Agent-Skills.md

Agent Skills: Model Tracking Protocol

MANDATORY tracking protocol for multi-model validation. Creates structured tracking tables BEFORE launching models, tracks progress during execution, and ensures complete results presentation. Use when running 2+ external AI models in parallel. Trigger keywords - "multi-model", "parallel review", "external models", "consensus", "model tracking".

UncategorizedID: madappgang/claude-code/model-tracking-protocol

Repository

madappgang/claude-code
MadAppGangLicense: MIT
24826

Install this agent skill to your local

pnpm dlx add-skill https://github.com/MadAppGang/claude-code/tree/HEAD/plugins/multimodel/skills/model-tracking-protocol

Skill Files

Browse the full folder contents for model-tracking-protocol.

Download Skill

Loading file tree…

plugins/multimodel/skills/model-tracking-protocol/SKILL.md

Skill Metadata

Name
model-tracking-protocol
Description
MANDATORY tracking protocol for multi-model validation. Creates structured tracking tables BEFORE launching models, tracks progress during execution, and ensures complete results presentation. Use when running 2+ external AI models in parallel. Trigger keywords - "multi-model", "parallel review", "external models", "consensus", "model tracking".

Model Tracking Protocol

Version: 1.0.0 Purpose: MANDATORY tracking protocol for multi-model validation to prevent incomplete reviews Status: Production Ready

Overview

This skill defines the MANDATORY tracking protocol for multi-model validation. It provides templates and procedures that make proper tracking unforgettable.

The Problem This Solves:

Agents often launch multiple external AI models but fail to:

  • Create structured tracking tables before launch
  • Collect timing and performance data during execution
  • Document failures with error messages
  • Perform consensus analysis comparing model findings
  • Present results in a structured format

The Solution:

This skill provides MANDATORY checklists, templates, and protocols that ensure complete tracking. Missing ANY of these steps = INCOMPLETE review.

Table of Contents

  1. MANDATORY Pre-Launch Checklist
  2. Tracking Table Templates
  3. Per-Model Status Updates
  4. Failure Documentation Protocol
  5. Consensus Analysis Requirements
  6. Results Presentation Template
  7. Common Failures and Prevention
  8. Integration Examples

MANDATORY Pre-Launch Checklist

You MUST complete ALL items before launching ANY external models.

This is NOT optional. If you skip this, your multi-model validation is INCOMPLETE.

Checklist (Copy and Complete)

PRE-LAUNCH VERIFICATION (complete before Task calls):

[ ] 1. SESSION_ID created: ________________________
[ ] 2. SESSION_DIR created: ________________________
[ ] 3. Tracking table written to: $SESSION_DIR/tracking.md
[ ] 4. Start time recorded: SESSION_START=$(date +%s)
[ ] 5. Model list confirmed (comma-separated): ________________________
[ ] 6. Per-model timing arrays initialized
[ ] 7. Code context written to session directory
[ ] 8. Tracking marker created: /tmp/.claude-multi-model-active

If ANY item is unchecked, STOP and complete it before proceeding.

Why Pre-Launch Matters

Without pre-launch setup, you will:

  • Lose timing data (cannot calculate speed accurately)
  • Miss failed model details (no structured place to record)
  • Skip consensus analysis (no model list to compare)
  • Present incomplete results (no tracking table to populate)

Pre-Launch Script Template

CRITICAL CONSENSUS FIX APPLIED: Use file-based detection instead of environment variables.

#!/bin/bash
# Run this BEFORE launching any Task calls

# 1. Create unique session
SESSION_ID="review-$(date +%Y%m%d-%H%M%S)-$(head -c 4 /dev/urandom | xxd -p)"
SESSION_DIR="/tmp/${SESSION_ID}"
mkdir -p "$SESSION_DIR"

# 2. Record start time
SESSION_START=$(date +%s)

# 3. Create tracking table
cat > "$SESSION_DIR/tracking.md" << EOF
# Multi-Model Tracking

## Session Info
- Session ID: ${SESSION_ID}
- Started: $(date -u +%Y-%m-%dT%H:%M:%SZ)
- Models Requested: [FILL]

## Model Status

| Model | Agent ID | Status | Start | End | Duration | Issues | Quality | Notes |
|-------|----------|--------|-------|-----|----------|--------|---------|-------|
| [MODEL 1] | | pending | | | | | | |
| [MODEL 2] | | pending | | | | | | |
| [MODEL 3] | | pending | | | | | | |

## Failures

| Model | Failure Type | Error Message | Retry? |
|-------|--------------|---------------|--------|

## Consensus

| Issue | Model 1 | Model 2 | Model 3 | Agreement |
|-------|---------|---------|---------|-----------|

EOF

# 4. Initialize timing arrays
declare -A MODEL_START_TIMES
declare -A MODEL_END_TIMES
declare -A MODEL_STATUS

# 5. Create tracking marker file (CRITICAL FIX)
# This allows hooks to detect that tracking is active
echo "$SESSION_DIR" > /tmp/.claude-multi-model-active

echo "Pre-launch setup complete. Session: $SESSION_ID"
echo "Directory: $SESSION_DIR"
echo "Tracking table: $SESSION_DIR/tracking.md"

Strict Mode (Optional)

For stricter enforcement, set:

export CLAUDE_STRICT_TRACKING=true

When enabled, hooks will BLOCK execution if tracking is not set up, rather than just warning.

Tracking Table Templates

Template A: Simple Model Tracking (3-5 models)

| Model | Status | Time | Issues | Quality | Cost |
|-------|--------|------|--------|---------|------|
| claude-embedded | pending | - | - | - | FREE |
| x-ai/grok-code-fast-1 | pending | - | - | - | - |
| qwen/qwen3-coder:free | pending | - | - | - | FREE |

Update as each completes:

| Model | Status | Time | Issues | Quality | Cost |
|-------|--------|------|--------|---------|------|
| claude-embedded | success | 32s | 8 | 95% | FREE |
| x-ai/grok-code-fast-1 | success | 45s | 6 | 87% | $0.002 |
| qwen/qwen3-coder:free | timeout | - | - | - | - |

Template B: Detailed Model Tracking (6+ models)

## Model Execution Status

### Summary
- Total Requested: 8
- Completed: 0
- In Progress: 0
- Failed: 0
- Pending: 8

### Detailed Status

| # | Model | Provider | Status | Start | Duration | Issues | Quality | Cost | Error |
|---|-------|----------|--------|-------|----------|--------|---------|------|-------|
| 1 | claude-embedded | Anthropic | pending | - | - | - | - | FREE | - |
| 2 | x-ai/grok-code-fast-1 | X-ai | pending | - | - | - | - | - | - |
| 3 | qwen/qwen3-coder:free | Qwen | pending | - | - | - | - | FREE | - |
| 4 | google/gemini-3-pro | Google | pending | - | - | - | - | - | - |
| 5 | openai/gpt-5.1-codex | OpenAI | pending | - | - | - | - | - | - |
| 6 | mistralai/devstral | Mistral | pending | - | - | - | - | FREE | - |
| 7 | deepseek/deepseek-r1 | DeepSeek | pending | - | - | - | - | - | - |
| 8 | anthropic/claude-sonnet | Anthropic | pending | - | - | - | - | - | - |

Template C: Session-Based Tracking File

Create this file at $SESSION_DIR/tracking.md:

# Multi-Model Validation Tracking
Session: ${SESSION_ID}
Started: ${TIMESTAMP}

## Pre-Launch Verification
- [x] Session directory created: ${SESSION_DIR}
- [x] Tracking table initialized
- [x] Start time recorded: ${SESSION_START}
- [x] Model list: ${MODEL_LIST}

## Model Status

| Model | Status | Start | Duration | Issues | Quality |
|-------|--------|-------|----------|--------|---------|
| claude | pending | - | - | - | - |
| grok | pending | - | - | - | - |
| gemini | pending | - | - | - | - |

## Failures
(populated as failures occur)

## Consensus
(populated after all complete)

Update Protocol

As each model completes, IMMEDIATELY update:

  1. Status: pending -> in_progress -> success/failed/timeout
  2. Duration: Calculate from start time
  3. Issues: Number of issues found
  4. Quality: Percentage if calculable
  5. Error: If failed, brief error message

DO NOT wait until all models finish. Update as each completes.

Per-Model Status Update Protocol

IMMEDIATELY After Each Model Completes

Do NOT wait until all models finish. Update tracking AS EACH COMPLETES.

Update Script

# Call this when each model completes
update_model_status() {
  local model="$1"
  local status="$2"
  local issues="${3:-0}"
  local quality="${4:-}"
  local error="${5:-}"

  local end_time=$(date +%s)
  local start_time="${MODEL_START_TIMES[$model]}"
  local duration=$((end_time - start_time))

  # Update arrays
  MODEL_END_TIMES["$model"]=$end_time
  MODEL_STATUS["$model"]="$status"

  # Log update to session tracking file
  echo "$(date -u +%Y-%m-%dT%H:%M:%SZ) - Model: $model, Status: $status, Duration: ${duration}s" >> "$SESSION_DIR/execution.log"

  # Update tracking table (append to tracking.md)
  echo "| $model | $status | ${duration}s | $issues | ${quality:-N/A} | ${error:-} |" >> "$SESSION_DIR/tracking.md"

  # Track performance in global statistics
  if [[ "$status" == "success" ]]; then
    track_model_performance "$model" "success" "$duration" "$issues" "$quality"
  else
    track_model_performance "$model" "$status" "$duration" 0 ""
  fi
}

# Usage examples:
update_model_status "claude-embedded" "success" 8 95
update_model_status "x-ai/grok-code-fast-1" "success" 6 87
update_model_status "some-model" "timeout" 0 "" "Exceeded 120s limit"
update_model_status "other-model" "failed" 0 "" "API 500 error"

Status Values

| Status | Meaning | Action | |--------|---------|--------| | pending | Not started | Wait | | in_progress | Currently executing | Monitor | | success | Completed successfully | Collect results | | failed | Error during execution | Document error | | timeout | Exceeded time limit | Note timeout | | cancelled | User cancelled | Note cancellation |

Real-Time Progress Display

Show user progress as models complete:

Model Status (3/5 complete):
✓ claude-embedded (32s, 8 issues)
✓ x-ai/grok-code-fast-1 (45s, 6 issues)
✓ qwen/qwen3-coder:free (52s, 5 issues)
⏳ openai/gpt-5.1-codex (in progress, 60s elapsed)
⏳ google/gemini-3-pro (in progress, 48s elapsed)

Failure Documentation Protocol

EVERY failed model MUST be documented with:

  1. Model name
  2. Failure type (timeout, API error, parse error, etc.)
  3. Error message (exact or summarized)
  4. Whether retry was attempted

Failure Report Template

## Failed Models Report

### Model: x-ai/grok-code-fast-1
- **Failure Type:** API Error
- **Error Message:** "500 Internal Server Error from OpenRouter"
- **Retry Attempted:** Yes, 1 retry, same error
- **Impact:** Review results based on 3/4 models instead of 4
- **Recommendation:** Check OpenRouter status, retry later

### Model: google/gemini-3-pro
- **Failure Type:** Timeout
- **Error Message:** "Exceeded 120s limit, response incomplete"
- **Retry Attempted:** No, time constraints
- **Impact:** Lost Gemini perspective, consensus based on remaining models
- **Recommendation:** Extend timeout to 180s for this model

Failure Categorization

| Category | Common Causes | Recovery | |----------|---------------|----------| | Timeout | Model slow, large input, network latency | Retry with extended timeout | | API Error | Provider down, rate limit, auth issue | Wait and retry, check API status | | Parse Error | Malformed response, encoding issue | Retry, simplify prompt | | Auth Error | Invalid API key, expired token | Check credentials | | Context Limit | Input too large for model | Reduce context, split task | | Rate Limit | Too many requests | Wait, implement backoff |

Failure Summary Table

Always include this in final results:

## Execution Summary

| Metric | Value |
|--------|-------|
| Models Requested | 8 |
| Successful | 5 (62.5%) |
| Failed | 3 (37.5%) |

### Failed Models

| Model | Failure | Recoverable? | Action |
|-------|---------|--------------|--------|
| grok-code-fast-1 | API 500 | Yes - retry later | Check OpenRouter status |
| gemini-3-pro | Timeout | Yes - extend limit | Use 180s timeout |
| deepseek-r1 | Auth Error | No - check key | Verify API key valid |

Writing Failures to Session Directory

# Document failure immediately when it occurs
document_failure() {
  local model="$1"
  local failure_type="$2"
  local error_msg="$3"
  local retry_attempted="${4:-No}"

  cat >> "$SESSION_DIR/failures.md" << EOF

### Model: $model
- **Failure Type:** $failure_type
- **Error Message:** "$error_msg"
- **Retry Attempted:** $retry_attempted
- **Timestamp:** $(date -u +%Y-%m-%dT%H:%M:%SZ)

EOF

  echo "Failure documented: $model ($failure_type)" >&2
}

# Usage:
document_failure "x-ai/grok-code-fast-1" "API Error" "500 Internal Server Error" "Yes, 1 retry"

Consensus Analysis Requirements

After ALL models complete (or max wait time), you MUST perform consensus analysis.

This is NOT optional. Even with 2 successful models, compare their findings.

Minimum Viable Consensus (2 models)

With only 2 models, consensus is simple:

  • AGREE: Both found the same issue
  • DISAGREE: Only one found the issue
| Issue | Model 1 | Model 2 | Consensus |
|-------|---------|---------|-----------|
| SQL injection | Yes | Yes | AGREE |
| Missing validation | Yes | No | Model 1 only |
| Weak hashing | No | Yes | Model 2 only |

Standard Consensus (3-5 models)

| Issue | Claude | Grok | Gemini | Agreement |
|-------|--------|------|--------|-----------|
| SQL injection | Yes | Yes | Yes | UNANIMOUS (3/3) |
| Missing validation | Yes | Yes | No | STRONG (2/3) |
| Rate limiting | Yes | No | No | DIVERGENT (1/3) |

Extended Consensus (6+ models)

For 6+ models, add summary statistics:

## Consensus Summary

- **Unanimous Issues (100%):** 3 issues
- **Strong Consensus (67%+):** 5 issues
- **Majority (50%+):** 2 issues
- **Divergent (<50%):** 4 issues

## Top 5 by Consensus

1. [6/6] SQL injection in search - FIX IMMEDIATELY
2. [6/6] Missing input validation - FIX IMMEDIATELY
3. [5/6] Weak password hashing - RECOMMENDED
4. [4/6] Missing rate limiting - CONSIDER
5. [3/6] Error handling gaps - INVESTIGATE

Consensus Analysis Script

# Perform consensus analysis on all model findings
analyze_consensus() {
  local session_dir="$1"
  local num_models="$2"

  echo "## Consensus Analysis" > "$session_dir/consensus.md"
  echo "" >> "$session_dir/consensus.md"
  echo "Based on $num_models model reviews:" >> "$session_dir/consensus.md"
  echo "" >> "$session_dir/consensus.md"

  # Read all review files and extract issues
  # (simplified - actual implementation would parse review markdown)
  for review in "$session_dir"/*-review.md; do
    echo "Processing: $review"
    # Extract issues, compare, categorize by agreement level
  done

  # Calculate consensus levels
  echo "### Consensus Levels" >> "$session_dir/consensus.md"
  echo "" >> "$session_dir/consensus.md"
  echo "- UNANIMOUS: All $num_models models agree" >> "$session_dir/consensus.md"
  echo "- STRONG: ≥67% of models agree" >> "$session_dir/consensus.md"
  echo "- MAJORITY: ≥50% of models agree" >> "$session_dir/consensus.md"
  echo "- DIVERGENT: <50% of models agree" >> "$session_dir/consensus.md"
}

NO Consensus Analysis = INCOMPLETE Review

If you present results without a consensus comparison, your review is INCOMPLETE.

Minimum Requirements:

  • ✅ Compare findings across ALL successful models
  • ✅ Categorize by agreement level (unanimous, strong, majority, divergent)
  • ✅ Prioritize issues by consensus + severity
  • ✅ Document in $SESSION_DIR/consensus.md

Results Presentation Template

Your final output MUST include ALL of these sections.

Required Output Format

## Multi-Model Review Complete

### Execution Summary

| Metric | Value |
|--------|-------|
| Session ID | review-20251224-143052-a3f2 |
| Session Directory | /tmp/review-20251224-143052-a3f2 |
| Models Requested | 5 |
| Successful | 4 (80%) |
| Failed | 1 (20%) |
| Total Duration | 68s (parallel) |
| Sequential Equivalent | 245s |
| Speedup | 3.6x |

### Model Performance

| Model | Time | Issues | Quality | Status | Cost |
|-------|------|--------|---------|--------|------|
| claude-embedded | 32s | 8 | 95% | Success | FREE |
| x-ai/grok-code-fast-1 | 45s | 6 | 87% | Success | $0.002 |
| qwen/qwen3-coder:free | 52s | 5 | 82% | Success | FREE |
| openai/gpt-5.1-codex | 68s | 7 | 89% | Success | $0.015 |
| mistralai/devstral | - | - | - | Timeout | - |

### Failed Models

| Model | Failure | Error |
|-------|---------|-------|
| mistralai/devstral | Timeout | Exceeded 120s limit |

### Top Issues by Consensus

1. **[UNANIMOUS]** SQL injection in search endpoint
   - Flagged by: claude, grok, qwen, gpt-5 (4/4)
   - Severity: CRITICAL
   - Action: FIX IMMEDIATELY

2. **[UNANIMOUS]** Missing input validation
   - Flagged by: claude, grok, qwen, gpt-5 (4/4)
   - Severity: CRITICAL
   - Action: FIX IMMEDIATELY

3. **[STRONG]** Weak password hashing
   - Flagged by: claude, grok, gpt-5 (3/4)
   - Severity: HIGH
   - Action: RECOMMENDED

### Detailed Reports

- Session directory: /tmp/review-20251224-143052-a3f2
- Consolidated review: /tmp/review-20251224-143052-a3f2/consolidated-review.md
- Individual reviews: /tmp/review-20251224-143052-a3f2/{model}-review.md
- Tracking data: /tmp/review-20251224-143052-a3f2/tracking.md
- Consensus analysis: /tmp/review-20251224-143052-a3f2/consensus.md

### Statistics Saved

- Performance data logged to: ai-docs/llm-performance.json

Missing Section Detection

Before presenting, verify ALL sections are present:

verify_output_complete() {
  local output="$1"

  local required=(
    "Execution Summary"
    "Model Performance"
    "Top Issues"
    "Detailed Reports"
    "Statistics"
  )

  local missing=()
  for section in "${required[@]}"; do
    if ! echo "$output" | grep -q "$section"; then
      missing+=("$section")
    fi
  done

  if [ ${#missing[@]} -gt 0 ]; then
    echo "ERROR: Missing required sections: ${missing[*]}" >&2
    return 1
  fi

  return 0
}

Checklist before presenting results:

  • [ ] Execution Summary (models requested/successful/failed)
  • [ ] Model Performance table (per-model times and quality)
  • [ ] Failed Models section (if any failed)
  • [ ] Top Issues by Consensus (prioritized list)
  • [ ] Detailed Reports (session directory, file paths)
  • [ ] Statistics confirmation (llm-performance.json updated)

Common Failures and Prevention

Failure 1: No Tracking Table Created

Symptom: Results presented as prose, not structured data

What went wrong:

"I ran 5 models. 3 succeeded and found various issues."
(No table, no structure)

Prevention:

  • Always run pre-launch script FIRST
  • Create $SESSION_DIR/tracking.md before Task calls
  • Populate table as models complete

Detection: SubagentStop hook warns if no tracking found

Failure 2: Timing Not Recorded

Symptom: "Duration: unknown" or missing speed stats

What went wrong:

# Launched models without recording start time
Task: reviewer1
Task: reviewer2
# No SESSION_START, cannot calculate duration!

Prevention:

# ALWAYS do this first
SESSION_START=$(date +%s)
MODEL_START_TIMES["model1"]=$SESSION_START

Detection: Hook checks for timing data in output

Failure 3: Failed Models Not Documented

Symptom: "2 of 8 succeeded" with no failure details

What went wrong:

"Launched 8 models. 2 succeeded."
(No info on why 6 failed)

Prevention:

# Immediately when model fails
document_failure "model-name" "Timeout" "Exceeded 120s" "No"

Detection: Hook checks for failure section when success < total

Failure 4: No Consensus Analysis

Symptom: Individual model results listed without comparison

What went wrong:

"Model 1 found: A, B, C
 Model 2 found: B, D, E"
(No comparison: which issues do they agree on?)

Prevention:

  • After all complete, ALWAYS run consolidation
  • Create consensus table comparing findings
  • Prioritize by agreement level

Detection: Hook checks for consensus keywords

Failure 5: Statistics Not Saved

Symptom: No record in ai-docs/llm-performance.json

What went wrong:

# Forgot to call tracking functions
# No record of this session

Prevention:

# ALWAYS call these
track_model_performance "model" "status" duration issues quality
record_session_stats total success failed parallel sequential speedup

Detection: Hook checks file modification time

Prevention Checklist

Before presenting results, verify:

[ ] Tracking table exists at $SESSION_DIR/tracking.md
[ ] Tracking table is populated with all model results
[ ] All model times recorded (or "timeout"/"failed" noted)
[ ] All failures documented in $SESSION_DIR/failures.md
[ ] Consensus analysis performed in $SESSION_DIR/consensus.md
[ ] Results match required output format
[ ] Statistics saved to ai-docs/llm-performance.json
[ ] Session directory contains all artifacts

Integration Examples

Example 1: Complete Multi-Model Review Workflow

#!/bin/bash
# Full multi-model review with complete tracking

# ============================================================================
# PHASE 1: PRE-LAUNCH (MANDATORY)
# ============================================================================

# 1. Create unique session
SESSION_ID="review-$(date +%Y%m%d-%H%M%S)-$(head -c 4 /dev/urandom | xxd -p)"
SESSION_DIR="/tmp/${SESSION_ID}"
mkdir -p "$SESSION_DIR"

# 2. Record start time
SESSION_START=$(date +%s)

# 3. Create tracking table
cat > "$SESSION_DIR/tracking.md" << EOF
# Multi-Model Validation Tracking

## Session: $SESSION_ID
Started: $(date -u +%Y-%m-%dT%H:%M:%SZ)

## Model Status
| Model | Status | Duration | Issues | Quality |
|-------|--------|----------|--------|---------|
EOF

# 4. Initialize timing arrays
declare -A MODEL_START_TIMES
declare -A MODEL_END_TIMES

# 5. Create tracking marker
echo "$SESSION_DIR" > /tmp/.claude-multi-model-active

# 6. Write code context
git diff > "$SESSION_DIR/code-context.md"

echo "Pre-launch complete. Session: $SESSION_ID"

# ============================================================================
# PHASE 2: MODEL EXECUTION (Parallel Task calls)
# ============================================================================

# Record start times for each model
MODEL_START_TIMES["claude-embedded"]=$(date +%s)
MODEL_START_TIMES["x-ai/grok-code-fast-1"]=$(date +%s)
MODEL_START_TIMES["qwen/qwen3-coder:free"]=$(date +%s)

# Launch all models in single message (parallel execution)
# (These would be actual Task calls in practice)
echo "Launching 3 models in parallel..."

# ============================================================================
# PHASE 3: RESULTS COLLECTION (as each completes)
# ============================================================================

# Update status immediately after each completes
update_model_status() {
  local model="$1" status="$2" issues="${3:-0}" quality="${4:-}"
  local end_time=$(date +%s)
  local duration=$((end_time - MODEL_START_TIMES["$model"]))

  echo "| $model | $status | ${duration}s | $issues | ${quality:-N/A} |" >> "$SESSION_DIR/tracking.md"
  track_model_performance "$model" "$status" "$duration" "$issues" "$quality"
}

# Example completions
update_model_status "claude-embedded" "success" 8 95
update_model_status "x-ai/grok-code-fast-1" "success" 6 87
update_model_status "qwen/qwen3-coder:free" "timeout"

# ============================================================================
# PHASE 4: CONSENSUS ANALYSIS (MANDATORY)
# ============================================================================

# Consolidate and compare findings
echo "Performing consensus analysis..."
# (Would launch consolidation agent here)

# ============================================================================
# PHASE 5: STATISTICS & PRESENTATION
# ============================================================================

# Calculate session stats
PARALLEL_TIME=52  # max of all durations
SEQUENTIAL_TIME=129  # sum of all durations
SPEEDUP=2.5

# Record session
record_session_stats 3 2 1 "$PARALLEL_TIME" "$SEQUENTIAL_TIME" "$SPEEDUP"

# Present results
cat << RESULTS
## Multi-Model Review Complete

Session: $SESSION_ID
Directory: $SESSION_DIR

Models: 3 requested, 2 successful, 1 failed

See tracking table: $SESSION_DIR/tracking.md
See consensus: $SESSION_DIR/consensus.md
Statistics saved to: ai-docs/llm-performance.json
RESULTS

# Cleanup marker
rm -f /tmp/.claude-multi-model-active

Example 2: Minimal 2-Model Comparison

# Simplest viable multi-model validation

# Pre-launch
SESSION_ID="review-$(date +%s)"
SESSION_DIR="/tmp/$SESSION_ID"
mkdir -p "$SESSION_DIR"
SESSION_START=$(date +%s)
echo "$SESSION_DIR" > /tmp/.claude-multi-model-active

# Launch
echo "Launching Claude + Grok..."
# Task: claude-embedded
# Task: claudish grok

# Track
track_model_performance "claude" "success" 32 8 95
track_model_performance "grok" "success" 45 6 87

# Consensus
echo "Issues both found: SQL injection, missing validation" > "$SESSION_DIR/consensus.md"

# Stats
record_session_stats 2 2 0 45 77 1.7

# Cleanup
rm -f /tmp/.claude-multi-model-active

Example 3: Handling Failures

# Multi-model with failure handling

# Pre-launch (same as Example 1)
# ... setup code ...

# Launch 4 models
# ... Task calls ...

# Model 1: Success
update_model_status "claude" "success" 32 8 95

# Model 2: Success
update_model_status "grok" "success" 45 6 87

# Model 3: Timeout
update_model_status "gemini" "timeout"
document_failure "gemini" "Timeout" "Exceeded 120s limit" "No"

# Model 4: API Error
update_model_status "gpt5" "failed"
document_failure "gpt5" "API Error" "500 from OpenRouter" "Yes, 1 retry"

# Proceed with 2 successful models
if [ "$SUCCESS_COUNT" -ge 2 ]; then
  echo "Proceeding with $SUCCESS_COUNT successful models"
  # Consensus with partial data
else
  echo "ERROR: Only $SUCCESS_COUNT succeeded, need minimum 2"
fi

Integration with Other Skills

With multi-model-validation

The multi-model-validation skill defines the execution patterns (4-Message Pattern, parallel execution, proxy mode). This skill (model-tracking-protocol) defines the tracking infrastructure.

Use together:

skills: orchestration:multi-model-validation, orchestration:model-tracking-protocol

Workflow:

  1. Read multi-model-validation for execution patterns
  2. Read model-tracking-protocol for tracking setup
  3. Pre-launch (tracking protocol)
  4. Execute (validation patterns)
  5. Track (protocol updates)
  6. Present (protocol templates)

With quality-gates

Use quality gates to ensure tracking is complete before proceeding:

# After tracking setup, verify completeness
if [ ! -f "$SESSION_DIR/tracking.md" ]; then
  echo "QUALITY GATE FAILED: No tracking table"
  exit 1
fi

# Before presenting results, verify all sections present
verify_output_complete "$OUTPUT" || exit 1

With task-orchestration

Track progress through multi-model phases:

Tasks:
1. [~] Pre-launch setup (tracking protocol)
2. [ ] Launch models (validation patterns)
3. [ ] Collect results (tracking updates)
4. [ ] Consensus analysis (protocol requirement)
5. [ ] Present results (protocol template)

Quick Reference

File-Based Tracking Marker (CONSENSUS FIX)

Create marker after pre-launch setup:

echo "$SESSION_DIR" > /tmp/.claude-multi-model-active

Check if tracking active (in hooks):

if [[ -f /tmp/.claude-multi-model-active ]]; then
  SESSION_DIR=$(cat /tmp/.claude-multi-model-active)
  [[ -f "$SESSION_DIR/tracking.md" ]] && echo "Tracking active"
fi

Remove marker when done:

rm -f /tmp/.claude-multi-model-active

Pre-Launch Commands

SESSION_ID="review-$(date +%Y%m%d-%H%M%S)-$(head -c 4 /dev/urandom | xxd -p)"
SESSION_DIR="/tmp/${SESSION_ID}"
mkdir -p "$SESSION_DIR"
SESSION_START=$(date +%s)
echo "$SESSION_DIR" > /tmp/.claude-multi-model-active

Tracking Commands

update_model_status "model" "status" issues quality
document_failure "model" "type" "error" "retry"
track_model_performance "model" "status" duration issues quality
record_session_stats total success failed parallel sequential speedup

Verification Commands

verify_output_complete "$OUTPUT"
[ -f "$SESSION_DIR/tracking.md" ] && echo "Tracking exists"
[ -f ai-docs/llm-performance.json ] && echo "Statistics saved"

Summary

This skill provides MANDATORY tracking infrastructure for multi-model validation:

  1. Pre-Launch Checklist - 8 items to complete before launching models
  2. Tracking Tables - Templates for 3-5 models and 6+ models
  3. Status Updates - Per-model completion tracking
  4. Failure Documentation - Required format for all failures
  5. Consensus Analysis - Comparing findings across models
  6. Results Template - Required output format
  7. Common Failures - Prevention strategies
  8. Integration Examples - Complete workflows

Key Innovation: File-based tracking marker (/tmp/.claude-multi-model-active) allows hooks to detect active tracking without relying on environment variables.

Use this skill when: Running 2+ external AI models in parallel for validation, review, or consensus analysis.

Missing tracking = INCOMPLETE validation.