dev-pipeline

Multi-agent dev pipeline plugin. Turns a Jira ticket into a reviewed, committed PR via a chain of specialized subagents. Memory-backed project registry for monorepos and multi-stack environments.

Pipeline

pipeline JIRA-1234
     ↓
[memory] — load project path, stack, quirks
     ↓
context-router  (haiku)  — picks context packs per agent
     ↓
pm              (haiku)  — ticket → structured spec + acceptance criteria
     ↓
architect       (sonnet) — spec + repo grep → implementation plan
     ↓
developer       (sonnet) — plan → branch + commits
     ↓
reviewer        (sonnet) — diff + acceptance criteria → verdict
     ↓
PR + Jira update + memory update

Setup

1. Required environment variables

export GITHUB_TOKEN=ghp_...
export JIRA_URL=https://yourcompany.atlassian.net
export JIRA_USERNAME=you@company.com
export JIRA_TOKEN=...
export CONFLUENCE_URL=https://yourcompany.atlassian.net
export CONFLUENCE_USERNAME=you@company.com
export CONFLUENCE_TOKEN=...   # same API token as JIRA_TOKEN if same Atlassian account

Add these to your shell profile (~/.zshrc or ~/.bashrc).

2. Memory storage

Memory persists automatically at ~/.dev-pipeline/memory.json. Create the directory:

mkdir -p ~/.dev-pipeline

3. Register your first project

Once the plugin is installed, say:

pipeline remember project

Claude will ask for alias, path, stack, Jira project key, and Confluence space key.

Usage

Run the pipeline

pipeline JIRA-1234
pipeline JIRA-1234 in acme-web

Manage projects

pipeline list projects
pipeline remember project
pipeline forget acme-web
pipeline update project acme-web

What gets stored in memory

• Project paths and aliases
• Stack details (turborepo shape, package manager, test command, known stacks)
• Context pack mappings per project
• Per-run learnings (hallucination patterns, deviations, verdict history)

Confluence as context source

The pipeline pulls ADRs, specs, and checklists directly from Confluence — no manual file management. The context-router searches the project's Confluence space using ticket keywords and routes relevant pages to each agent.

To make a Confluence page discoverable by the pipeline, add labels to it in Confluence:

• adr or architecture → routed to architect
• spec or product → routed to pm
• security or checklist → routed to reviewer
• convention or style → routed to developer

The context-router also searches by text match, so labeled pages rank higher but any page mentioning the ticket's keywords can be picked up.

Confluence space key is stored per project in memory. Use pipeline remember project to set it, or pipeline update project <alias> to change it.

Hooks (repo-level)

The file-write allowlist hook and token logger are repo-specific — add them to your target repo's .claude/settings.json. Example hooks are in the original dev-pipeline scaffold.

Cost estimate

~$0.20–0.60 per medium ticket (vs. $2–4 for a single Sonnet agent with full context).