.0

A step-gated micro-change protocol CLI for coding agents: propose, approve, execute, verify — one small step at a time.

stepgate

A step-gated micro-change protocol for coding agents.

Coding agents (Claude Code, Codex, …) working on large or loosely-scoped tasks tend to mix contexts, silently expand scope, touch files unrelated to the original request, or declare success without real evidence. stepgate makes the correct workflow — investigate, propose, wait for approval, execute only what was approved, verify, suggest the next step — the single natural path of action, so that deviating from it is visible and recorded, never silent.

stepgate is not an enforcement tool. It never blocks your code, your editor, your commits, or git. You can always edit anything, commit anything, and cancel any agent session at any time. What it gives you is structure and an honest, append-only trail of what was proposed, approved, executed, and verified — across every agent working on the repo, even concurrently.

Install

pipx install stepgate   # recommended: one global install per machine
# or: pip install stepgate

Then, at the root of any project:

stepgate init

This creates .stepgate/ (state + append-only history) and injects an instruction block into AGENTS.md/CLAUDE.md telling agents to use the protocol. It is idempotent and never overwrites anything you wrote in those files — only the text between its own markers is ever touched. If your project already has a domain guardrails document, point to it with --guardrails GUARDRAILS.md and the generated block will reference it.

The protocol

Every micro-change is a proposal moving through a state machine, validated by the CLI itself:

PENDING ──approve──► APPROVED ──exec-log──► EXECUTED ──verify──► VERIFIED ──close──► CLOSED
   │
   └─reject──► REJECTED          (any non-terminal state) ──abandon──► ABANDONED

A proposal covers six points, each written as natural flowing prose (not a telegraphic form): what will change now, why this step comes first, where (files/contracts/flows touched), how it will be implemented, the expected result, and the verification that will demonstrate it.

Key rule: a micro-change reduces the scope of execution, never the depth of investigation — the agent still investigates everything it needs to understand before proposing, even if it will only implement a small piece.

A real cycle

You ask your agent to fix a race condition. It investigates, then proposes:

stepgate propose --agent claude --file plan.json

You read it rendered as prose, and approve with a tweak:

stepgate show
stepgate approve --adjust --note "rename the function to apply_sanity_loss_atomic"

The agent implements only what was approved, logs it (a git diff --stat is captured automatically as objective evidence), and verifies:

stepgate exec-log --summary "atomic decrement migration + typing" --files "migrations/013.sql,types.ts"
stepgate verify --evidence "type-check ok, simulated concurrency test passed"

You close the cycle; the agent suggests — but does not start — the next step:

stepgate close
stepgate next --suggest "wire SalaJogo.tsx to the new function via supabase.rpc(...)"

The suggestion stays visible in stepgate status until a new proposal is opened. The full trail lives in stepgate history — chronological, across all sessions and agents, append-only.

Commands

Command	Purpose
stepgate init	Create .stepgate/, inject the agent instruction block
stepgate propose --agent X --file plan.json	Register a micro-change plan (PENDING)
stepgate show	Render the active proposal as readable prose
stepgate approve [--adjust --scope ... --note ...]	Approve (optionally with reduced scope)
stepgate reject --note "..."	Reject a pending proposal
stepgate exec-log --summary "..." --files "..."	Record execution (+ automatic git diff --stat)
stepgate verify --evidence "..."	Record verification evidence
stepgate close	Close a verified micro-change
stepgate abandon --reason "..."	Cleanly abandon from any non-terminal state
stepgate next --suggest "..."	Record a next-step suggestion (opens nothing)
stepgate status	Current session + aggregated project view
stepgate history [--session X] [--since DATE]	Append-only, cross-session log
stepgate doctor	Report corrupted/invalid state files (fixes nothing)

Multiple agents can work concurrently: each session has its own state file (.stepgate/sessions/claude-2026-07-09-1.json — human-readable names, not hashes), writes are file-locked, and propose/status warn — informationally, never blockingly — when two active proposals touch the same files.

Design principles

• Never block. No git hooks, no file locks on your code, no commit gates. The file lock in .stepgate/ protects only stepgate's own state files.
• Never act silently. No auto-repair, no auto-expiry of stale sessions, no rewriting of history. doctor diagnoses; a human decides.
• Non-interactive by design. No prompts or confirmations, so it behaves identically in a terminal, a desktop app, or an IDE side-panel extension.
• Structure when you opt in. Editing code without a proposal is a legitimate flow, not an error. The state machine gives shape to the agent workflow — it is never a requirement for the code to change.

License

MIT © Leo Costta