Pipeline

Skill akashgit/remote-factory/skills/pipeline

Design and execute a custom multi-agent pipeline for any goal. Analyzes the goal, selects appropriate specialist agents, designs a DAG of steps with dependencies, and executes them via factory CLI with gate decisions between steps. Use when the user says 'run a pipeline for X', 'orchestrate X', or wants a custom multi-agent workflow.From its SKILL.md

Install

npx -y skills add akashgit/remote-factory --skill pipeline

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

3.6 KB, as published. Nobody here has run it

Pipeline — Dynamic Multi-Agent Orchestrator

You design and execute custom multi-agent pipelines to accomplish the user's goal.

The user wants: $ARGUMENTS

Prerequisites

The factory CLI must be installed:

command -v factory >/dev/null 2>&1 || uv tool install "${CLAUDE_PLUGIN_ROOT}"
mkdir -p .factory/pipeline

Your Agents

Spawn specialists via the CLI. Each agent gets a fresh context window.

factory agent <role> --task "<task description>" --project "$(pwd)" [--timeout N]

Role	Purpose
researcher	Web research, codebase analysis, domain studies
strategist	Generate prioritized hypotheses from observations
builder	Implement code changes on a feature branch, open PRs
qa	Health checks, code review, adversarial QA verification
archivist	Record findings to `.factory/archive/`
distiller	Refine vague ideas into buildable specs

Invocation Rules

Each factory agent call is synchronous and blocking — it returns only when the agent finishes. Do not shell-background (&) individual commands.

To run steps in parallel, issue multiple factory agent commands as separate bash tool calls in the same message turn. Claude Code executes them concurrently. This is parallel tool calls, not shell backgrounding.

The runner captures agent stdout to .factory/reviews/<role>-latest.md.

Phase 1: Design the Pipeline

Understand the goal — what outcome is desired? Which agents are needed?

Inspect project state:

factory detect "$(pwd)"
cat .factory/config.json 2>/dev/null

Write the pipeline plan to .factory/pipeline/plan.md:

## Pipeline: <goal summary>

### Steps

| Step | Role | Task Summary | Depends On |
|------|------|-------------|-----------|
| S1 | researcher | ... | - |
| S2 | qa | ... | - |
| S3 | strategist | ... | S1, S2 |
| ... | ... | ... | ... |

### Gate Rules
- After S1: PROCEED if ...; REDIRECT if ...
- After S3: PROCEED if ...; ABORT if ...

Design Principles

Minimize invocations — only agents needed for this goal
Maximize parallelism — steps whose dependencies are all satisfied and that don't depend on each other can be issued as parallel tool calls
Mandatory archival — always include at least one archivist step at the end
Gate rules — define PROCEED/REDIRECT/ABORT criteria for critical transitions

Phase 2: Execute the Pipeline

Process steps in topological order:

Identify next batch — steps whose dependencies are all complete
Build task strings — incorporate output from prior steps by reading .factory/reviews/<role>-latest.md
Invoke agents — single or parallel batch
Read output — cat .factory/reviews/<role>-latest.md
Apply gate rule:
- PROCEED: Move to next step
- REDIRECT: Re-invoke with corrections (max 2 per step)
- ABORT: Skip downstream steps, jump to summary
Repeat until done

Error Recovery

Agent timeout: retry once with shorter scope
Agent failure: check output, decide REDIRECT or ABORT
2 consecutive failures: ABORT pipeline

Final Summary

Write .factory/pipeline/summary.md with goal, status, step results, and key findings.