AI Model Performance Dashboard - Statuslines

Features

• Multi-provider model support (Claude, GPT, Gemini, Grok, Llama)
• Context occupancy percentage with color-coded warnings
• Truncation risk alerts based on real-world model reliability limits (65%)
• Time to First Token (TTFT) estimation for latency monitoring
• Tokens per minute rate calculation for throughput tracking
• Formatted number display with thousands separators
• 2025 verified context limits for all major models
• Multi-line dashboard for comprehensive metrics at a glance

Use Cases

• Production workflows requiring performance monitoring
• Multi-model comparison and optimization
• Preventing context truncation in long conversations
• Tracking token consumption rates for cost management
• Identifying latency issues before they impact workflow
• SLA monitoring for enterprise AI deployments
• Performance regression detection across model versions

Requirements

• Claude Code CLI installed and configured
• Bash shell available (bash 4.0+ recommended for arithmetic operations)
• jq command-line JSON processor (jq 1.6+ recommended for safe extraction with // defaults)
• date command with epoch conversion support (macOS: -j flag, Linux: -d flag)
• Terminal with ANSI color code support (256-color mode recommended for traffic light colors)
• Write access to /tmp directory (optional, for caching performance metrics)

Configuration

{
  "statusLine": {
    "type": "command",
    "command": "$CLAUDE_PROJECT_DIR/.claude/statuslines/ai-model-performance-dashboard.sh"
  }
}

Examples

Enhanced Performance Dashboard with Cost Tracking

Extended version with cost per token and total cost metrics

#!/usr/bin/env bash

# Enhanced AI Model Performance Dashboard with Cost Tracking

input=$(cat)

model=$(echo "$input" | jq -r '.model.id // .model.display_name // "unknown"')
prompt_tokens=$(echo "$input" | jq -r '.session.promptTokens // 0')
total_tokens=$(echo "$input" | jq -r '.session.totalTokens // 0')
cost=$(echo "$input" | jq -r '.cost.total_cost_usd // 0')
session_start=$(echo "$input" | jq -r '.session.startTime // ""')

# Model context limits and pricing (2025)
case "$model" in
  *"claude-sonnet-4"*|*"claude-opus-4"*)
    context_limit=1000000
    model_display="Claude Sonnet 4"
    cost_per_1k_input=0.003
    cost_per_1k_output=0.015
    ;;
  *"gpt-4.1"*|*"gpt-4-turbo"*)
    context_limit=1000000
    model_display="GPT-4.1 Turbo"
    cost_per_1k_input=0.01
    cost_per_1k_output=0.03
    ;;
  *)
    context_limit=100000
    model_display="$model"
    cost_per_1k_input=0.001
    cost_per_1k_output=0.002
    ;;
esac

occupancy=$((prompt_tokens * 100 / context_limit))
reliable_limit=$((context_limit * 65 / 100))

if [ $prompt_tokens -gt $reliable_limit ]; then
  truncation="⚠ TRUNCATION RISK"
  TRUNC_COLOR="\033[38;5;196m"
else
  truncation="✓ Safe"
  TRUNC_COLOR="\033[38;5;46m"
fi

if [ $occupancy -lt 50 ]; then
  OCC_COLOR="\033[38;5;46m"
elif [ $occupancy -lt 80 ]; then
  OCC_COLOR="\033[38;5;226m"
else
  OCC_COLOR="\033[38;5;196m"
fi

# Calculate tokens/min
if [ -n "$session_start" ]; then
  current_time=$(date +%s)
  start_epoch=$(date -j -f "%Y-%m-%dT%H:%M:%SZ" "$session_start" +%s 2>/dev/null || date -d "$session_start" +%s 2>/dev/null || echo "$current_time")
  elapsed_seconds=$((current_time - start_epoch))
  tokens_per_min=$((elapsed_seconds > 0 ? total_tokens * 60 / elapsed_seconds : 0))
else
  tokens_per_min=0
fi

formatNumber() { printf "%'d" "$1" 2>/dev/null || echo "$1"; }

RESET="\033[0m"
BOLD="\033[1m"

echo -e "${BOLD}📊 ${model_display}${RESET}"
echo -e "${OCC_COLOR}Occupancy: ${occupancy}%${RESET} ($(formatNumber $prompt_tokens)/$(formatNumber $context_limit)) | ${TRUNC_COLOR}${truncation}${RESET}"
echo -e "Rate: $(formatNumber $tokens_per_min) tok/min | Cost: $${cost} | Total: $(formatNumber $total_tokens)"

Performance Dashboard with Custom Reliable Limit

Configurable truncation threshold for different risk tolerances

#!/usr/bin/env bash

# AI Model Performance Dashboard with Custom Reliable Limit
# Configure via RELIABLE_LIMIT_PERCENT environment variable (default 65%)

input=$(cat)

RELIABLE_LIMIT_PERCENT=${RELIABLE_LIMIT_PERCENT:-65}

model=$(echo "$input" | jq -r '.model.id // .model.display_name // "unknown"')
prompt_tokens=$(echo "$input" | jq -r '.session.promptTokens // 0')
total_tokens=$(echo "$input" | jq -r '.session.totalTokens // 0')

# Model context limits
case "$model" in
  *"claude-sonnet-4"*|*"claude-opus-4"*)
    context_limit=1000000
    model_display="Claude Sonnet 4"
    ;;
  *"gpt-4.1"*|*"gpt-4-turbo"*)
    context_limit=1000000
    model_display="GPT-4.1 Turbo"
    ;;
  *"gemini-1.5-pro"*)
    context_limit=2000000
    model_display="Gemini 1.5 Pro"
    ;;
  *)
    context_limit=100000
    model_display="$model"
    ;;
esac

occupancy=$((prompt_tokens * 100 / context_limit))
reliable_limit=$((context_limit * RELIABLE_LIMIT_PERCENT / 100))

if [ $prompt_tokens -gt $reliable_limit ]; then
  truncation="⚠ TRUNCATION RISK"
  TRUNC_COLOR="\033[38;5;196m"
else
  truncation="✓ Safe"
  TRUNC_COLOR="\033[38;5;46m"
fi

if [ $occupancy -lt 50 ]; then
  OCC_COLOR="\033[38;5;46m"
elif [ $occupancy -lt 80 ]; then
  OCC_COLOR="\033[38;5;226m"
else
  OCC_COLOR="\033[38;5;196m"
fi

formatNumber() { printf "%'d" "$1" 2>/dev/null || echo "$1"; }

RESET="\033[0m"
BOLD="\033[1m"

echo -e "${BOLD}📊 ${model_display}${RESET}"
echo -e "${OCC_COLOR}Occupancy: ${occupancy}%${RESET} ($(formatNumber $prompt_tokens)/$(formatNumber $context_limit)) | ${TRUNC_COLOR}${truncation}${RESET} (limit: ${RELIABLE_LIMIT_PERCENT}%)"
echo -e "Total: $(formatNumber $total_tokens) tokens"

Performance Dashboard Installation Example

Complete setup script with model limit updates

#!/bin/bash
# Installation script for AI Model Performance Dashboard

mkdir -p .claude/statuslines

cat > .claude/statuslines/ai-model-performance-dashboard.sh << 'SCRIPT_EOF'
#!/usr/bin/env bash
# Paste the full statusline script from the primary example above before running this installer.
SCRIPT_EOF

chmod +x .claude/statuslines/ai-model-performance-dashboard.sh

# Add to settings.json
if [ ! -f .claude/settings.json ]; then
  echo '{"statusLine":{"type":"command","command":"$CLAUDE_PROJECT_DIR/.claude/statuslines/ai-model-performance-dashboard.sh"}}' > .claude/settings.json
else
  jq '.statusLine = {"type":"command","command":"$CLAUDE_PROJECT_DIR/.claude/statuslines/ai-model-performance-dashboard.sh"}' .claude/settings.json > .claude/settings.json.tmp
  mv .claude/settings.json.tmp .claude/settings.json
fi

echo "AI Model Performance Dashboard installed successfully!"
echo "Note: Update model context limits in script as new models are released."

Troubleshooting

Occupancy percentage always showing 0% or incorrect values

Verify session.promptTokens exists in JSON: jq .session.promptTokens. Check context_limit set correctly for model. Test calculation: echo $((100000 * 100 / 1000000)) (should be 10). Ensure integer division working. Verify prompt_tokens is numeric: echo "$prompt_tokens" | grep -E '^[0-9]+$'.

Model not recognized, showing default context_limit 100000

Check model string extraction: jq -r '.model.id // .model.display*name'. Verify case statement matches model name pattern. Test model matching: echo "$model" | grep -i 'claude-sonnet-4'. Add custom model: Add case for *"your-model"_) context_limit=X ;; before default case. Check model name format in Claude Code JSON output.

Tokens per minute showing 0 or extremely high numbers

Verify session.startTime format: jq -r '.session.startTime'. Check date parsing works: macOS (date -j -f '%Y-%m-%dT%H:%M:%SZ' '2025-10-23T10:00:00Z' +%s) or Linux (date -d '2025-10-23T10:00:00Z' +%s). Ensure elapsed_seconds > 0. Check timezone handling. Verify session_start is not empty.

Number formatting with commas not working (showing raw numbers)

Check printf thousands separator support: printf "%'d" 1000000. If unsupported, formatNumber falls back to raw echo. Enable locale: export LC_NUMERIC=en_US.UTF-8. Test: locale | grep LC_NUMERIC. Verify printf supports ' flag (POSIX extension). Use alternative: sed 's/([0-9]{1,3})$/\1/' for basic formatting.

Truncation warning appearing too early or too late

Adjust reliable_limit calculation (currently 65% of context_limit). Research shows models fail 30-35% before claimed limits. Increase for conservative: reliable_limit=$((context_limit * 50 / 100)). Decrease for aggressive: reliable_limit=$((context_limit * 75 / 100)). Set via environment variable: export RELIABLE_LIMIT_PERCENT=70. Test with known token counts.

TTFT (Time to First Token) estimation always shows same value

TTFT is estimated based on prompt_tokens, not actual measured latency. For accurate TTFT, integrate with Claude Code API response timing. Current estimation: <10k tokens (~0.8s), 10k-50k (~1.2s), >50k (~2.5s). Consider adding actual timing measurement if Claude Code exposes response start time in JSON.

Color codes not displaying correctly in terminal

Verify terminal supports 256-color mode: echo $TERM (should include '256color' or 'xterm-256color'). Check ANSI code format: Use \033[38;5;Xm for 256-color mode. Test colors: echo -e '\033[38;5;46mGreen\033[0m'. Verify echo -e flag is used. Check terminal emulator color support settings.

Multi-line dashboard not displaying correctly (showing as single line)

Verify echo -e flag is used for escape sequence interpretation. Check terminal supports multi-line statuslines (Claude Code feature). Test: echo -e "Line 1\nLine 2". Some terminals may collapse multi-line output. Consider single-line format if multi-line not supported.

• Source repositorygithub.com 2026-07-13T15:19:32-07:00
• Documentationcode.claude.com