syndarix/docs/adrs/ADR-012-cost-tracking.md

ADR-012: Cost Tracking and Budget Management

Status: Accepted Date: 2025-12-29 Deciders: Architecture Team Related Spikes: SPIKE-010

---

Context

Syndarix agents make potentially expensive LLM API calls. Without proper cost tracking and budget enforcement, projects could incur unexpected charges. We need:

• Real-time cost visibility
• Per-project budget enforcement
• Cost optimization strategies
• Historical analytics

Decision Drivers

• Visibility: Real-time cost tracking per agent/project
• Control: Budget enforcement with soft/hard limits
• Optimization: Identify and reduce unnecessary costs
• Attribution: Clear cost allocation for billing

Decision

Implement multi-layered cost tracking using:

1. LiteLLM Callbacks for real-time usage capture
2. Redis for budget enforcement
3. PostgreSQL for persistent analytics
4. SSE Events for dashboard updates

Implementation

Cost Attribution Hierarchy

Organization (Billing Entity)
  └── Project (Cost Center)
        └── Sprint (Time-bounded Budget)
              └── Agent Instance (Worker)
                    └── LLM Request (Atomic Cost Unit)

LiteLLM Callback

from litellm.integrations.custom_logger import CustomLogger

class SyndarixCostLogger(CustomLogger):
    async def async_log_success_event(self, kwargs, response_obj, start_time, end_time):
        agent_id = kwargs.get("metadata", {}).get("agent_id")
        project_id = kwargs.get("metadata", {}).get("project_id")
        model = kwargs.get("model")
        cost = kwargs.get("response_cost", 0)
        usage = response_obj.usage

        # Real-time budget check (Redis)
        await self.budget_service.increment(
            project_id=project_id,
            cost=cost,
            tokens=usage.total_tokens
        )

        # Persistent record (async queue to PostgreSQL)
        await self.usage_queue.enqueue({
            "agent_id": agent_id,
            "project_id": project_id,
            "model": model,
            "prompt_tokens": usage.prompt_tokens,
            "completion_tokens": usage.completion_tokens,
            "cost_usd": cost,
            "timestamp": datetime.utcnow()
        })

        # Check budget status
        budget_status = await self.budget_service.check_status(project_id)
        if budget_status == "exceeded":
            await self.notify_budget_exceeded(project_id)

Budget Enforcement

class BudgetService:
    async def check_budget(self, project_id: str) -> BudgetStatus:
        """Check current budget status."""
        budget = await self.get_budget(project_id)
        usage = await self.redis.get(f"cost:{project_id}:daily")

        percentage = (usage / budget.daily_limit) * 100

        if percentage >= 100 and budget.enforcement == "hard":
            return BudgetStatus.BLOCKED
        elif percentage >= 100:
            return BudgetStatus.EXCEEDED
        elif percentage >= 80:
            return BudgetStatus.WARNING
        elif percentage >= 50:
            return BudgetStatus.APPROACHING
        else:
            return BudgetStatus.OK

    async def enforce(self, project_id: str) -> bool:
        """Returns True if request should proceed."""
        status = await self.check_budget(project_id)

        if status == BudgetStatus.BLOCKED:
            raise BudgetExceededException(project_id)

        if status in [BudgetStatus.EXCEEDED, BudgetStatus.WARNING]:
            # Auto-downgrade to cheaper model
            await self.set_model_override(project_id, "cost-optimized")

        return True

Database Schema

CREATE TABLE token_usage (
    id UUID PRIMARY KEY,
    agent_id UUID,
    project_id UUID NOT NULL,
    model VARCHAR(100) NOT NULL,
    prompt_tokens INTEGER NOT NULL,
    completion_tokens INTEGER NOT NULL,
    total_tokens INTEGER NOT NULL,
    cost_usd DECIMAL(10, 6) NOT NULL,
    timestamp TIMESTAMPTZ NOT NULL
);

CREATE TABLE project_budgets (
    id UUID PRIMARY KEY,
    project_id UUID NOT NULL UNIQUE,
    daily_limit_usd DECIMAL(10, 2) DEFAULT 50.00,
    weekly_limit_usd DECIMAL(10, 2) DEFAULT 250.00,
    monthly_limit_usd DECIMAL(10, 2) DEFAULT 1000.00,
    enforcement VARCHAR(20) DEFAULT 'soft',  -- 'soft', 'hard'
    alert_thresholds JSONB DEFAULT '[50, 80, 100]'
);

-- Materialized view for analytics
CREATE MATERIALIZED VIEW daily_cost_summary AS
SELECT
    project_id,
    DATE(timestamp) as date,
    SUM(cost_usd) as total_cost,
    SUM(total_tokens) as total_tokens,
    COUNT(*) as request_count
FROM token_usage
GROUP BY project_id, DATE(timestamp);

Cost Model Prices

Model	Input ($/1M)	Output ($/1M)	Notes
Claude Opus 4.5	$15.00	$75.00	Highest reasoning capability
GPT 5.1 Codex max	$12.00	$60.00	Code generation specialist
Gemini 3 Pro	$3.50	$10.50	Strong multimodal
Gemini 3 Flash	$0.35	$1.05	Fast inference
Qwen3-235B	$2.00	$6.00	Cost-effective (or $0 self-hosted)
DeepSeek V3.2	$0.00	$0.00	Self-hosted, open weights

Cost Optimization Strategies

Strategy	Savings	Implementation
Semantic caching	15-30%	Redis cache for repeated queries
Model cascading	60-80%	Start with Gemini Flash, escalate to Opus
Prompt compression	10-20%	Remove redundant context
Self-hosted fallback	100% for some	DeepSeek V3.2/Qwen3 for non-critical tasks
Task-appropriate routing	40-60%	Route code tasks to GPT 5.1 Codex, simple to Flash

Consequences

Positive

• Complete cost visibility at all levels
• Automatic budget enforcement
• Cost optimization reduces spend significantly
• Real-time dashboard updates

Negative

• Redis dependency for real-time tracking
• Additional complexity in LLM gateway

Mitigation

• Redis already required for other features
• Clear separation of concerns in cost tracking module

Compliance

This decision aligns with:

• FR-801: Real-time cost tracking
• FR-802: Budget configuration (soft/hard limits)
• FR-803: Budget alerts
• FR-804: Cost analytics
• NFR-602: Logging and monitoring (cost observability)
• BR-002: Cost overruns from API usage (risk mitigation)

---

This ADR establishes the cost tracking and budget management architecture for Syndarix.