session-cost（会话成本）

Session Cost

Analyze OpenClaw session logs for token usage, costs, and performance metrics grouped by agent and model.

By default, scans all agents in ~/.openclaw/agents/. Each agent's sessions are read from ~/.openclaw/agents/<name>/sessions/.

Quick Start

# Summary across all agents
node /home/claw/.openclaw/workspace/skills/session-cost/scripts/session-cost.js

# Show all session details
node /home/claw/.openclaw/workspace/skills/session-cost/scripts/session-cost.js --details

# Show details for a specific session (searches across all agents)
node /home/claw/.openclaw/workspace/skills/session-cost/scripts/session-cost.js --details abc123

Options

• --path <dir> — Directory to scan for .jsonl files (overrides agent auto-discovery)
• --agent <name> — Filter by agent name (e.g., main, codegen)
• --offset <time> — Only include sessions from the last N units (30m, 2h, 7d)
• --provider <name> — Filter by model provider (anthropic, openai, ollama, etc.)
• --details [session-id] — Show per-session details. Optionally pass a session ID to show just that session (searches across all agents for <id>.jsonl)
• --table — Show details in compact table format (use with --details)
• --format <type> — Output format: text (default), json, or discord
• --json — Shorthand for --format json (backwards compat)
• --help, -h — Show help message

Examples

# Last 24 hours summary
node /home/claw/.openclaw/workspace/skills/session-cost/scripts/session-cost.js --offset 24h

# Only the main agent
node /home/claw/.openclaw/workspace/skills/session-cost/scripts/session-cost.js --agent main

# Last 7 days, JSON output
node /home/claw/.openclaw/workspace/skills/session-cost/scripts/session-cost.js --offset 7d --json

# Discord-friendly format (for bots/chat)
node /home/claw/.openclaw/workspace/skills/session-cost/scripts/session-cost.js --format discord

# Discord format with filters
node /home/claw/.openclaw/workspace/skills/session-cost/scripts/session-cost.js --format discord --offset 24h --provider anthropic

# Filter by provider
node /home/claw/.openclaw/workspace/skills/session-cost/scripts/session-cost.js --provider anthropic

# All sessions in compact table format
node /home/claw/.openclaw/workspace/skills/session-cost/scripts/session-cost.js --details --table

# Custom path with details (overrides agent discovery, scans exact directory)
node /home/claw/.openclaw/workspace/skills/session-cost/scripts/session-cost.js --path /other/dir --details

# Single session detail (found automatically across agents)
node /home/claw/.openclaw/workspace/skills/session-cost/scripts/session-cost.js --details 9df7a399-8254-411b-a875-e7337df73d29

# Anthropic sessions from last 24h in table format
node /home/claw/.openclaw/workspace/skills/session-cost/scripts/session-cost.js --provider anthropic --offset 24h --details --table

Output Format

Text Summary (Default)

Results are grouped by agent, then by model within each agent. A grand total section shows per-agent subtotals and a combined total.

Found 52 .jsonl files across 2 agents, 52 matched

====================================================================================================
SUMMARY BY AGENT
====================================================================================================

Agent: main

  anthropic/claude-sonnet-4-5-20250929
  --------------------------------------------------------------------------------
    Sessions: 30
    Tokens:   1,234,567 (input: 900,000, output: 334,567)
    Cache:    read: 500,000 tokens, write: 200,000 tokens
    Cost:     $12.3456
      Input:       $5.4000
      Output:      $5.0185
      Cache read:  $1.5000  (included in total, discounted rate)
      Cache write: $0.4271  (included in total)

  anthropic/claude-opus-4-6
  --------------------------------------------------------------------------------
    Sessions: 5
    Tokens:   250,000 (input: 180,000, output: 70,000)
    ...

Agent: codegen

  anthropic/claude-sonnet-4-5-20250929
  --------------------------------------------------------------------------------
    Sessions: 17
    ...

====================================================================================================
GRAND TOTAL
====================================================================================================
  main                 — 35 sessions, $15.8200
  codegen              — 17 sessions, $8.5600

All agents (2)
--------------------------------------------------------------------------------
  Sessions: 52
  Tokens:   ...
  Cost:     $24.3800
  ...

When only a single agent is present, the grand total shows "All models (N)" instead.

Text Details (--details)

Shows per-session breakdown (session ID, agent, model, duration, timestamps, tokens, cache, cost) followed by the agent/model summary.

Table Format (--details --table)

Compact table view. When multiple agents are present, an Agent column is included.

SESSION DETAILS
============================================================================================================================================
Agent          Model                              Duration    Tokens        Cache               Cost        Session
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
main           anthropic/claude-sonnet-4.5        45 min      128.5K        15.2K / 8.1K        $0.3245     abc123def456
codegen        anthropic/claude-sonnet-4.5        12 min      45.3K         2.1K / 1.5K         $0.8921     xyz789abc012

With a single agent, the Agent column is omitted and the table matches the previous format.

JSON (--format json)

Results are nested by agent. Each agent contains its model summaries and an agent-level totals object. A top-level grandTotal aggregates across all agents.

{
  "agents": {
    "main": {
      "models": {
        "anthropic/claude-sonnet-4-5-20250929": {
          "sessions": 30,
          "tokens": { "input": 900000, "output": 334567, "total": 1234567 },
          "cache": { "read": 500000, "write": 200000 },
          "cost": { "total": 12.3456, "input": 5.4, "output": 5.0185, "cacheRead": 1.5, "cacheWrite": 0.4271 }
        }
      },
      "totals": {
        "sessions": 35,
        "tokens": { "input": 1080000, "output": 404567, "total": 1484567 },
        "cache": { "read": 600000, "write": 250000 },
        "cost": { "total": 15.82, ... }
      }
    },
    "codegen": {
      "models": { ... },
      "totals": { ... }
    }
  },
  "grandTotal": {
    "sessions": 52,
    "tokens": { "input": 1500000, "output": 600000, "total": 2100000 },
    "cache": { "read": 800000, "write": 350000 },
    "cost": { "total": 24.38, ... }
  }
}

Discord (--format discord)

Optimized for chat platforms (Discord, Slack, etc.) - concise, markdown-friendly, no tables:

💰 **Usage Summary**
(last 24h)

**Total Cost:** $24.38
**Total Tokens:** 2.1M
**Sessions:** 52

**By Agent:**
• main: $15.82 (35 sessions)
• codegen: $8.56 (17 sessions)

**By Provider:**
• anthropic: $22.50 (1.9M tokens)
• openai: $1.88 (200K tokens)

**Top Models:**
• anthropic/claude-sonnet-4.5: $18.20 (1.5M tokens)
• anthropic/claude-opus-4: $4.30 (400K tokens)
• openai/gpt-4o: $1.88 (200K tokens)

The "By Agent" section is shown only when multiple agents are present.

Output Fields

• Agent — Agent name (derived from directory under ~/.openclaw/agents/)
• Sessions — Number of session files analyzed
• Tokens — Total, input, and output token counts
• Cache — Cache read and write token counts
• Cost — Total cost broken down by input, output, cache read, and cache write
• Duration — Session duration in minutes (details mode)
• Timestamps — First and last activity timestamps (details mode)

Notes

• When --path is provided, it overrides agent auto-discovery and scans exactly that directory. The agent name is inferred from the path (e.g., .../agents/main/sessions → "main").
• --agent and --provider filters can be combined (e.g., --agent main --provider anthropic).
• Single session lookup (--details <id>) searches across all discovered agents to find the session file.
• *.trajectory.jsonl files are skipped — they use a different schema and would otherwise show up as unknown/unknown rows with zero tokens.
• Local models (Ollama, etc.) that genuinely report zero token usage display (not reported by model) next to the token counts so they can be distinguished from sessions with no assistant activity.