Example: Vibecheck Assembly Line

A user submits a GitHub repository URL (the Task Card). Eight specialised Claude Code Agents work in sequence, each writing a structured report to a shared stage git repository. The SYNTHESIS Agent aggregates all reports into a scored Vibecheck Report, which is reviewed by a human before the DELIVER Agent publishes it to the Agentics artifact store.

INTAKE → ARCHITECTURE → CODE QUALITY → SECURITY → TESTING → DOCUMENTATION
                                                                    ↓
DELIVER ← [Gate: HUMAN REVIEW] ← SYNTHESIS ─────────────────────────

Title:   Vibecheck: my-startup/backend

Description:
  Repository:  https://github.com/my-startup/backend
  Branch:      main
  Stack:       Node.js + PostgreSQL   (optional — agent auto-detects)
  Context:     About to open-source this. Want a thorough review first.
  Concerns:    Auth implementation, SQL injection risks

Each Agent's prompt.md receives {{task.title}} and {{task.description}} at dispatch time.

The .agentics/ folder is split across several files — there is no single monolithic config.

{
  "id": "vibecheck",
  "name": "Vibecheck",
  "description": "Comprehensive AI-powered code review pipeline — 8 specialist stations covering intake, architecture, code quality, security, testing, documentation, synthesis, and delivery.",
  "version": "1.0.0"
}

[
  { "from": "intake",        "to": "architecture",  "condition": "success" },
  { "from": "architecture",  "to": "code-quality",  "condition": "success" },
  { "from": "code-quality",  "to": "security",      "condition": "success" },
  { "from": "security",      "to": "testing",       "condition": "success" },
  { "from": "testing",       "to": "documentation", "condition": "success" },
  { "from": "documentation", "to": "synthesis",     "condition": "success" },
  { "from": "synthesis",     "to": "deliver",       "condition": "success", "gate": "human-review" }
]

[
  {
    "id": "human-review",
    "name": "Human Review",
    "description": "A human reviewer checks the compiled Vibecheck Report (SYNTHESIS output) before the DELIVER agent posts results to GitHub. Approve to publish, reject to stop delivery.",
    "requiresApproval": true
  }
]

{
  "id": "intake",
  "name": "INTAKE",
  "trigger": {
    "type": "dispatch_worker",
    "labels": ["vibecheck"],
    "idleTimeoutMinutes": 5,
    "maxTimeoutMinutes": 25,
    "createStageRepo": true
  }
}

Conduct intake analysis for this Vibecheck submission.

Project: {{task.title}}
{{task.description}}

Your current working directory is the repository being reviewed — it was cloned by the runner before you started. Do not commit or modify files here.

Clone the stage git repository to $STAGE_DIR:

git -c "http.extraHeader=Authorization: Bearer $STAGE_GIT_TOKEN" clone "$STAGE_GIT_URL" "$STAGE_DIR"
git -C "$STAGE_DIR" config http.extraHeader "Authorization: Bearer $STAGE_GIT_TOKEN"

Write your report to $STAGE_DIR/reports/intake.md, then commit and push:

git -C "$STAGE_DIR" add reports/
git -C "$STAGE_DIR" commit -m "feat(intake): initial analysis"
git -C "$STAGE_DIR" push origin main

Explore the repository (your CWD) and write reports/intake.md covering:

• Tech stack and versions detected
• Project type and purpose
• Key entry points and architecture overview
• Dependency summary
• Size metrics (file count, LOC by language)
• Any immediate red flags
• Recommended focus areas for downstream stations

**`system.md`** — appended to the agent's system prompt to define its role:

```markdown
You are a Technical Intelligence Officer specialising in rapid codebase comprehension.
Your role is to deeply understand a submitted project and produce a precise, structured brief
that downstream agents will rely on. Be factual. Cite specific files and line counts.
If something is unclear or undocumented, say so explicitly — never assume.
Set up the stage workspace, write your brief, commit it, then call stop_broadcast.

All 8 stations follow this same three-file pattern. The only differences are station.json timeouts and the content of prompt.md / system.md.

At INTAKE, the platform creates a fresh stage git repository for this run. Every subsequent agent clones it, writes a report, and pushes. The git history is a complete, ordered audit trail of the review.

The runner pre-clones the submitted repository into each worker's CWD — agents do not need to fetch the target repo themselves.

INTAKE:         CWD = submitted repo (pre-cloned by runner)
                clone stage repo → $STAGE_DIR
                writes $STAGE_DIR/reports/intake.md
                git push

ARCHITECTURE:   CWD = submitted repo
                clone stage repo → $STAGE_DIR
                reads  $STAGE_DIR/reports/intake.md
                writes $STAGE_DIR/reports/architecture.md
                git push

SECURITY:       CWD = submitted repo
                clone stage repo → $STAGE_DIR
                reads  $STAGE_DIR/reports/intake.md + architecture.md
                writes $STAGE_DIR/reports/security.md
                git push

... (CODE QUALITY, TESTING, DOCUMENTATION) ...

SYNTHESIS:      clone stage repo → $STAGE_DIR
                reads  $STAGE_DIR/reports/*.md
                writes $STAGE_DIR/vibecheck-report.md
                git push

DELIVER:        clone stage repo → $STAGE_DIR
                reads  $STAGE_DIR/vibecheck-report.md
                POSTs  artifact to Agentics API
                writes $STAGE_DIR/reports/delivery-result.md
                git push

SYNTHESIS produces a weighted composite score across five dimensions:

Each dimension is scored 0–10. A total Vibe Score below 6.0 / 10 is a failing grade.

DELIVER does not post a GitHub comment. It POSTs a structured artifact to the Agentics REST API:

POST ${AGENTICS_BASE_URL}/api/owners/${AGENTICS_OWNER}/projects/${AGENTICS_PROJECT_NAME}/artifacts
Authorization: Bearer $AGENTICS_TOKEN
Content-Type: application/json

{
  "type": "vibecheck",
  "name": "Vibecheck — <project>",
  "data": {
    "vibeScore": 7.4,
    "categories": {
      "security":      { "score": 8 },
      "codeQuality":   { "score": 7 },
      "architecture":  { "score": 7 },
      "testing":       { "score": 6 },
      "documentation": { "score": 9 }
    },
    "reportMarkdown": "..."
  }
}

The response includes an id used to construct the public artifact URL:

https://agentics.dk/project/{owner}/{project}/artifacts/{id}

This page displays the Vibe Score as a diploma-style number with per-category score bars and the full report in a collapsible section.

The repo ships a reusable skill at plugin/skills/deliver-vibecheck/SKILL.md. This is an example of how assembly lines can bundle custom skills that agents load at runtime — the DELIVER station's prompt.md references it directly:

Follow the deliver-vibecheck skill (see plugin/skills/deliver-vibecheck/SKILL.md)

This pattern keeps complex multi-step instructions out of the prompt and makes them reusable across assembly lines.