source-command-flatline-review

name	source-command-flatline-review
description	Manual invocation of Flatline Protocol for adversarial multi-model review. Use when auto-trigger is disabled or to re-run after document changes. Also provides rollback capability for autonomous integrations.

source-command-flatline-review

Use this skill when the user asks to run the migrated source command flatline-review.

Command Template

Flatline Review

Purpose

Manually invoke the Flatline Protocol for adversarial multi-model review of planning documents.

Usage

# Review current PRD
/flatline-review prd

# Review current SDD
/flatline-review sdd

# Review specific document
/flatline-review grimoires/loa/sdd.md

# Quick skeptic-only review
/flatline-review prd --skeptic-only

# Dry run to validate without API calls
/flatline-review sdd --dry-run

When to Use

• After making significant document changes
• When auto-trigger is disabled in config
• To get a fresh perspective on a stuck decision
• Before finalizing planning documents
• When you want to re-run after addressing feedback

Workflow

Step 1: Resolve Document

# Auto-detect document from argument
case "$argument" in
    prd)
        doc="grimoires/loa/prd.md"
        phase="prd"
        ;;
    sdd)
        doc="grimoires/loa/sdd.md"
        phase="sdd"
        ;;
    sprint)
        doc="grimoires/loa/sprint.md"
        phase="sprint"
        ;;
    *)
        # Assume it's a path
        doc="$argument"
        # Infer phase from filename
        phase=$(basename "$doc" | sed 's/\.md$//' | grep -oE 'prd|sdd|sprint' || echo "prd")
        ;;
esac

Step 2: Run Flatline Protocol

result=$(.claude/scripts/flatline-orchestrator.sh \
    --doc "$doc" \
    --phase "$phase" \
    ${skip_knowledge:+--skip-knowledge} \
    ${budget:+--budget "$budget"} \
    --json)

Step 3: Display Results

Present results to user:

HIGH_CONSENSUS items (score >700 both models):

• These improvements have strong agreement
• Consider integrating them directly
• After integration, display: "Multi-model review working as designed. /feedback if you disagree." (Only when at least 1 HIGH_CONSENSUS item was integrated)

DISPUTED items (delta >300):

• Models disagreed significantly
• Review each one and decide

BLOCKERS (skeptic concern >700):

• Critical concerns that may block progress
• Must address before finalizing

Step 4: Save Results

mkdir -p grimoires/loa/a2a/flatline
echo "$result" | jq . > "grimoires/loa/a2a/flatline/${phase}-review.json"

Output Format

{
    "phase": "prd",
    "document": "grimoires/loa/prd.md",
    "timestamp": "2026-02-03T12:00:00Z",
    "consensus_summary": {
        "high_consensus_count": 5,
        "disputed_count": 2,
        "low_value_count": 3,
        "blocker_count": 1,
        "model_agreement_percent": 73
    },
    "high_consensus": [...],
    "disputed": [...],
    "blockers": [...],
    "metrics": {
        "total_latency_ms": 95000,
        "cost_cents": 245
    }
}

Skeptic-Only Mode

When --skeptic-only is specified:

• Only run skeptic reviews (no improvements)
• Faster execution (~40% of full cost)
• Focus on finding problems rather than suggestions
• Good for quick sanity checks

Integration with Planning Commands

This command provides manual control. For automatic integration:

1. Enable in .loa.config.yaml:

   flatline_protocol:
     enabled: true
     auto_trigger: true
   

2. Flatline will run automatically after /plan-and-analyze, /architect, /sprint-plan

Configuration

# .loa.config.yaml
flatline_protocol:
  enabled: true
  auto_trigger: false  # Set to true for automatic execution

  models:
    primary: "opus"
    secondary: "gpt-5.3-codex"

  thresholds:
    high_consensus: 700
    dispute_delta: 300
    low_value: 400
    blocker: 700

  budget:
    max_tokens_per_phase: 100000
    warn_at_percent: 80

Error Handling

Error	Cause	Resolution
"Document not found"	Invalid path	Check document path
"Flatline disabled"	Config disabled	Enable in .loa.config.yaml
"API error"	Model unavailable	Check API keys, retry later
"Budget exceeded"	Cost limit hit	Increase budget or use --skeptic-only
"Timeout"	Slow API response	Retry or increase timeout

Rollback Commands (v1.22.0)

Rollback auto-integrated changes from autonomous Flatline execution.

Usage

# Rollback single integration by ID
/flatline-review --rollback abc123-001-f7e8d9

# Rollback entire run
/flatline-review --rollback --run-id flatline-run-abc123

# Direct snapshot restore
/flatline-review --rollback --snapshot 20260203_143000_a1b2c3d4

# Force rollback despite divergence
/flatline-review --rollback abc123-001-f7e8d9 --force

# List rollback options for a run
/flatline-review --rollback --run-id flatline-run-abc123 --dry-run

Rollback Workflow

When --rollback is specified:

if [[ -n "$rollback" ]]; then
    if [[ -n "$snapshot" ]]; then
        # Direct snapshot restore
        .claude/scripts/flatline-rollback.sh snapshot --snapshot-id "$snapshot" ${force:+--force}
    elif [[ -n "$run_id" ]]; then
        # Full run rollback
        .claude/scripts/flatline-rollback.sh run --run-id "$run_id" ${dry_run:+--dry-run} ${force:+--force}
    else
        # Single integration rollback
        .claude/scripts/flatline-rollback.sh single --integration-id "$rollback" ${run_id:+--run-id "$run_id"} ${force:+--force}
    fi
    exit $?
fi

Divergence Detection

Before rollback, the system checks if the document has been modified since integration:

• Expected hash: Stored at integration time
• Current hash: Calculated from current file

If diverged:

• Without --force: Rollback refused with warning
• With --force: Creates backup and proceeds

Backup Creation

Before any rollback, a backup is created:

• Path: {document}.pre-rollback-{timestamp}
• Permissions: 600 (owner only)
• Always created unless document doesn't exist

Interactive Confirmation

In interactive mode, user confirmation is requested before rollback:

Document has been modified since integration.
Current hash:  a1b2c3d4...
Expected hash: e5f6g7h8...

This may overwrite changes made after the integration.
Continue with rollback? [y/N]

In autonomous mode, --force is required to override divergence.

Related

• /plan-and-analyze - Creates PRD (auto-triggers Flatline if enabled)
• /architect - Creates SDD (auto-triggers Flatline if enabled)
• /sprint-plan - Creates sprint plan (auto-triggers Flatline if enabled)
• /gpt-review - Single-model GPT review (simpler alternative)
• /run sprint-plan - Autonomous sprint execution (uses autonomous Flatline)