Adversarial Pipeline

Adversarial Pipeline

DISABLED: This skill chains into hostile_reviewer (Stage 2 gate), which is currently disabled pending eval framework validation. Do NOT invoke this skill. It will produce incorrect results (silent-pass gate). Re-enable when the eval framework validation closes.

Overview

Three-stage autonomous pipeline that catches wrong-approach patterns before tickets are created. Chains design_to_plan, hostile_reviewer, and plan_to_tickets as background agents. The adversarial gate blocks ticket creation until the plan has been stress-tested.

Anti-patterns caught: ONEX not Docker, typed Pydantic not strings, contracts not hardcoding, OAuth not SSO, topics from contract.yaml not topics.py. See plugins/onex/prompts/adversarial-rubric.md for the full rubric.

Stages

Stage 1 — Design Agent (background)

Dispatch a background agent to run design_to_plan on the provided topic.

/onex:dispatch_worker role=designer
  Invoke /design_to_plan --topic "<topic>" --no-launch
  Save the output plan path to checkpoint key: adversarial_pipeline.plan_path
  Report: plan file path saved to checkpoint

Wait for Stage 1 agent to complete. Read adversarial_pipeline.plan_path from checkpoint before proceeding.

Stage 2 — Adversarial Gate (background)

Dispatch a background agent to run hostile_reviewer --static on the plan output.

/onex:dispatch_worker role=adversarial-reviewer
  Read plan path from checkpoint key: adversarial_pipeline.plan_path
  Invoke /hostile_reviewer --pr <PR_number> --repo <owner/repo>
  # Note: --static mode does not accept --file or --rubric; use PR mode for plan review.
  Count findings in reviewer output.
  IF findings < 3:
    ESCALATE to user: "Adversarial gate: fewer than 3 issues found — plan may be under-scrutinized. Attach reviewer output and pause pipeline."
    DO NOT proceed to Stage 3.
  ELSE:
    Save findings summary to checkpoint key: adversarial_pipeline.findings
    Save plan path to checkpoint key: adversarial_pipeline.gated_plan_path
    Report: gate passed, N findings recorded

Wait for Stage 2 agent to complete. Read adversarial_pipeline.findings and adversarial_pipeline.gated_plan_path from checkpoint. If gate did not pass (escalation fired), surface findings to user and stop — do not proceed to Stage 3.

Stage 3 — Ticket Creator (background)

Only dispatched after Stage 2 gate passes.

/onex:dispatch_worker role=ticket-creator
  Read plan path from checkpoint key: adversarial_pipeline.gated_plan_path
  Invoke /plan_to_tickets --plan-file <plan_path> [--project <project>] [--dry-run if --dry-run set]
  Report: Linear epic URL and ticket count

Execution Rules

• Every stage runs as a background agent via dispatch_worker. No foreground implementation.
• Stage 3 is gated — it MUST NOT run if Stage 2 escalated.
• The pipeline uses /onex:checkpoint to pass state between stages. Never pass paths via inline text substitution.
• If any stage fails (non-escalation error), surface the error to user and stop.
• Do not reimplement design_to_plan, hostile_reviewer, or plan_to_tickets logic here. Invoke the skills.

Invocation

/adversarial-pipeline "design a cross-repo dependency analyzer"
/adversarial-pipeline "design a real-time alerting system" --project "Platform"
/adversarial-pipeline "design a unified auth layer" --dry-run

Acceptance Criteria

• End-to-end run produces a Linear epic.
• Adversarial stage catches ≥1 ONEX pattern violation in a test run.
• All stages run as background agents.
• Rubric file exists at plugins/onex/prompts/adversarial-rubric.md.