Token Cost Ledger

Reference

• ID: S-OPS-08
• Category: Ops & Risk
• Inputs: token usage data, company.config.yaml (ai.cost_budget_monthly, ai.cost_tracking_enabled, ai.cost_alert_threshold_percent)
• Outputs: ledger entries in cogs/ai-ledger/entries.jsonl, summary in cogs/ai-ledger/summary.md
• Used by: Ops & Risk Agent, Orchestrator Agent (post-ship advisory)
• Tool scripts: ./tools/ops/token-ledger.sh

Purpose

Track AI token consumption and costs as Cost of Goods Sold (COGS). Provides a lightweight, git-trackable cost accounting system for solo developers using AI agents. Supports logging individual entries, generating period summaries, showing per-feature cost breakdowns, monitoring budget health, and exporting data for financial reporting.

When to Use

• After completing a ship flow or significant AI-assisted work session
• When reviewing monthly AI spend against budget
• When preparing investor updates or financial reports (feeds into pricing-unit-economics)
• When deciding whether to use Opus vs Sonnet for a task (cost awareness)
• When company.stage transitions (budgets should be reassessed)

Procedure

Step 1: Check Configuration

Read company.config.yaml and extract:

• ai.cost_budget_monthly — monthly spend cap
• ai.cost_tracking_enabled — whether tracking is active (default: true if budget is set)
• ai.cost_alert_threshold_percent — warning threshold (default: 80)

If ai.cost_budget_monthly is empty, suggest the user set a budget with guidance by stage:

• idea: $25-50/month (exploration, light usage)
• mvp: $50-200/month (active development, ship flows)
• growth: $200-1000/month (multiple features, more agent usage)
• scale: $1000+/month (heavy automation, multiple concurrent flows)

Step 2: Subcommand Routing

Subcommand: log

Log a token cost entry:

bash
1./tools/ops/token-ledger.sh log \
2  --model "claude-sonnet-4-20250514" \
3  --input-tokens 12500 \
4  --output-tokens 3200 \
5  --cache-read 8000 \
6  --cache-write 1500 \
7  --agent engineering \
8  --category ship-flow \
9  --feature PRD-001 \
10  --session "ship-user-auth-20260301" \
11  --notes "RFC generation for user auth"

If the user provides raw numbers from their session, help them fill in the fields. Cost is auto-calculated from the model rates in standards/ops/token-cost-tracking.md when not provided via --cost.

After logging, the tool shows budget impact:

Entry logged: $0.08
  Budget: $13.82 / $100.00 (13.8%)

Subcommand: report

Generate a cost report:

bash
1./tools/ops/token-ledger.sh summary --period monthly

Present the results with analysis:

1. Total spend by category (ship-flow, agent-session, research, ad-hoc)
2. Total spend by model (which models are costing the most)
3. Total spend by agent (which agents are most expensive)
4. Budget health with color coding

Also regenerates cogs/ai-ledger/summary.md for human review.

If the user has investor reporting set up, note that token costs feed into the infrastructure cost line item in standards/ops/investor-reporting-template.md.

Subcommand: budget-check

Quick budget health check:

bash
1./tools/ops/token-ledger.sh budget

Present results:

• Current month spend vs budget
• Projected end-of-month spend (based on daily average)
• Recommendation: stay the course / reduce usage / increase budget

Subcommand: feature-cost

Show total cost of building a specific feature:

bash
1./tools/ops/token-ledger.sh feature-cost PRD-001

Present the feature cost report:

• Total cost with entry count
• Breakdown by agent (who spent what)
• Breakdown by phase (ship-flow, research, etc.)
• Timeline showing each entry chronologically

This is the key command for understanding "how much did this feature cost to build?"

Step 3: Cost Optimization Suggestions

Based on the report data, suggest optimizations:

• Model downgrade: If Opus is used for template-driven tasks, suggest Sonnet
• Cache utilization: If cache_read_tokens is consistently 0, suggest enabling prompt caching
• Category rebalancing: If research spend exceeds ship-flow spend, suggest more focused sessions
• Budget adjustment: If consistently under 50% budget, suggest reducing to free up capital

Refer to the full optimization playbook in standards/ops/token-cost-tracking.md.

Step 4: Cross-Reference with Unit Economics

When generating reports, if artifacts exist in artifacts/finance/, note how AI COGS relates to:

• Gross margin (AI costs reduce margin)
• CAC (AI-assisted customer acquisition costs)
• Infrastructure costs in the investor template

Typical Session Cost Estimates

Reference ranges for planning (based on Claude Opus/Sonnet pricing):

These are estimates only. Actual costs depend on codebase size, artifact count, and conversation length.

Cross-References

• pricing-unit-economics — consumes AI cost data for infrastructure cost calculations
• ship — advisory integration at Step 7.5 (suggest logging session cost after release)
• investor-reporting-template — AI COGS feeds into burn rate reporting

Quality Checklist

• company.config.yaml was read for budget and tracking config
• Ledger entry has all required fields (timestamp, model, tokens, cost, category)
• Cost calculation matches model rate table (if auto-calculated)
• Budget status uses color coding (green/yellow/red)
• Summary report regenerates cogs/ai-ledger/summary.md
• Empty ledger is handled gracefully
• Feature ID is tagged when applicable (especially during ship flows)
• Cross-reference to unit economics noted when applicable