Skip to main content
Agent mode reads your GitHub backlog, selects the highest-priority issue, generates a spec, runs the full Ralph loop, reviews the diff against the spec, and auto-merges when checks pass — then moves to the next issue. Zero intervention required.

​
How it works

1

Read backlog

Fetches open GitHub issues and ranks them by priority labels (P0 > P1 > P2) and dependency order.
2

Assess feature state

For each issue, determines the right action: start fresh, generate a plan, resume an existing implementation, skip if already shipped, or comment on an existing PR.
3

Generate spec

Creates an implementation-ready spec from the issue context (title, body, labels) — same quality as wiggum new.
4

Run Ralph loop

Executes the full 5-phase Ralph loop: planning → implementation → E2E testing → verification → PR review.
5

Review and merge

Reviews the diff against the spec. In merge mode, auto-merges the PR when all checks pass, then moves to the next issue.

​
Flags

​
string
Comma-separated GitHub labels to filter issues. Only issues with at least one matching label are processed.
​
string
Comma-separated issue numbers to process. Restricts agent to specific issues instead of the full backlog.
​
number
Stop after completing this many issues.
​
number
Stop after this many orchestrator steps (tool calls). Safety limit to prevent runaway execution.
​
string
default:"manual"
Controls what happens after loop completion:
  • manual — Creates PR, stops for human review
  • auto — Auto-reviews the diff, creates PR
  • merge — Auto-reviews, creates PR, and auto-merges when checks pass
​
string
Override the AI model for the orchestrator.
​
boolean
default:"false"
Simulate the full workflow without executing loops or creating PRs. Useful for testing issue selection and spec generation.
​
boolean
default:"false"
Force headless streaming output. Skips the TUI even in interactive terminals.
​
boolean
default:"false"
Run GitHub connectivity checks to debug authentication or permission issues.

​
Routing behavior

​
Examples

​
Feature state assessment

Before processing each issue, the agent assesses the current state and decides the right action: This prevents duplicate work and ensures the agent picks up where it left off after interruptions.