Repair Orchestrator

Repair Orchestrator

The escalating repair pipeline. Driven by quality.json findings.

Inputs

• workspace/{task_id}/quality.json (one or more gates failed)
• workspace/{task_id}/review.json (if independent-reviewer ran)
• state.repair_iteration (current attempt count; cap = 4)

5 escalation levels

Level 1 — SURGICAL (cheapest, fastest)
  Edit specific lines indicated by quality.json.repairs[].instruction
  Use Edit tool only (preserve everything else)
  If quality improves ≥3 score → progress; loop back at Level 1 again
  If improvement  3:
            return 3  # Stage-level fix
        return 2  # Try again
    
    # Stage rewrite didn't help → full regen
    if prior_level == 3:
        return 4
    
    # Full regen didn't help → from scratch
    if prior_level == 4:
        return 5

What each level fixes

Level 1: Surgical — best for

• Banned word hits
• Em-dash hits
• Citation marker → APA replacement
• Minor wording polish

Level 2: Section-rewrite — best for

• One H2 missing Citation Capsule
• One section badly off voice
• Word count outside band in 1-2 sections
• Format doesn't match for 1-2 sections

Level 3: Stage-rewrite — best for

• Structural: H2 hierarchy wrong → outline-architect
• Multiple unverified citations → fact-check-and-citation
• Voice inconsistent across most sections → humanizer (rewrite mode)

Level 4: Full-regen — best for

• AI-Slop score consistently high
• Many sections fail quality
• Multiple gates failing

Level 5: From-scratch — last resort

• Article topic just doesn't work in current angle
• Try alternative title from angle.alternative_titles_considered
• Completely new outline

Cost

Each level has roughly cumulative cost (worst case 4 rounds):

• Level 1 × 4 iterations: ~$0.10 (cheap edits)
• Level 1 + 1 × Level 2: ~$0.25 (one section rewrite)
• Up to Level 3: ~$0.50 (one stage rerun)
• Up to Level 4: ~$1.50 (full build rerun)
• Up to Level 5: ~$2.50 (mostly full pipeline rerun)

Cap is set so total repair never exceeds ~$2.50 over baseline article cost.

Output

After each escalation, write workspace/{task}/repair-log.json (append):

{
  "iterations": [
    {
      "iteration": 1,
      "level": 1,
      "trigger": "AI-Slop score 24 > 20 threshold",
      "actions": [
        {"type": "edit", "line": 47, "before": "crucial", "after": "central"}
      ],
      "score_before": 24, "score_after": 18,
      "duration_seconds": 12,
      "verdict": "improved, but still failing other gates"
    },
    {
      "iteration": 2,
      "level": 1,
      "trigger": "Em-dash count 1 > 0",
      "actions": [...],
      "score_before": 18, "score_after": 16,
      "verdict": "ship"
    }
  ],
  "final_level": 1,
  "total_iterations": 2,
  "outcome": "passed"
}

Handoff

If passed:

• recommended_next_skill: independent-reviewer (final blessing)
• → phase-publish if reviewer approves

If exhausted (4 rounds done, still failing):

• completion_status: BLOCKED
• Return workspace/{task}/repair-log.json + best-of-N draft to user
• User decides: accept with caveats, abort, or take over manually

See also

• schemas/quality.schema.json (input format)
• agents/editor-in-chief.md (reverse-callback for Level 2)
• agents/writer.md (Level 2 dispatched agent)
• subskills/build/section-drafter/SKILL.md (Level 3/4 dispatched skill)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: XuanRanL
• Source: XuanRanL/loamwright-SEO-Skill
• License: Apache-2.0
• Homepage: https://loamwrightseo.com

Install and usage instructions live in the source repository linked above.