flow-next-pilot

Advances one ready spec through one pipeline stage per tick and emits a terminal PILOT_VERDICT line.

692|55|Updated Dec 26, 2025
One-click install
npx skills add https://github.com/gmickel/gmickel-claude-marketplace --skill flow-next-pilot
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: flow-next-pilot
Source: https://github.com/gmickel/gmickel-claude-marketplace/tree/main/plugins/flow-next/codex/skills/flow-next-pilot
Command: npx skills add https://github.com/gmickel/gmickel-claude-marketplace --skill flow-next-pilot

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

Autonomous build loops drift, re-run finished work, or silently stall when no conductor enforces a strict one-spec-one-stage contract. This Skill conducts a single tick of the flow-next pipeline: it selects one ready spec, classifies its current stage, dispatches exactly one stage skill, verifies state advanced, and ends with a parseable PILOT_VERDICT line that host loop primitives like /loop or /goal can drive.

Core Features & Use Cases

  • Single-tick stage conductor: Selects one ready spec with dependency, claim-collision, and strike-ledger checks, then classifies it into plan, plan-review, work, qa, or make-pr and dispatches exactly one stage skill.
  • Verdict contract for drivers: Every tick ends with one terminal PILOT_VERDICT line (ADVANCED, NO_WORK, DEFERRED_TO_LAND, BLOCKED, NEEDS_HUMAN) so /goal and /loop stop clauses can route or halt unattended runs.
  • Optional backlog mode: When pilot.autonomy is set to backlog, it widens selection to tracker-promoted issues, triages items by agent judgment, and parks gaps as async questions without ever merging or authoring specs.
  • Use Case: Point a Claude Code /goal loop at your backlog with the instruction to keep running pilot until it prints PILOT_VERDICT=NO_WORK, and each tick advances one spec one stage toward a draft PR.

Quick Start

Ask the agent to run the flow-next pilot on your ready specs, optionally with a review backend such as --review=codex, and let it emit its PILOT_VERDICT after one tick.

Frequently Asked Questions about flow-next-pilot

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run an autonomous build loop over ready specs?

Invoke the pilot skill once per tick, or wrap it in a host loop primitive such as /loop in Claude Code or /goal in Claude Code or Codex. Each tick selects one ready spec, dispatches one stage skill, and ends with a PILOT_VERDICT line the driver can use as a stop condition.

What does the PILOT_VERDICT line mean?

PILOT_VERDICT is the single terminal line every tick prints, with values ADVANCED, NO_WORK, DEFERRED_TO_LAND, BLOCKED, or NEEDS_HUMAN plus the spec id and stage. Drivers grep it to decide whether to continue looping, route to land, or stop for human input.

Can pilot run unattended with a tracker like Linear or GitLab?

Yes, when pilot.autonomy is set to backlog or the --backlog flag is passed, pilot unions tracker-promoted issues into selection via tracker-sync read ops. It supports Linear, GitHub, GitLab, and Jira adapters and never merges or authors specs in this mode.

Why does pilot refuse to run with a dirty working tree?

Pilot hard-stops with NEEDS_HUMAN when git status shows changes outside the .flow directory at tick start. This guard prevents the conductor from dispatching stage work on top of uncommitted state it cannot attribute or safely verify.

What is the difference between pilot and Ralph in flow-next?

Pilot and Ralph are alternative autonomous drivers that must never be nested. Ralph is an external shell loop with receipt plumbing, while pilot is an in-session single-tick conductor designed for host loop primitives; pilot exits immediately if FLOW_RALPH or REVIEW_RECEIPT_PATH is set.

When does pilot stop and ask for a human?

Pilot emits NEEDS_HUMAN for crash-class conditions such as a dirty tree, a failed gh probe at the all-done branch, a stale in-progress claim, or a NOT_RETRYABLE review verdict. It never asks interactive questions mid-tick; ambiguity always maps to a terminal verdict.