amr/org-agent-contrib
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
5f2b7e8d1659f0d76ed4a6f55bff5daa5af9a75d
org-agent-contrib/org-skill-token-accountant.org

16 KiB
Raw Blame History

SKILL: Token Accountant Agent (Universal Literate Note) Token Management & Model Optimization Research Token Optimization - $50 Monthly Budget Token Optimization Token Optimization - Quick Start Token Optimization Strategy

Overview

The Token Accountant is the governor of the Neural Engine. It manages the cost, reliability, and routing of LLM providers. Its primary mission is to ensure the PSF operates at maximum intelligence with minimum marginal cost by aggressively prioritizing subsidized free models when appropriate.

Phase A: Demand (PRD)

1. Purpose

Autonomously manage the provider cascade and model selection to optimize for cost, speed, and reliability.

Phase B: Blueprint (PROTOCOL)

1. Architectural Intent

Maintain a state-aware provider cascade that routes around "pain" (failures) and dynamically selects models based on task complexity.

2. Semantic Interfaces

Routing and Pain Management 

(in-package :opencortex)

(defvar *provider-pain-table* (make-hash-table :test 'equal))

(defun token-accountant-record-pain (provider)
  "Marks a provider as 'pained' (failed). It will be de-prioritized."
  (setf (gethash provider *provider-pain-table*) (+ (get-universal-time) 600)) ; 10 min penalty
  (harness-log "ACCOUNTANT - Provider ~a de-prioritized due to failure." provider))

(defun token-accountant-get-cascade (context)
  "Returns a dynamic list of providers, routing around pained ones. Uses standardized gateway keywords."
  (let ((all-providers '(:openrouter :groq :gemini-api :ollama))
        (healthy nil)
        (pained nil)
        (now (get-universal-time)))
    (dolist (p all-providers)
      (if (> (or (gethash p *provider-pain-table*) 0) now)
          (push p pained)
          (push p healthy)))
    (append (nreverse healthy) (nreverse pained))))

(defun token-accountant-get-model-for-provider (provider &optional context)
  "Returns the recommended model for the provider, prioritizing free/subsidized models. Updated April 2026."
  (let ((complexity (ignore-errors (uiop:symbol-call :opencortex.skills.org-skill-router :router-classify-complexity context))))
    (case provider
      (:openrouter
       (case complexity
         (:REASONING "meta-llama/llama-3.3-70b-instruct:free") ; High fidelity, zero cost
         (:COGNITION "qwen/qwen3.6-plus:free")               ; Latest interaction, zero cost
         (t "meta-llama/llama-3.2-3b-instruct:free")))       ; Ultra-fast reflex, zero cost
      (:groq
       (case complexity
         (:REASONING "llama-3.3-70b-versatile")
         (t "llama-3.1-8b-instant")))
      (:gemini-api
       "gemini-1.5-flash-latest")
      (t nil))))

(defun token-accountant-patch-kernel ()
  "Hot-patches the harness's cascade and model selector to use our dynamic logic."
  (setf opencortex:*provider-cascade* #'token-accountant-get-cascade)
  (setf opencortex::*model-selector-fn* #'token-accountant-get-model-for-provider))

Registration 

(progn
  (token-accountant-patch-kernel)
  (defskill :skill-token-accountant
    :priority 100
    :trigger (lambda (context) 
               (let ((sensor (getf (getf context :payload) :sensor)))
                 (or (eq sensor :tool-error) (eq sensor :cost-audit))))
    :neuro (lambda (context) nil)
    :symbolic (lambda (action context) 
                (let ((p (getf (getf context :payload) :provider)))
                  (when p (token-accountant-record-pain p))
                  action))))

Documentation (Token Optimization)

research.org

Token Management Strategy Research

Initial Findings

OpenRouter Free Tier

Google AI Studio (Gemini)

  • Free tier available
  • Limits: 60 requests/minute, 300K tokens/day
  • No credit card required
  • Every API key gets these limits

Research Questions

  1. Which providers offer free or low-cost tiers?
  2. What are the rate limits and quotas?
  3. Which models are best for which use cases?
  4. How to optimize context windows?
  5. What is the cost per token breakdown?

To Research Further

Provider Free Tier Paid Tier Best For
Google Gemini 300K tokens/day Pay per use? General, coding
OpenRouter Varies by model Per-request Routing, variety
OpenAI ? ? GPT-4 quality
Anthropic ? ? Claude capabilities
Mistral ? ? Open weights
Local Hardware cost Free Privacy, control

Token Optimization Strategies to Explore

  1. Tiered Model Usage

    • Simple tasks: Fast/cheap models
    • Complex tasks: Stronger models
    • Fallback: Lower tier if higher fails

  2. Context Compression

    • Summarize long contexts
    • Use RAG instead of full context
    • Prune old conversation

  3. Caching

    • Cache common responses
    • Reuse embeddings
    • Batch requests

  4. Hybrid Approach

    • Local models for simple queries
    • Cloud APIs for complex tasks
    • Manual review for critical outputs

X Account Access

Pending: X account access via Google login Blocker: Requires OTP from user per security rule (SOUL.md) Action needed: User provides OTP, I complete OAuth, access bookmarks

budget-50.org

Budget: $50/Month

Budget Breakdown

Tier Provider Allocation Tokens Est. Use Case
FREE Google Gemini $0 ~9M/month 90% of work
CHEAP OpenRouter $20 ~6M tokens Fallback, complex tasks
PREMIUM Claude/GPT-4o $25 ~500K tokens Critical decisions
BUFFER Various $5 Emergency Overruns, testing

Daily Free Allowance

  • Google Gemini: 300K tokens/day = 9M/month = $0
  • This covers 90-95% of expected workload

Paid Tier Allocation ($45)

  • $20 → OpenRouter (Qwen, Mistral, Llama)

    • ~6M tokens at $0.003/1K
    • Use when: Gemini rate limited, need different model

  • $25 → Premium models (Claude, GPT-4o)

    • ~500K tokens at $0.05/1K average
    • Use when: Architecture decisions, critical code review, final validation

  • $5 → Buffer

    • Handle overruns
    • Emergency access
    • Testing new models

Hard Limits

Provider Monthly Cap Alert At
OpenRouter $20 $16 (80%)
Premium $25 $20 (80%)
Total $50 $45 (90%)

Daily Tracking

Target: Monitor consumption every session

``` IF daily_cost > $1.50: → Switch to Gemini only → Defer premium tasks

IF weekly_cost > $12: → Review usage patterns → Find optimization opportunities ```

Emergency Protocol

If approaching $50 limit before month end:

  1. Halt all paid API calls
  2. Switch to Gemini-only mode
  3. Queue premium tasks for next month
  4. Consider local inference setup

Cost-Per-Task Guidelines

Task Type Max Cost Preferred Model
Quick lookup $0.00 Gemini
Code review $0.01 Gemini/OpenRouter
Feature design $0.05 OpenRouter
Architecture review $0.10 Claude/GPT-4o
Emergency debug $0.20 Best available

Optimization Imperative

With $50/month, waste is not affordable:

  • ❌ No speculative queries
  • ❌ No "just curious" premium calls
  • ❌ No repeated similar prompts
  • ✅ Always use Gemini first
  • ✅ Batch similar requests
  • ✅ Cache embeddings locally
  • ✅ Summarize long contexts

Monthly Review

  1. Compare actual vs. projected usage
  2. Adjust model routing rules
  3. Identify expensive query patterns
  4. Plan next month's allocation

Break-Even Analysis

At $50/month = $600/year:

  • Option A: Continue APIs (flexible, managed)

  • Option B: Local inference (~$800 hardware, $0 ongoing)

    • Break-even: 16 months
    • Risk: Hardware failure, maintenance

Recommendation: Stick with APIs until $100+/month, then evaluate hardware.

Questions for Human Partner

  1. Is $50 firm or flexible in emergencies?
  2. What happens if we hit limit mid-critical-task?
  3. Preference for which premium model? (Claude vs GPT-4 vs both)
  4. Should I track and report costs per project?
  5. Any tasks that are "unlimited budget" critical?

README.org

Cost-effective LLM usage through smart routing, context compression, and multi-provider strategies.

Token Optimization

Strategy and implementation for minimizing LLM costs while maintaining quality.

Project Tasks

See the actionable tasks for this project in GTD.org > Projects > Token Optimization

Key Documents

Current Focus

  • Multi-provider setup (Gemini primary, OpenRouter fallback)
  • Usage tracking and budget alerts
  • Smart routing by task type
  • Context compression techniques

quick-start.org

Quick Reference for Daily Use

Rule of Thumb

What you need Use this Cost
Quick answer, formatting, lookup Gemini Flash FREE
Code review, analysis Gemini Pro FREE
Complex problem solving Claude Haiku / Qwen $
Critical architecture decision GPT-4o $$

Free Tier Limits (Daily)

Provider Tokens Requests Reset
Google AI Studio 300,000 60/min Daily
OpenRouter Free Varies Limited -

Current Recommendation

→ Use Google Gemini exclusively until hitting 250K tokens/day → Then add OpenRouter fallback → Only use GPT-4 for final reviews

This will reduce token costs by ~90%

Next Steps

  1. Configure Gemini as primary (already partially done)
  2. Add quota tracking
  3. Set alerts at 80% of free limits
  4. Implement tiered routing

Savings Potential: $100-500/month → $10-50/month

plan.org

Executive Summary

Goal: Minimize inference costs while maximizing capability

Current approach: Single default model → Multi-tier, multi-provider strategy

Three-Tier Model Strategy

Tier 1: Fast/Cheap (80% of queries)

  • Purpose: Simple tasks, formatting, lookups
  • Models: Google Gemini Flash, Local models
  • Cost: $0-0.000001 per 1K tokens
  • Speed: Fastest

Tier 2: Balanced (18% of queries)

  • Purpose: Complex reasoning, code generation, analysis
  • Models: Gemini Pro, Claude Haiku, Llama 3 70B
  • Cost: $0.0001-0.003 per 1K tokens
  • Speed: Medium

Tier 3: High-Performance (2% of queries)

  • Purpose: Critical decisions, complex architecture, final review
  • Models: GPT-4, Claude Opus, Gemini Ultra
  • Cost: $0.01-0.03 per 1K tokens
  • Speed: Slower

Provider Analysis

Google AI Studio (Primary Recommended)

Model Free Tier Rate Limit Best For
Gemini 2.0 Flash 300K tokens/day 60 req/min Quick tasks, coding
Gemini 1.5 Flash 300K tokens/day 60 req/min Fast responses
Gemini 1.5 Pro 300K tokens/day 60 req/min Complex tasks

Cost: FREE (within limits)

OpenRouter.Aggregated (Secondary)

Model Price/1K tokens Context Reliability
Qwen 3 235B $0.0001-0.0003 128K High
Mistral Large $0.002-0.006 128K High
Llama 4 405B $0.0002-0.0005 128K Medium
Free tier models $0 Varies Variable

OpenAI (Tier 3 only)

  • GPT-4: $0.03/1K tokens (expensive)
  • GPT-4o: $0.005/1K tokens (better value)
  • Use sparingly for critical tasks only

Local Inference (Long-term goal)

  • Hardware: $1000-5000 initial investment
  • Ongoing: $0 (electricity only)
  • Models: Llama 3, Mistral, DeepSeek
  • Best for: High-volume, privacy-sensitive work

Context Optimization Strategies

1. Context Windows by Task Type

Task Type Optimal Context Compression Savings
Code review 4K-8K Truncate old files 50%
Documentation 8K-16K Summarize sections 30%
Research 16K-32K Chunk + RAG 70%
Architecture 32K-128K Maintain full 0%

2. Conversation Pruning

  • Remove "thinking" blocks from history
  • Summarize conversation every 10 turns
  • Archive old sessions to external storage

3. RAG vs. Full Context

  • Rule: < 5K tokens of context → Full
  • Rule: > 10K tokens of context → Use embeddings/RAG
  • Savings: 60-80% on large document tasks

Request Optimization

Batching Strategy

  • Group similar requests (3-5 per batch)
  • Same model, same parameters
  • Shared overhead costs

Caching Strategy

  • Cache embeddings for repeated contexts
  • Store common completions (templates)
  • Reuse code snippet suggestions

Streaming vs. Non-Stream

  • Streaming: Better UX, but higher token overhead
  • Non-stream: More efficient for programmatic use
  • Recommendation: Non-stream for background tasks

Smart Routing Rules

Automatic Selection Logic

``` IF task_type == "simple_lookup" OR "formatting": → Gemini Flash (free)

ELIF task_type == "code_generation" AND complexity < 3: → Gemini Pro (free tier)

ELIF task_type == "complex_reasoning" OR "architecture": → Claude Sonnet or GPT-4o

ELIF task_type == "final_review" OR "critical_decision": → GPT-4 or Claude Opus ```

Fallback Chain

  1. Try Gemini (free)
  2. If rate limited → OpenRouter (cheap)
  3. If quality insufficient → GPT-4o
  4. If critical failure → GPT-4

Concrete Implementation

Config Structure (openclaw.json)

```json { "models": { "defaults": { "primary": "google-gemini-cli/gemini-2.0-flash", "fallbacks": [ "openrouter/qwen/qwen3-235b-a22b", "google-gemini-cli/gemini-1.5-pro", "openai/gpt-4o" ] }, "providers": { "google-gemini-cli": { "freeTier": true, "dailyLimit": 300000, "rateLimit": 60 }, "openrouter": { "freeTierModels": ["openrouter/auto"], "budgetLimit": 500 }, "openai": { "budgetLimit": 200, "useFor": ["critical", "architecture"] } } } } ```

Monitoring & Alerts

  • Track daily token usage per provider
  • Alert at 80% of free tier limits
  • Monthly budget review and adjustment

Cost Projections

Current Unknown Usage → Optimized

Scenario Monthly Tokens Current Cost Optimized Cost Savings
Light (< 1M) 1M $50-100 $0-10 90%
Medium (1-5M) 3M $200-500 $20-100 80%
Heavy (5-20M) 10M $1000-3000 $200-500 80%

Immediate Actions

Week 1: Setup

  • Configure Gemini as primary provider
  • Set up OpenRouter fallback
  • Implement basic usage tracking
  • Document current baseline

Week 2: Implement

  • Add smart routing logic
  • Implement context compression
  • Set up budget alerts
  • A/B test model choices

Week 3: Optimize

  • Analyze usage patterns
  • Fine-tune routing rules
  • Tune context windows
  • Document findings

Week 4: Scale

  • Full multi-provider setup
  • Implement full caching
  • Maximize free tier usage
  • Plan for paid tiers if needed

Long-term: Local Inference Path

Minimum Viable Setup

  • Hardware: RTX 4090 or Apple Silicon M3 Max
  • Software: Ollama + OpenClaw integration
  • Cost: ~$2000-4000 one-time
  • Break-even: 3-6 months vs. API costs

Full Self-Hosted

  • Hardware: Dual RTX 4090 or 2x Mac Studio
  • Models: Llama 3 70B, Mixtral 8x22B
  • Cost: ~$8000-12000
  • For: Privacy, unlimited inference, control
Reference in New Issue View Git Blame Copy Permalink