rrce-workflow 0.3.7 → 0.3.8

package/docs/opencode-guide-optimization-addendum.md

@@ -1,392 +0,0 @@
-# RRCE Optimization Guide - Token Usage Best Practices
-
-**Version:** 2.0 (Optimized)  
-**Date:** January 2026
-
-This addendum documents the token optimization improvements made to RRCE workflow and best practices for efficient usage.
-
----
-
-## 🎯 What Changed?
-
-### Token Reduction Summary
-
-| Component | Before | After | Reduction |
-|-----------|--------|-------|-----------|
-| Research Prompt | 332 lines (~15K tokens) | 80 lines (~4K tokens) | **73%** |
-| Planning Prompt | 307 lines (~14K tokens) | 80 lines (~4K tokens) | **71%** |
-| Executor Prompt | 300+ lines (~14K tokens) | 100 lines (~5K tokens) | **64%** |
-| Orchestrator Prompt | 351 lines (~16K tokens) | 120 lines (~5K tokens) | **69%** |
-
-### Overall Workflow Improvements
-
-- **Full workflow token usage**: 150K → 53K tokens (**65% reduction**)
-- **Cost per workflow**: $0.45 → $0.16 (**64% savings**)
-- **Latency**: ~40% faster (fewer round-trips, smaller prompts)
-
----
-
-## ✅ Recommended Usage Patterns
-
-### **Pattern 1: Direct Subagent Invocation** (Recommended for 90% of use cases)
-
-For most work, invoke subagents **directly** using `@rrce_*` syntax:
-
-#### Research Phase
-```
-@rrce_research_discussion TASK_SLUG=user-auth REQUEST="Add JWT-based authentication"
-```
-
-**Benefits:**
-- 70% fewer tokens (no orchestrator overhead)
-- Better prompt caching
-- More interactive control
-- Faster responses
-
-#### Planning Phase
-```
-@rrce_planning_discussion TASK_SLUG=user-auth
-```
-
-**Prerequisites:** Research must be complete
-
-#### Execution Phase
-```
-@rrce_executor TASK_SLUG=user-auth
-```
-
-**Prerequisites:** Research AND planning must be complete
-
----
-
-### **Pattern 2: Orchestrator** (Only for full automation)
-
-Use orchestrator **only** when you want complete hands-off automation:
-
-```
-@rrce_orchestrator "Implement user authentication feature from research to deployment"
-```
-
-**The orchestrator will:**
-1. Auto-detect required phases (research → plan → execute)
-2. Pre-fetch context once (avoids redundant searches)
-3. Use session reuse for caching (60-80% token reduction)
-4. Auto-progress through phases without prompts
-5. Return final synthesized results
-
-**When to use orchestrator:**
-- Implementing complete features end-to-end
-- Want zero-interaction automation
-- Running batch workflows
-
-**When NOT to use orchestrator:**
-- Single-phase work (just research/planning)
-- Interactive workflows (want to review each phase)
-- Debugging or iterative development
-
----
-
-## 🔥 New Features: Session Reuse & Smart Caching
-
-### Session Reuse
-
-Agents now support **session continuity** across multiple invocations:
-
-```
-# First invocation
-@rrce_research_discussion TASK_SLUG=feature-x REQUEST="..."
-
-# Agent responds with questions...
-# You answer in the SAME chat session...
-
-# Second response uses CACHED prompt (90% reduction!)
-```
-
-**How it works:**
-- OpenCode automatically assigns `promptCacheKey` = `sessionID`
-- After first turn, system prompt is cached
-- Subsequent turns only send new user messages
-- Works with Anthropic Claude (cache_control) and OpenAI (prompt_cache_key)
-
-**Example token usage:**
-- Turn 1: 4K prompt + 1K user = 5K tokens
-- Turn 2: 0.4K prompt (cached!) + 1K user = 1.4K tokens
-- Turn 3: 0.4K prompt (cached!) + 1K user = 1.4K tokens
-- **Total:** 8K tokens (vs. 15K without caching)
-
----
-
-### Smart Knowledge Caching
-
-Agents now cache knowledge searches:
-
-**Old behavior (inefficient):**
-```
-Turn 1: Search knowledge → Ask questions
-Turn 2: Search knowledge AGAIN → Ask more questions
-Turn 3: Search knowledge AGAIN → Generate brief
-```
-
-**New behavior (optimized):**
-```
-Turn 1: Search knowledge ONCE → Store results → Ask questions
-Turn 2: Reference cached findings → Ask more questions
-Turn 3: Reference cached findings → Generate brief
-```
-
-**Savings:** ~5K tokens per session (no redundant searches)
-
----
-
-## 💰 Model Configuration (User Choice)
-
-**RRCE does NOT force specific models.** You choose what works for you!
-
-The optimization works with **ANY model** you configure in OpenCode. What we enable:
-
-| Feature | What It Does | Works With |
-|---------|--------------|------------|
-| **Provider Caching** | `setCacheKey: true` for all providers | Anthropic, OpenAI, OpenRouter, Google |
-| **Prompt Caching** | Automatic via `promptCacheKey = sessionID` | Any model that supports caching |
-| **Slim Prompts** | 70-93% smaller prompts | ALL models |
-| **Session Reuse** | Orchestrator reuses sessions | ALL models |
-
-### Recommended Models (Optional)
-
-If you want cost optimization, consider:
-
-| Agent | Recommendation | Rationale |
-|-------|----------------|-----------|
-| **Research** | Haiku/Mini | Q&A doesn't need heavy reasoning |
-| **Planning** | Sonnet/GPT-4o | Task breakdown needs reasoning |
-| **Executor** | Sonnet/GPT-4o | Code generation needs power |
-
-**To set models per agent (optional):**
-```json
-{
-  "agent": {
-    "rrce_research_discussion": {
-      "model": "anthropic/claude-haiku-4-20250514"
-    }
-  }
-}
-```
-
-**Configuration location:** `opencode.json` in project root or `~/.config/opencode/opencode.json`
-
----
-
-## 🚀 Best Practices for Token Efficiency
-
-### 1. Use Direct Subagent Invocation
-
-**❌ Inefficient:**
-```
-@rrce_orchestrator "Research user authentication requirements"
-```
-Token cost: ~20K (orchestrator overhead + delegation + research)
-
-**✅ Efficient:**
-```
-@rrce_research_discussion TASK_SLUG=user-auth REQUEST="Research authentication"
-```
-Token cost: ~5K (direct invocation, prompt caching)
-
----
-
-### 2. Answer Questions in Same Session
-
-**❌ Inefficient:** Starting new chat for each answer
-- Each new chat = full prompt reload
-- No caching benefit
-
-**✅ Efficient:** Continue in same session
-- Prompt cached after turn 1
-- 90% token reduction on subsequent turns
-
----
-
-### 3. Let Agents Cache Knowledge
-
-**❌ Inefficient:** Asking agent to "re-search" on each turn
-
-**✅ Efficient:** Trust the hybrid approach
-- Agent searches once (first turn)
-- References findings thereafter
-- Only re-searches if you introduce new scope
-
----
-
-### 4. Use Orchestrator for Full Workflows Only
-
-**❌ Inefficient:** Using orchestrator for single phases
-```
-@rrce_orchestrator "Just do research on X"
-```
-
-**✅ Efficient:** Direct invocation for single phases
-```
-@rrce_research_discussion TASK_SLUG=x REQUEST="Research X"
-```
-
-**✅ Efficient:** Orchestrator for full automation
-```
-@rrce_orchestrator "Implement feature X from research to deployment"
-```
-
----
-
-## 📊 Measuring Token Usage
-
-### Via OpenCode Logs
-
-Token usage is logged in each API response. Look for:
-
-```
-Input tokens: 4,234
-Cache creation tokens: 3,800  (first turn only)
-Cache read tokens: 3,800      (subsequent turns)
-Output tokens: 512
-```
-
-**Calculation:**
-- First turn cost: `input_tokens + output_tokens`
-- Cached turn cost: `(input_tokens - cache_read_tokens) + output_tokens`
-
-### Via Provider Dashboard
-
-- **Anthropic Console:** https://console.anthropic.com → Usage
-- **OpenAI Dashboard:** https://platform.openai.com/usage
-
----
-
-## 🔧 Troubleshooting
-
-### "Agent uses too many tokens"
-
-**Check:**
-1. Are you using direct invocation? (More efficient than orchestrator)
-2. Are you continuing in same session? (Enables caching)
-3. Is agent re-searching knowledge? (Should only search once)
-
-**Fix:**
-- Use `@rrce_*` directly for single phases
-- Keep conversations in same session
-- Review agent config (`opencode.json`) - verify Haiku for research
-
----
-
-### "Cache not activating"
-
-**Symptoms:** Token usage stays high on subsequent turns
-
-**Common causes:**
-1. **New sessions:** Each new chat = new session = no cache
-2. **Prompt modifications:** Changing agent config invalidates cache
-3. **Different models:** Each model has separate cache
-
-**Fix:**
-- Continue in same session (don't start new chat)
-- Avoid editing agent prompts mid-session
-- Verify `promptCacheKey` in logs
-
----
-
-### "Research agent asks too many questions"
-
-**Default:** Hybrid approach (ask critical questions, document rest as assumptions)
-
-**If too verbose:**
-Edit `agent-core/prompts/research_discussion.md`:
-```markdown
-**STOP after 1 round.** Document remaining ambiguity as assumptions.
-```
-
-**If too brief:**
-```markdown
-**Up to 3 rounds** if needed for critical clarification.
-```
-
----
-
-## 📈 Performance Benchmarks
-
-### Before Optimization (v1.0)
-
-| Workflow | Tokens | Cost | Time |
-|----------|--------|------|------|
-| Research only (3 rounds) | 66K | $0.20 | ~45s |
-| Research → Planning | 110K | $0.33 | ~75s |
-| Full workflow | 150K | $0.45 | ~120s |
-
-### After Optimization (v2.0)
-
-| Workflow | Tokens | Cost | Time |
-|----------|--------|------|------|
-| Research only (3 rounds) | 16K | $0.004 | ~25s |
-| Research → Planning | 35K | $0.11 | ~40s |
-| Full workflow | 53K | $0.16 | ~70s |
-
-**Improvements:**
-- 76% fewer tokens for research
-- 68% fewer tokens for planning
-- 65% fewer tokens for full workflow
-- 98% cost reduction for research (Haiku!)
-- 42% faster execution
-
----
-
-## 🎓 Advanced Optimizations
-
-### Custom Agent Configuration
-
-Override defaults in `opencode.json`:
-
-```json
-{
-  "agent": {
-    "rrce_research_discussion": {
-      "model": "anthropic/claude-opus-4",  // More powerful if needed
-      "temperature": 0.1,                  // More deterministic
-      "maxSteps": 3                        // Limit iterations
-    }
-  }
-}
-```
-
-### Project-Specific Settings
-
-Create `.opencode/opencode.json` in your project:
-
-```json
-{
-  "agent": {
-    "rrce_research_discussion": {
-      "model": "anthropic/claude-haiku-4-20250514"
-    }
-  }
-}
-```
-
-This overrides global settings for this project only.
-
----
-
-## 📚 Related Documentation
-
-- [Main RRCE Guide](./opencode-guide.md)
-- [Architecture Documentation](./architecture.md)
-- [Migration Guide](./MIGRATION-v2.md)
-- [OpenCode Docs](https://opencode.ai/docs)
-
----
-
-## 🤝 Contributing
-
-Found additional optimizations? Submit a PR or issue:
-- GitHub: https://github.com/rryando/rrce-workflow
-
----
-
-**Last Updated:** January 2026  
-**Version:** 2.0