chore: verify and stabilize main branch (#42)

* chore: update package-lock.json after npm install

* chore: remove unused ajv-cli, keep vitest at stable 1.6.1

- Removed ajv-cli (unused dev dependency, had high severity vulnerability)
- Kept vitest at 1.6.1 (stable, upgrading to v4 requires extensive refactoring)
- Reduced vulnerabilities from 6 to 4 (all moderate, dev-only)
- All remaining vulnerabilities are in esbuild/vite (dev server only, no production impact)
- Build and tests verified working

* feat(evals): add execution-balance validation to 21 OpenAgent tests and enhance agent system

Major improvements to evaluation framework and agent capabilities:

## Execution Balance Evaluator Integration
- Added execution-balance rule to 21 OpenAgent tests (2.7% → 31.5% coverage)
- Validates "read before execute" pattern across critical rules, workflows, and edge cases
- Tests now check: read operations before execution tools, healthy read/exec ratio (≥1.0)
- Coverage includes: approval-gate, context-loading, stop-on-failure, report-first tests

## Agent System Enhancements
- Enhanced OpenCoder agent with critical rules enforcement (approval gates, context loading)
- Added structured delegation rules and workflow phases
- Improved error handling with REPORT→PROPOSE→REQUEST→FIX pattern

## New Agent Creation System
- Added research-backed agent creation templates and guides
- Includes 8 test templates covering planning, context, incremental execution, tools
- Automated agent scaffolding with /create-agent command
- Complete documentation in .opencode/command/openagents/new-agents/

## OpenCoder Test Suite
- Added 8 comprehensive test cases for OpenCoder agent
- Tests cover: planning, context-loading, delegation, error-handling, implementation
- Includes DEBUG_GUIDE.md and QUICK_TEST_GUIDE.md for developers

## Documentation & Tooling
- Added DEVELOPMENT.md guide for contributors
- Enhanced CONTRIBUTING.md with quick reference and common commands
- Improved event-logger with DEBUG_VERBOSE mode for detailed output
- Added debug scripts: show-test-conversation.sh, run-test-verbose.sh
- Removed deprecated EXTERNAL_PR_GUIDE.md

## Version Bump
- Bumped version to 0.0.3 (patch)

Files changed: 60+ files (21 test updates, 16 new agent templates, 8 new tests, docs, tooling)

* feat(ci): add PR title validation for semantic versioning

Add automated PR title validation to ensure conventional commit format.
This enables proper automatic version bumping when PRs are merged.

Features:
- Validates PR titles against conventional commit patterns
- Shows expected version bump (major/minor/patch)
- Posts helpful comment with examples if validation fails
- Supports all conventional commit types (feat, fix, docs, etc.)
- Supports breaking changes (feat!, fix!)
- Supports pre-release tags ([alpha], [beta], [rc])

When PR titles follow the format, version bumping works correctly:
- feat: → minor bump (0.3.0 → 0.4.0)
- fix: → patch bump (0.3.0 → 0.3.1)
- feat!: → major bump (0.3.0 → 1.0.0)

* fix(ci): improve PR checks workflow with sequential execution

Reorganize PR checks to run in logical sequence and skip unnecessary checks:

Changes:
- Run PR title validation first (fast, always required)
- Detect changed files to determine which checks to run
- Only run build checks if evals/ files changed
- Add comprehensive summary job showing all check results
- Prevent cache errors when evals files not changed

Benefits:
- Faster PR checks (skip unnecessary builds)
- Clear sequential execution (title → changes → build)
- Better error messages and summaries
- Reduced CI/CD costs

* fix(ci): resolve PR #42 issues and refactor prompt architecture

## Fixed PR #42 Issues

1. **Prompt Validation** - Refactored to new architecture where agent files are canonical defaults
2. **Package-lock.json** - Fixed workflow to use root package-lock (npm workspaces)
3. **Git Merge Base** - Added fetch-depth: 0 to PR checks workflow

## New Prompt Architecture

**Before:**
- Agent files had to match .opencode/prompts/<agent>/default.md
- Caused validation failures when updating prompts
- Redundant duplication

**After:**
- Agent files (.opencode/agent/*.md) = Canonical defaults (source of truth)
- Prompt variants (.opencode/prompts/<agent>/<model>.md) = Model-specific optimizations
- No more default.md files needed

## Changes

### Workflows
- pr-checks.yml: Added fetch-depth: 0 for full git history
- validate-test-suites.yml: Fixed package-lock path for npm workspaces
- validate-registry.yml: Updated validation messaging

### Scripts
- validate-pr.sh: Validates prompt library structure (rejects default.md files)
- use-prompt.sh: Handles 'default' as agent file, variants as model-specific
- test-prompt.sh: Tests default variant correctly, saves results to prompts/results/

### Documentation
- .opencode/prompts/README.md: Updated architecture explanation
- docs/contributing/CONTRIBUTING.md: Updated prompt workflow
- scripts/development/demo.sh: Updated structure display

## Results

- 12 files changed, 231 insertions(+), 811 deletions(-)
- Net reduction: 580 lines (cleaner codebase)
- All validations passing
- Results directory structure preserved
- Backwards compatible with existing test results

.github/EXTERNAL_PR_GUIDE.md

@@ -1,174 +0,0 @@
-# External PR Guide
-
-This guide explains how GitHub Actions workflows behave for external contributors (fork PRs).
-
-## For External Contributors (Fork PRs)
-
-When you submit a PR from a fork, workflows run but **cannot auto-commit fixes** due to GitHub security restrictions.
-
-### What Runs Automatically
-
-✅ **Fast Build Check (< 2 minutes)**
-- TypeScript compilation check
-- YAML test suite validation
-- Quick feedback on code quality
-
-✅ **Registry Validation**
-- Checks if `registry.json` matches actual files
-- Validates all paths are correct
-- Checks prompts use defaults
-- **Cannot auto-commit** (you'll need to fix locally)
-
-❌ **What Doesn't Run on PRs**
-- Full AI agent tests (too expensive, run manually by maintainers)
-- Auto-version bumping (happens after merge)
-
-### What You Need to Do
-
-If the validation workflow reports issues, you'll need to fix them locally:
-
-#### 1. Build Failures
-
-If TypeScript compilation fails:
-
-```bash
-# Build locally to see errors
-cd evals/framework
-npm install
-npm run build
-
-# Fix the errors, then commit
-git add .
-git commit -m "fix: resolve build errors"
-git push
-```
-
-#### 2. Registry Issues
-
-If new components are detected (you'll get a comment on your PR):
-
-```bash
-# Run auto-detect locally
-./scripts/registry/auto-detect-components.sh --auto-add
-
-# Commit the updated registry
-git add registry.json
-git commit -m "chore: update registry with new components"
-git push
-```
-
-#### 3. Prompt Issues
-
-If prompts don't match defaults:
-
-```bash
-# Restore default prompts
-./scripts/prompts/use-prompt.sh <agent-name> default
-
-# Commit the changes
-git add .opencode/agent/
-git commit -m "chore: restore default prompts"
-git push
-```
-
-#### 4. Path Issues
-
-If registry paths are invalid:
-
-```bash
-# Validate and see suggestions
-./scripts/validate-registry.sh --fix
-
-# Manually fix paths in registry.json
-# Then commit
-git add registry.json
-git commit -m "fix: correct registry paths"
-git push
-```
-
-### What Happens Next
-
-1. **Fix Issues Locally**: If validation fails, fix issues and push
-2. **Fast Feedback**: Build checks complete in < 2 minutes
-3. **Maintainer Review**: A maintainer will review your code
-4. **Merge**: Once everything passes, your PR will be merged!
-5. **Auto-Release**: After merge, version is auto-bumped and release created (you don't need to do anything!)
-
----
-
-## For Maintainers
-
-### Handling External PRs
-
-External PRs run normally but **cannot auto-commit** due to GitHub security (can't push to fork branches).
-
-### Simple Overrides
-
-If you need to override checks:
-
-**Skip Validation:**
-1. Go to **Actions** → **Validate Registry on PR**
-2. Click **"Run workflow"**
-3. Enable **"Skip validation checks"**
-4. Run on the PR branch
-
-**Skip Version Bump:**
-1. Go to **Actions** → **Post-Merge Automation**
-2. Click **"Run workflow"**
-3. Enable **"Skip version bump"**
-4. Run manually
-
-### Workflow Behavior
-
-| PR Type | Build Check | Registry Validation | Auto-Commit | AI Tests |
-|---------|------------|-------------------|-------------|----------|
-| **Internal PR** (branch) | ✅ Yes | ✅ Yes | ✅ Yes | ❌ No* |
-| **External PR** (fork) | ✅ Yes | ✅ Yes | ❌ No** | ❌ No* |
-
-*AI tests don't run on PRs (too expensive) - maintainers run manually if needed  
-**Cannot push to fork branches - contributor must fix locally
-
-### Manual Testing External PRs
-
-If you want to test an external PR locally:
-
-```bash
-# Fetch the PR
-gh pr checkout <PR_NUMBER>
-
-# Run build check
-cd evals/framework
-npm install
-npm run build
-npm run validate:suites:all
-
-# Run validation
-./scripts/registry/validate-registry.sh -v
-./scripts/prompts/validate-pr.sh
-
-# Auto-fix registry if needed
-./scripts/registry/auto-detect-components.sh --auto-add
-
-# Run AI tests (optional - only if needed)
-npm run test:ci
-
-# If everything passes, approve and merge
-gh pr review <PR_NUMBER> --approve
-gh pr merge <PR_NUMBER>
-```
-
-**Note:** AI tests are optional for PRs. The build check is usually sufficient.
-
----
-
-## Workflow Files
-
-- **`.github/workflows/pr-checks.yml`**: Fast build validation (< 2 min)
-- **`.github/workflows/validate-registry.yml`**: Registry validation with override
-- **`.github/workflows/post-merge.yml`**: Auto-versioning after merge (main only)
-
-## Questions?
-
-- **Contributors**: See [CONTRIBUTING.md](../docs/contributing/CONTRIBUTING.md)
-- **Maintainers**: See [WORKFLOW_GUIDE.md](../WORKFLOW_GUIDE.md)
-- **Issues**: Open a [GitHub Issue](https://github.com/darrenhinde/OpenAgents/issues)

.github/workflows/README.md

@@ -1,207 +0,0 @@
-# Workflow Guide
-
-This directory contains GitHub Actions workflows for automated testing, validation, and releases.
-
-## Workflow Overview
-
-### For Contributors (PR Workflows)
-
-#### `pr-checks.yml` - Fast Build Validation
-**Triggers:** Pull requests to main/dev with evals changes  
-**Duration:** < 2 minutes  
-**What it does:**
-- ✅ TypeScript compilation check
-- ✅ YAML test suite validation
-- ✅ Fast feedback on code quality
-
-**Fork-friendly:** ✅ Yes - read-only checks
-
----
-
-#### `validate-registry.yml` - Registry Validation
-**Triggers:** Pull requests to main/dev  
-**Duration:** < 1 minute  
-**What it does:**
-- ✅ Validates registry.json paths
-- ✅ Auto-detects new components
-- ✅ Validates prompts use defaults
-- ✅ **Internal PRs:** Auto-commits registry updates
-- ✅ **Fork PRs:** Posts comment with instructions
-
-**Fork-friendly:** ✅ Yes - detects forks and provides guidance
-
----
-
-### For Maintainers (Post-Merge Workflows)
-
-#### `post-merge.yml` - Auto-Versioning & Releases
-**Triggers:** Push to main branch  
-**Duration:** < 1 minute  
-**What it does:**
-- ✅ Auto-bumps version based on conventional commits
-- ✅ Updates CHANGELOG.md
-- ✅ Creates git tag
-- ✅ Publishes GitHub release
-
-**Conventional Commit Examples:**
-```bash
-feat: add new feature        # → Minor bump (0.1.0 → 0.2.0)
-fix: bug fix                 # → Patch bump (0.1.0 → 0.1.1)
-feat!: breaking change       # → Major bump (0.1.0 → 1.0.0)
-[alpha] experimental         # → Alpha bump (0.1.0-alpha.1)
-```
-
-**Manual override:** Can skip version bump via workflow_dispatch
-
----
-
-#### `update-registry.yml` - Registry Auto-Update
-**Triggers:** Push to main with .opencode changes  
-**Duration:** < 1 minute  
-**What it does:**
-- ✅ Auto-detects new components
-- ✅ Updates registry.json
-- ✅ Commits changes to main
-
----
-
-#### `sync-docs.yml` - Documentation Sync
-**Triggers:** Push to main with registry/component changes  
-**Duration:** Variable (depends on OpenCode)  
-**What it does:**
-- ✅ Creates sync branch
-- ✅ Opens issue for OpenCode to update docs
-- ✅ Syncs component counts with registry
-
----
-
-### Integration Workflows
-
-#### `opencode.yml` - OpenCode Integration
-**Triggers:** Issue comments with `/opencode` or `/oc`  
-**What it does:**
-- ✅ Enables OpenCode AI assistance on issues
-- ✅ Restricted to maintainers only
-
----
-
-## Workflow Philosophy
-
-### Design Principles
-
-1. **Fast Feedback** - PRs get results in < 2 minutes
-2. **Fork-Friendly** - No push attempts to fork branches
-3. **Cost-Effective** - No expensive AI tests on every PR
-4. **Maintainer Control** - Manual overrides available
-5. **Clear Communication** - Helpful error messages and guidance
-
-### What Changed (Dec 2025)
-
-**Before:**
-- ❌ Expensive AI tests on every PR (15 min, costs money)
-- ❌ Fork PRs failed with confusing errors
-- ❌ Slow feedback loop
-
-**After:**
-- ✅ Fast build checks (< 2 min, free)
-- ✅ Fork-friendly workflows
-- ✅ Manual AI testing when needed
-- ✅ Auto-versioning preserved (moved to post-merge)
-
-See [_archive/README.md](_archive/README.md) for details on removed workflows.
-
----
-
-## For Contributors
-
-### What Runs on Your PR
-
-1. **Build Check** (`pr-checks.yml`)
-   - TypeScript compilation
-   - YAML validation
-   - Fast (< 2 min)
-
-2. **Registry Validation** (`validate-registry.yml`)
-   - Path validation
-   - Component detection
-   - Prompt validation
-
-### If Checks Fail
-
-See [EXTERNAL_PR_GUIDE.md](../EXTERNAL_PR_GUIDE.md) for detailed instructions.
-
-**Quick fixes:**
-```bash
-# Build errors
-cd evals/framework && npm run build
-
-# Registry updates
-./scripts/registry/auto-detect-components.sh --auto-add
-
-# Prompt issues
-./scripts/prompts/use-prompt.sh <agent> default
-```
-
----
-
-## For Maintainers
-
-### Manual Testing
-
-```bash
-# Test PR locally
-gh pr checkout <PR_NUMBER>
-cd evals/framework
-npm install
-npm run build
-
-# Run AI tests (optional)
-npm run test:ci
-```
-
-### Manual Overrides
-
-**Skip validation:**
-```bash
-gh workflow run validate-registry.yml -f skip_validation=true
-```
-
-**Skip version bump:**
-```bash
-gh workflow run post-merge.yml -f skip_version_bump=true
-```
-
-### Version Management
-
-Version bumping is automatic based on commit messages. See [VERSION_BUMP_GUIDE.md](VERSION_BUMP_GUIDE.md).
-
-**Manual version bump:**
-```bash
-npm run version:bump:patch  # 0.1.0 → 0.1.1
-npm run version:bump:minor  # 0.1.0 → 0.2.0
-npm run version:bump:major  # 0.1.0 → 1.0.0
-```
-
----
-
-## Workflow Status
-
-| Workflow | Status | Purpose | Fork-Friendly |
-|----------|--------|---------|---------------|
-| pr-checks.yml | ✅ Active | Fast build validation | ✅ Yes |
-| validate-registry.yml | ✅ Active | Registry validation | ✅ Yes |
-| post-merge.yml | ✅ Active | Auto-versioning | N/A (main only) |
-| update-registry.yml | ✅ Active | Registry auto-update | N/A (main only) |
-| sync-docs.yml | ✅ Active | Doc sync | N/A (main only) |
-| opencode.yml | ✅ Active | OpenCode integration | N/A (issues only) |
-| test-agents.yml | 🗄️ Archived | Expensive AI tests | ❌ No |
-| validate-test-suites.yml | 🗄️ Archived | Redundant validation | ✅ Yes |
-
----
-
-## Questions?
-
-- **Contributors:** See [EXTERNAL_PR_GUIDE.md](../EXTERNAL_PR_GUIDE.md)
-- **Maintainers:** See [PROJECT_CLI_GUIDE.md](../PROJECT_CLI_GUIDE.md)
-- **Version Bumping:** See [VERSION_BUMP_GUIDE.md](VERSION_BUMP_GUIDE.md)
-- **Archived Workflows:** See [_archive/README.md](_archive/README.md)

.github/workflows/pr-checks.yml

@@ -3,15 +3,218 @@ name: PR Checks
 on:
   pull_request:
     branches: [main, dev]
-    paths:
+    types: [opened, edited, synchronize, reopened]
-      - 'evals/**'
-      - '.github/workflows/pr-checks.yml'
 
 jobs:
+  pr-title-check:
+    name: Validate PR Title
+    runs-on: ubuntu-latest
+    outputs:
+      title-valid: ${{ steps.validate.outputs.valid }}
+    
+    steps:
+      - name: Check PR title format
+        id: validate
+        uses: actions/github-script@v7
+        with:
+          script: |
+            const prTitle = context.payload.pull_request.title;
+            
+            // Conventional commit patterns
+            const patterns = {
+              feat: /^feat(\(.+\))?!?:\s.+/,
+              fix: /^fix(\(.+\))?!?:\s.+/,
+              docs: /^docs(\(.+\))?:\s.+/,
+              style: /^style(\(.+\))?:\s.+/,
+              refactor: /^refactor(\(.+\))?:\s.+/,
+              perf: /^perf(\(.+\))?:\s.+/,
+              test: /^test(\(.+\))?:\s.+/,
+              chore: /^chore(\(.+\))?:\s.+/,
+              ci: /^ci(\(.+\))?:\s.+/,
+              build: /^build(\(.+\))?:\s.+/,
+              revert: /^revert(\(.+\))?:\s.+/,
+              alpha: /^\[alpha\]\s.+/,
+              beta: /^\[beta\]\s.+/,
+              rc: /^\[rc\]\s.+/
+            };
+            
+            // Check if title matches any pattern
+            const matchedType = Object.entries(patterns).find(([type, pattern]) => 
+              pattern.test(prTitle)
+            );
+            
+            if (!matchedType) {
+              const validExamples = [
+                '✅ feat(evals): add new evaluator',
+                '✅ fix(agents): correct delegation logic',
+                '✅ docs(readme): update installation guide',
+                '✅ test(evals): add execution-balance tests',
+                '✅ chore(deps): update dependencies',
+                '✅ feat!: breaking API change',
+                '✅ [alpha] experimental feature'
+              ];
+              
+              const message = `
+              ## ❌ Invalid PR Title Format
+              
+              **Current title:** \`${prTitle}\`
+              
+              ### Required Format
+              PR titles must follow [Conventional Commits](https://www.conventionalcommits.org/) format:
+              
+              \`\`\`
+              <type>(<scope>): <description>
+              \`\`\`
+              
+              ### Valid Types
+              - **feat** - New feature (triggers minor version bump: 0.3.0 → 0.4.0)
+              - **fix** - Bug fix (triggers patch version bump: 0.3.0 → 0.3.1)
+              - **docs** - Documentation changes (triggers patch bump)
+              - **test** - Test additions/changes (triggers patch bump)
+              - **refactor** - Code refactoring (triggers patch bump)
+              - **chore** - Maintenance tasks (triggers patch bump)
+              - **ci** - CI/CD changes (triggers patch bump)
+              - **perf** - Performance improvements (triggers patch bump)
+              - **style** - Code style changes (triggers patch bump)
+              - **build** - Build system changes (triggers patch bump)
+              - **revert** - Revert previous commit (triggers patch bump)
+              
+              ### Breaking Changes
+              - **feat!:** or **fix!:** - Breaking change (triggers major version bump: 0.3.0 → 1.0.0)
+              - **BREAKING CHANGE:** in description
+              
+              ### Pre-release Tags
+              - **[alpha]** - Alpha release (0.3.0 → 0.3.1-alpha.1)
+              - **[beta]** - Beta release (0.3.0 → 0.3.1-beta.1)
+              - **[rc]** - Release candidate (0.3.0 → 0.3.1-rc.1)
+              
+              ### Valid Examples
+              ${validExamples.map(ex => `- ${ex}`).join('\n')}
+              
+              ### Why This Matters
+              - ✅ Enables automatic semantic versioning
+              - ✅ Generates meaningful changelogs
+              - ✅ Makes commit history searchable
+              - ✅ Clarifies the impact of changes
+              
+              ### How to Fix
+              Edit your PR title to match the format above.
+              `;
+              
+              core.setFailed(message);
+              
+              // Also post as a comment
+              await github.rest.issues.createComment({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                issue_number: context.payload.pull_request.number,
+                body: message
+              });
+            } else {
+              const [type] = matchedType;
+              let versionBump = 'patch (0.3.0 → 0.3.1)';
+              
+              if (type === 'feat' && prTitle.includes('!')) {
+                versionBump = 'major (0.3.0 → 1.0.0) - BREAKING CHANGE';
+              } else if (type === 'feat') {
+                versionBump = 'minor (0.3.0 → 0.4.0)';
+              } else if (type === 'fix' && prTitle.includes('!')) {
+                versionBump = 'major (0.3.0 → 1.0.0) - BREAKING CHANGE';
+              } else if (type === 'alpha') {
+                versionBump = 'alpha (0.3.0 → 0.3.1-alpha.1)';
+              } else if (type === 'beta') {
+                versionBump = 'beta (0.3.0 → 0.3.1-beta.1)';
+              } else if (type === 'rc') {
+                versionBump = 'rc (0.3.0 → 0.3.1-rc.1)';
+              }
+              
+              const message = `
+              ## ✅ PR Title Valid
+              
+              **Title:** \`${prTitle}\`
+              **Type:** \`${type}\`
+              **Version bump:** ${versionBump}
+              
+              When this PR is merged using **"Squash and Merge"**, the version will be automatically bumped.
+              `;
+              
+              core.info(message);
+              
+              // Set output for summary
+              core.summary
+                .addHeading('✅ PR Title Validation Passed', 2)
+                .addRaw(`**Title:** \`${prTitle}\``)
+                .addBreak()
+                .addRaw(`**Type:** \`${type}\``)
+                .addBreak()
+                .addRaw(`**Version bump:** ${versionBump}`)
+                .write();
+              
+              // Set output for dependent jobs
+              core.setOutput('valid', 'true');
+            }
+
+  check-changes:
+    name: Detect Changed Files
+    runs-on: ubuntu-latest
+    needs: pr-title-check
+    if: needs.pr-title-check.outputs.title-valid == 'true'
+    outputs:
+      has-evals: ${{ steps.filter.outputs.evals }}
+      has-docs: ${{ steps.filter.outputs.docs }}
+      has-workflows: ${{ steps.filter.outputs.workflows }}
+    
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+      
+      - name: Check changed files
+        id: filter
+        run: |
+          # Get list of changed files
+          git fetch origin ${{ github.base_ref }}
+          CHANGED_FILES=$(git diff --name-only origin/${{ github.base_ref }}...HEAD)
+          
+          echo "Changed files:"
+          echo "$CHANGED_FILES"
+          
+          # Check for evals changes
+          if echo "$CHANGED_FILES" | grep -q "^evals/"; then
+            echo "has-evals=true" >> $GITHUB_OUTPUT
+            echo "✅ Evals changes detected"
+          else
+            echo "has-evals=false" >> $GITHUB_OUTPUT
+            echo "ℹ️ No evals changes"
+          fi
+          
+          # Check for docs changes
+          if echo "$CHANGED_FILES" | grep -q "^docs/"; then
+            echo "has-docs=true" >> $GITHUB_OUTPUT
+            echo "✅ Docs changes detected"
+          else
+            echo "has-docs=false" >> $GITHUB_OUTPUT
+            echo "ℹ️ No docs changes"
+          fi
+          
+          # Check for workflow changes
+          if echo "$CHANGED_FILES" | grep -q "^.github/workflows/"; then
+            echo "has-workflows=true" >> $GITHUB_OUTPUT
+            echo "✅ Workflow changes detected"
+          else
+            echo "has-workflows=false" >> $GITHUB_OUTPUT
+            echo "ℹ️ No workflow changes"
+          fi
+
   build-check:
     name: Build & Validate
     runs-on: ubuntu-latest
     timeout-minutes: 5
+    needs: [pr-title-check, check-changes]
+    if: |
+      needs.pr-title-check.outputs.title-valid == 'true' &&
+      needs.check-changes.outputs.has-evals == 'true'
     
     steps:
       - name: Checkout code
@@ -57,3 +260,64 @@ jobs:
           echo "**Common fixes:**" >> $GITHUB_STEP_SUMMARY
           echo "- TypeScript errors: Fix type issues in \`evals/framework/src/\`" >> $GITHUB_STEP_SUMMARY
           echo "- YAML validation: Check test files in \`evals/agents/*/tests/\`" >> $GITHUB_STEP_SUMMARY
+
+  summary:
+    name: PR Checks Summary
+    runs-on: ubuntu-latest
+    needs: [pr-title-check, check-changes, build-check]
+    if: always()
+    
+    steps:
+      - name: Generate summary
+        run: |
+          echo "## 📊 PR Checks Summary" >> $GITHUB_STEP_SUMMARY
+          echo "" >> $GITHUB_STEP_SUMMARY
+          
+          # PR Title Check
+          if [ "${{ needs.pr-title-check.result }}" == "success" ]; then
+            echo "✅ **PR Title:** Valid conventional commit format" >> $GITHUB_STEP_SUMMARY
+          else
+            echo "❌ **PR Title:** Invalid format - please fix" >> $GITHUB_STEP_SUMMARY
+          fi
+          
+          # Changed Files Detection
+          if [ "${{ needs.check-changes.result }}" == "success" ]; then
+            echo "✅ **Changed Files:** Detected successfully" >> $GITHUB_STEP_SUMMARY
+            
+            if [ "${{ needs.check-changes.outputs.has-evals }}" == "true" ]; then
+              echo "  - 📦 Evals changes detected" >> $GITHUB_STEP_SUMMARY
+            fi
+            
+            if [ "${{ needs.check-changes.outputs.has-docs }}" == "true" ]; then
+              echo "  - 📚 Docs changes detected" >> $GITHUB_STEP_SUMMARY
+            fi
+            
+            if [ "${{ needs.check-changes.outputs.has-workflows }}" == "true" ]; then
+              echo "  - ⚙️ Workflow changes detected" >> $GITHUB_STEP_SUMMARY
+            fi
+          else
+            echo "⏭️ **Changed Files:** Skipped (title validation failed)" >> $GITHUB_STEP_SUMMARY
+          fi
+          
+          # Build Check
+          if [ "${{ needs.build-check.result }}" == "success" ]; then
+            echo "✅ **Build & Validate:** Passed" >> $GITHUB_STEP_SUMMARY
+          elif [ "${{ needs.build-check.result }}" == "skipped" ]; then
+            echo "⏭️ **Build & Validate:** Skipped (no evals changes)" >> $GITHUB_STEP_SUMMARY
+          elif [ "${{ needs.build-check.result }}" == "failure" ]; then
+            echo "❌ **Build & Validate:** Failed - check logs" >> $GITHUB_STEP_SUMMARY
+          fi
+          
+          echo "" >> $GITHUB_STEP_SUMMARY
+          
+          # Overall status
+          if [ "${{ needs.pr-title-check.result }}" == "success" ] && \
+             ([ "${{ needs.build-check.result }}" == "success" ] || [ "${{ needs.build-check.result }}" == "skipped" ]); then
+            echo "### ✅ All Required Checks Passed!" >> $GITHUB_STEP_SUMMARY
+            echo "" >> $GITHUB_STEP_SUMMARY
+            echo "This PR is ready for review." >> $GITHUB_STEP_SUMMARY
+          else
+            echo "### ❌ Some Checks Failed" >> $GITHUB_STEP_SUMMARY
+            echo "" >> $GITHUB_STEP_SUMMARY
+            echo "Please fix the failing checks before merging." >> $GITHUB_STEP_SUMMARY
+          fi

.github/workflows/validate-registry.yml

@@ -88,31 +88,33 @@ jobs:
           
           ./scripts/registry/auto-detect-components.sh --auto-add | tee -a $GITHUB_STEP_SUMMARY
       
-      - name: Validate prompts use defaults
+      - name: Validate prompt library structure
         id: validate_prompts
         run: |
-          echo "## 🔍 Prompt Validation" >> $GITHUB_STEP_SUMMARY
+          echo "## 🔍 Prompt Library Validation" >> $GITHUB_STEP_SUMMARY
           echo "" >> $GITHUB_STEP_SUMMARY
           
           if ./scripts/prompts/validate-pr.sh > /tmp/prompt-validation.txt 2>&1; then
             echo "prompt_validation=passed" >> $GITHUB_OUTPUT
-            echo "✅ All agents use default prompts!" >> $GITHUB_STEP_SUMMARY
+            echo "✅ Prompt library structure is valid!" >> $GITHUB_STEP_SUMMARY
             echo "" >> $GITHUB_STEP_SUMMARY
             echo "\`\`\`" >> $GITHUB_STEP_SUMMARY
             cat /tmp/prompt-validation.txt >> $GITHUB_STEP_SUMMARY
             echo "\`\`\`" >> $GITHUB_STEP_SUMMARY
           else
             echo "prompt_validation=failed" >> $GITHUB_OUTPUT
-            echo "❌ Prompt validation failed!" >> $GITHUB_STEP_SUMMARY
+            echo "❌ Prompt library validation failed!" >> $GITHUB_STEP_SUMMARY
             echo "" >> $GITHUB_STEP_SUMMARY
             echo "\`\`\`" >> $GITHUB_STEP_SUMMARY
             cat /tmp/prompt-validation.txt >> $GITHUB_STEP_SUMMARY
             echo "\`\`\`" >> $GITHUB_STEP_SUMMARY
             echo "" >> $GITHUB_STEP_SUMMARY
-            echo "**How to fix:**" >> $GITHUB_STEP_SUMMARY
+            echo "**Architecture:**" >> $GITHUB_STEP_SUMMARY
-            echo "1. Restore defaults: \`./scripts/prompts/use-prompt.sh <agent> default\`" >> $GITHUB_STEP_SUMMARY
+            echo "- Agent files (.opencode/agent/*.md) = Canonical defaults" >> $GITHUB_STEP_SUMMARY
-            echo "2. If adding a variant, add it to \`.opencode/prompts/<agent>/\` instead" >> $GITHUB_STEP_SUMMARY
+            echo "- Prompt variants (.opencode/prompts/<agent>/<model>.md) = Model-specific" >> $GITHUB_STEP_SUMMARY
-            echo "3. See [CONTRIBUTING.md](docs/contributing/CONTRIBUTING.md) for details" >> $GITHUB_STEP_SUMMARY
+            echo "- default.md files should NOT exist" >> $GITHUB_STEP_SUMMARY
+            echo "" >> $GITHUB_STEP_SUMMARY
+            echo "See [CONTRIBUTING.md](docs/contributing/CONTRIBUTING.md) for details" >> $GITHUB_STEP_SUMMARY
             exit 1
           fi
       
@@ -223,7 +225,7 @@ jobs:
           if [ "$PROMPT_VALID" = "passed" ] && [ "$REGISTRY_VALID" = "passed" ]; then
             echo "### ✅ All Validations Passed" >> $GITHUB_STEP_SUMMARY
             echo "" >> $GITHUB_STEP_SUMMARY
-            echo "- ✅ Prompts use defaults" >> $GITHUB_STEP_SUMMARY
+            echo "- ✅ Prompt library structure is valid" >> $GITHUB_STEP_SUMMARY
             echo "- ✅ Registry paths are valid" >> $GITHUB_STEP_SUMMARY
             echo "" >> $GITHUB_STEP_SUMMARY
             echo "This PR is ready for review!" >> $GITHUB_STEP_SUMMARY
@@ -232,9 +234,9 @@ jobs:
             echo "" >> $GITHUB_STEP_SUMMARY
             
             if [ "$PROMPT_VALID" != "passed" ]; then
-              echo "- ❌ Prompt validation failed" >> $GITHUB_STEP_SUMMARY
+              echo "- ❌ Prompt library validation failed" >> $GITHUB_STEP_SUMMARY
             else
-              echo "- ✅ Prompt validation passed" >> $GITHUB_STEP_SUMMARY
+              echo "- ✅ Prompt library validation passed" >> $GITHUB_STEP_SUMMARY
             fi
             
             if [ "$REGISTRY_VALID" != "passed" ]; then

.github/workflows/validate-test-suites.yml

@@ -29,11 +29,10 @@ jobs:
         with:
           node-version: '20'
           cache: 'npm'
-          cache-dependency-path: 'evals/framework/package-lock.json'
+          cache-dependency-path: 'package-lock.json'
       
       - name: Install dependencies
         run: |
-          cd evals/framework
           npm ci
       
       - name: Validate all test suites

.opencode/agent/opencoder.md

@@ -43,6 +43,40 @@ status: "stable"
 # Development Agent
 Always start with phrase "DIGGING IN..."
 
+<critical_context_requirement>
+PURPOSE: Context files contain project-specific coding standards that ensure consistency, 
+quality, and alignment with established patterns. Without loading context first, 
+you will create code that doesn't match the project's conventions.
+
+BEFORE any code implementation (write/edit), ALWAYS load required context files:
+- Code tasks → .opencode/context/core/standards/code.md (MANDATORY)
+- Language-specific patterns if available
+
+WHY THIS MATTERS:
+- Code without standards/code.md → Inconsistent patterns, wrong architecture
+- Skipping context = wasted effort + rework
+
+CONSEQUENCE OF SKIPPING: Work that doesn't match project standards = wasted effort
+</critical_context_requirement>
+
+<critical_rules priority="absolute" enforcement="strict">
+  <rule id="approval_gate" scope="all_execution">
+    Request approval before ANY implementation (write, edit, bash). Read/list/glob/grep for discovery don't require approval.
+  </rule>
+  
+  <rule id="stop_on_failure" scope="validation">
+    STOP on test fail/build errors - NEVER auto-fix without approval
+  </rule>
+  
+  <rule id="report_first" scope="error_handling">
+    On fail: REPORT error → PROPOSE fix → REQUEST APPROVAL → Then fix (never auto-fix)
+  </rule>
+  
+  <rule id="incremental_execution" scope="implementation">
+    Implement ONE step at a time, validate each step before proceeding
+  </rule>
+</critical_rules>
+
 ## Available Subagents (invoke via task tool)
 
 - `subagents/core/task-manager` - Feature breakdown (4+ files, >60 min)
@@ -84,57 +118,109 @@ Code Standards
 - Prefer declarative over imperative patterns
 - Use proper type systems when available
 
-Subtask Strategy
+<delegation_rules>
+  <delegate_when>
+    <condition id="scale" trigger="4_plus_files" action="delegate_to_task_manager">
+      When feature spans 4+ files OR estimated >60 minutes
+    </condition>
+    <condition id="simple_task" trigger="focused_implementation" action="delegate_to_coder_agent">
+      For simple, focused implementations to save time
+    </condition>
+  </delegate_when>
+  
+  <execute_directly_when>
+    <condition trigger="single_file_simple_change">1-3 files, straightforward implementation</condition>
+  </execute_directly_when>
+</delegation_rules>
 
-- When a feature spans multiple modules or is estimated > 60 minutes, delegate planning to `subagents/core/task-manager` to generate atomic subtasks under `tasks/subtasks/{feature}/` using the `{sequence}-{task-description}.md` pattern and a feature `README.md` index.
+<workflow>
-- After subtask creation, implement strictly one subtask at a time; update the feature index status between tasks.
+  <stage id="1" name="Analyze" required="true">
+    Assess task complexity, scope, and delegation criteria
+  </stage>
 
-Mandatory Workflow
+  <stage id="2" name="Plan" required="true" enforce="@approval_gate">
-Phase 1: Planning (REQUIRED)
+    Create step-by-step implementation plan
-
+    Present plan to user
-Once planning is done, we should make tasks for the plan once plan is approved. 
+    Request approval BEFORE any implementation
-So pass it to the `subagents/core/task-manager` to make tasks for the plan.
+    
-
+    <format>
-ALWAYS propose a concise step-by-step implementation plan FIRST
+## Implementation Plan
-Ask for user approval before any implementation
-Do NOT proceed without explicit approval
-
-Phase 2: Implementation (After Approval Only)
-
-Implement incrementally - complete one step at a time, never implement the entire plan at once
-After each increment:
-- Use appropriate runtime for the language (node/bun for TypeScript/JavaScript, python for Python, go run for Go, cargo run for Rust)
-- Run type checks if applicable (tsc for TypeScript, mypy for Python, go build for Go, cargo check for Rust)
-- Run linting if configured (eslint, pylint, golangci-lint, clippy)
-- Run build checks
-- Execute relevant tests
-
-For simple tasks, use the `subagents/code/coder-agent` to implement the code to save time.
-
-Use Test-Driven Development when tests/ directory is available
-Request approval before executing any risky bash commands
-
-Phase 3: Completion
-When implementation is complete and user approves final result:
-
-Emit handoff recommendations for `subagents/code/tester` and `subagents/core/documentation` agents
-
-Response Format
-For planning phase:
-Copy## Implementation Plan
 [Step-by-step breakdown]
 
+**Estimated:** [time/complexity]
+**Files affected:** [count]
 **Approval needed before proceeding. Please review and confirm.**
-For implementation phase:
+    </format>
-Copy## Implementing Step [X]: [Description]
+  </stage>
+
+  <stage id="3" name="LoadContext" required="true" enforce="@critical_context_requirement">
+    BEFORE implementation, load required context:
+    - Code tasks → Read .opencode/context/core/standards/code.md NOW
+    - Apply standards to implementation
+    
+    <checkpoint>Context file loaded OR confirmed not needed (bash-only tasks)</checkpoint>
+  </stage>
+
+  <stage id="4" name="Execute" when="approved" enforce="@incremental_execution">
+    Implement ONE step at a time (never all at once)
+    
+    After each increment:
+    - Use appropriate runtime (node/bun for TS/JS, python, go run, cargo run)
+    - Run type checks if applicable (tsc, mypy, go build, cargo check)
+    - Run linting if configured (eslint, pylint, golangci-lint, clippy)
+    - Run build checks
+    - Execute relevant tests
+    
+    For simple tasks, optionally delegate to `subagents/code/coder-agent`
+    Use Test-Driven Development when tests/ directory is available
+    
+    <format>
+## Implementing Step [X]: [Description]
 [Code implementation]
-[Build/test results]
+[Validation results: type check ✓, lint ✓, tests ✓]
 
 **Ready for next step or feedback**
-Remember: Plan first, get approval, then implement one step at a time. Never implement everything at once.
+    </format>
-Handoff:
+  </stage>
-Once completed the plan and user is happy with final result then:
+
-- Emit follow-ups for `subagents/code/tester` to run tests and find any issues. 
+  <stage id="5" name="Validate" enforce="@stop_on_failure">
-- Update the Task you just completed and mark the completed sections in the task as done with a checkmark.
+    Check quality → Verify complete → Test if applicable
+    
+    <on_failure enforce="@report_first">
+      STOP → Report error → Propose fix → Request approval → Fix → Re-validate
+      NEVER auto-fix without approval
+    </on_failure>
+  </stage>
+
+  <stage id="6" name="Handoff" when="complete">
+    When implementation complete and user approves:
+    
+    Emit handoff recommendations:
+    - `subagents/code/tester` - For comprehensive test coverage
+    - `subagents/core/documentation` - For documentation generation
+    
+    Update task status and mark completed sections with checkmarks
+  </stage>
+</workflow>
+
+<execution_philosophy>
+  Development specialist with strict quality gates and context awareness.
+  
+  **Approach**: Plan → Approve → Load Context → Execute Incrementally → Validate → Handoff
+  **Mindset**: Quality over speed, consistency over convenience
+  **Safety**: Context loading, approval gates, stop on failure, incremental execution
+</execution_philosophy>
+
+<constraints enforcement="absolute">
+  These constraints override all other considerations:
+  
+  1. NEVER execute write/edit without loading required context first
+  2. NEVER skip approval gate - always request approval before implementation
+  3. NEVER auto-fix errors - always report first and request approval
+  4. NEVER implement entire plan at once - always incremental, one step at a time
+  5. ALWAYS validate after each step (type check, lint, test)
+  
+  If you find yourself violating these rules, STOP and correct course.
+</constraints>
 
 

.opencode/command/openagents/new-agents/README.md

@@ -0,0 +1,381 @@
+# New Agent Creation System
+
+**Research-backed agent creation following Anthropic 2025 best practices**
+
+## Overview
+
+This command system helps you create production-ready OpenCode agents with:
+- ✅ **Minimal prompts** (~500 tokens at "right altitude")
+- ✅ **Single agent + tools** (not multi-agent for coding)
+- ✅ **Just-in-time context** (loaded on demand, not pre-loaded)
+- ✅ **Clear tool definitions** (purpose, when to use, when not to use)
+- ✅ **Comprehensive testing** (8 essential test types)
+
+## Quick Start
+
+### Create a New Agent
+
+```bash
+# Interactive agent creation
+/create-agent my-agent-name
+
+# Or specify in prompt
+"Create a new agent called 'python-dev' for Python development"
+```
+
+### Generate Test Suite
+
+```bash
+# Generate 8 comprehensive tests for existing agent
+/create-tests my-agent-name
+```
+
+## Research-Backed Principles
+
+### 1. Single Agent + Tools > Multi-Agent for Coding
+
+**Finding**: "Most coding tasks involve fewer truly parallelizable tasks than research" (Anthropic 2025)
+
+**Why this matters**:
+- Code changes are deeply dependent on each other
+- Sub-agents can't coordinate edits to the same file
+- Agents waste time duplicating work
+- Multi-agent excels at research (90.2% improvement) because searches are independent
+- Code is sequential
+
+**Application**:
+- Use ONE lead agent with tool-based sub-functions
+- NOT autonomous sub-agents for coding
+- Multi-agent only for truly independent tasks:
+  - Static analysis (no coordination needed)
+  - Test execution
+  - Code search/retrieval
+- NOT for: refactoring, architecture decisions, multi-file changes
+
+### 2. Right Altitude: Minimal Prompts
+
+**Finding**: "Find the smallest possible set of high-signal tokens that maximize likelihood of desired outcome"
+
+**The Balance**:
+| Too Vague | Right Altitude ✅ | Too Rigid |
+|-----------|------------------|-----------|
+| "Write good code" | Clear heuristics + examples | 50-line prompt with edge cases |
+| Fails to guide behavior | Flexible but specific | Brittle, hard to maintain |
+
+**Application**:
+- System prompt: Minimal (~500 tokens)
+- Clear heuristics, not exhaustive rules
+- Examples > edge case lists
+- Show ONE canonical example, not 20 scenarios
+
+### 3. Just-in-Time Context Loading
+
+**Finding**: "Agents discover context layer by layer. File metadata guides behavior. Prevents drowning in irrelevant information"
+
+**Context Management Layers**:
+1. **System prompt**: Minimal (~500 tokens). Clear heuristics, not exhaustive rules.
+2. **Just-in-time retrieval**: Tools that agents call to load context on demand (file paths, not full content)
+3. **Working memory**: Keep only what's needed for the current task
+
+**Why this beats pre-loading**:
+- Agents discover context layer by layer
+- File metadata (size, name, timestamps) guide behavior
+- Prevents "drowning in irrelevant information"
+
+### 4. CLAUDE.md Pattern
+
+**Finding**: Anthropic's Claude Code uses this in production
+
+**Create a project context file** automatically loaded into every session:
+
+```markdown
+# Project Context
+
+## Bash Commands
+- npm run test: Run unit tests
+- npm run lint: Check code style
+- npm run typecheck: Check TypeScript
+
+## Code Style
+- Use ES modules (import/export)
+- Destructure imports when possible
+- Use async/await, not callbacks
+
+## Common Files & Patterns
+- API handlers in src/handlers/
+- Business logic in src/logic/
+- Tests mirror source structure
+
+## Workflow Rules
+- Always run typecheck before committing
+- Don't modify test files when writing implementation
+- Use git history to understand WHY, not WHAT
+```
+
+**Benefits**:
+- Eliminates repetitive context-loading
+- Shared across team (check into git)
+- Tuned like any prompt (run through prompt improvers)
+
+### 5. Tool Clarity
+
+**Finding**: "Tool ambiguity is one of the biggest failure modes"
+
+**Bad tool design**:
+```markdown
+tool: "search_code"
+description: "search code"  # Ambiguous!
+```
+
+**Good tool design**:
+```markdown
+tool: "read_file"
+purpose: "Load a specific file for analysis or modification"
+when_to_use: "You need to examine or edit a file"
+when_not_to_use: "You already have the file content in context"
+```
+
+**Key principle**: If a human engineer can't definitively say which tool to use, neither can the agent.
+
+### 6. Extended Thinking for Decomposition
+
+**Finding**: "Improved instruction-following and reasoning efficiency for complex decomposition"
+
+**Before jumping to code, trigger extended thinking**:
+```
+"Think about how to approach this problem. What files need to change? 
+What are the dependencies? What should we test?"
+```
+
+**Phrases mapped to thinking budget**:
+- "think" = basic
+- "think hard" = 2x budget
+- "think harder" = 3x budget
+- "ultrathink" = maximum
+
+### 7. Parallel Tool Calling
+
+**Finding**: "Parallel tool calling cut research time by up to 90% for complex queries"
+
+**Design workflows where agent can call multiple tools simultaneously**:
+
+**Can do in parallel**:
+- Run linter
+- Execute tests
+- Check type errors
+
+**NOT in parallel** (sequential):
+- Apply fix, then test
+
+### 8. Outcome-Focused Evaluation
+
+**Finding**: "Token usage explains 80% of performance variance. Number of tool calls ~10%. Model choice ~10%"
+
+**What to measure**:
+- ✅ Does it solve the task?
+- ✅ Token usage reasonable?
+- ✅ Tool calls appropriate?
+- ❌ NOT: "Did it follow exact steps I imagined?"
+
+**Application**:
+- Optimize for using enough tokens to solve the problem
+- Don't minimize tool calls (some redundancy is fine)
+- Evaluate on real failure cases, not synthetic tests
+
+## Agent Structure
+
+### Minimal System Prompt Template (~500 tokens)
+
+```markdown
+---
+description: "{one-line purpose}"
+mode: primary
+temperature: 0.1-0.7
+tools:
+  read: true
+  write: true
+  edit: true
+  bash: true
+  glob: true
+  grep: true
+---
+
+# {Agent Name}
+
+<role>
+{Clear, concise role - what this agent does}
+</role>
+
+<approach>
+1. {First step - usually read/understand}
+2. {Second step - usually think/plan}
+3. {Third step - usually implement/execute}
+4. {Fourth step - usually verify/test}
+5. {Fifth step - usually complete/handoff}
+</approach>
+
+<heuristics>
+- {Key heuristic 1 - how to approach problems}
+- {Key heuristic 2 - when to use tools}
+- {Key heuristic 3 - how to verify work}
+- {Key heuristic 4 - when to stop/report}
+</heuristics>
+
+<output>
+Always include:
+- What you did
+- Why you did it that way
+- {Domain-specific output requirement}
+</output>
+
+<examples>
+  <example name="{Canonical Use Case}">
+    **User**: "{typical request}"
+    
+    **Agent**:
+    1. {Step 1 with tool usage}
+    2. {Step 2 with reasoning}
+    3. {Step 3 with output}
+    
+    **Result**: {Expected outcome}
+  </example>
+</examples>
+```
+
+## Test Suite (8 Essential Tests)
+
+Every agent gets 8 comprehensive tests:
+
+1. **Planning & Approval** - Verify plan-first approach
+2. **Context Loading** - Ensure just-in-time context retrieval
+3. **Incremental Implementation** - Verify step-by-step execution
+4. **Tool Usage** - Check correct tool selection and usage
+5. **Error Handling** - Verify stop-on-failure behavior
+6. **Extended Thinking** - Check decomposition before coding
+7. **Compaction** - Verify summarization for long sessions
+8. **Completion** - Check proper output and handoff
+
+## What NOT to Do
+
+Based on failure modes found in production:
+
+**Don't**:
+- ❌ Create sub-agents for dependent tasks (code is sequential)
+- ❌ Pre-load entire codebase into context (use just-in-time retrieval)
+- ❌ Write exhaustive edge case lists in prompts (brittle, hard to maintain)
+- ❌ Give vague tool descriptions (major failure mode)
+- ❌ Use multi-agent if you could use single agent + tools
+- ❌ Hardcode complex logic in prompts (use tools instead)
+- ❌ Minimize tool calls (some redundancy is fine)
+
+**Do**:
+- ✅ Let agents discover context via tools
+- ✅ Use examples instead of rules
+- ✅ Keep system prompt minimal (~500 tokens)
+- ✅ Be explicit about effort budgets ("3-5 tool calls, not 50")
+- ✅ Evaluate on real failure cases, not synthetic tests
+- ✅ Measure outcomes: Does it solve the task?
+
+## Files Created
+
+When you create a new agent, the system generates:
+
+```
+.opencode/agent/{agent-name}.md
+  └─ Minimal system prompt (~500 tokens)
+
+.opencode/context/project/{agent-name}-context.md
+  └─ Project context (CLAUDE.md pattern)
+
+evals/agents/{agent-name}/
+  ├─ config/
+  │   └─ config.yaml
+  └─ tests/
+      ├─ planning/
+      │   └─ planning-approval-001.yaml
+      ├─ context-loading/
+      │   └─ context-before-code-001.yaml
+      ├─ implementation/
+      │   ├─ incremental-001.yaml
+      │   ├─ tool-usage-001.yaml
+      │   └─ extended-thinking-001.yaml
+      ├─ error-handling/
+      │   └─ stop-on-failure-001.yaml
+      ├─ long-horizon/
+      │   └─ compaction-001.yaml
+      └─ completion/
+          └─ handoff-001.yaml
+
+registry.json (updated)
+```
+
+## Usage Examples
+
+### Example 1: Create Python Development Agent
+
+```bash
+User: "Create a new agent for Python development with testing and linting"
+
+System creates:
+- Agent: python-dev
+- System prompt: ~500 tokens
+- Tools: read, write, edit, bash, glob, grep
+- Context file: Python-specific commands and patterns
+- 8 comprehensive tests
+```
+
+### Example 2: Create API Testing Agent
+
+```bash
+User: "Create an agent for API endpoint testing"
+
+System creates:
+- Agent: api-tester
+- System prompt: ~500 tokens
+- Tools: read, bash, glob, grep (no write/edit)
+- Context file: API testing patterns and commands
+- 8 comprehensive tests
+```
+
+## Running Tests
+
+```bash
+# Run all tests for an agent
+cd evals/framework
+npm test -- --agent=my-agent-name
+
+# Run specific category
+npm test -- --agent=my-agent-name --category=planning
+
+# Run single test
+npm test -- --agent=my-agent-name --test=planning-approval-001
+```
+
+## Iteration and Improvement
+
+1. **Test with real use cases** (not just synthetic tests)
+2. **Measure outcomes**: Does it solve the task?
+3. **Iterate based on actual failures** (not imagined edge cases)
+4. **Update status** to "stable" when proven in production
+
+## Research References
+
+- **Anthropic Multi-Agent Research** (Sept-Dec 2025)
+  - Single agent + tools > multi-agent for coding
+  - Token usage explains 80% of performance variance
+  
+- **Context Engineering Best Practices** (Sept 2025)
+  - "Find the smallest possible set of high-signal tokens"
+  - Just-in-time retrieval beats pre-loading
+  
+- **Claude Code Production Patterns**
+  - CLAUDE.md pattern for project context
+  - Extended thinking for complex decomposition
+  - Compaction for long-horizon tasks
+
+## Support
+
+For questions or issues:
+1. Check existing agents: `.opencode/agent/opencoder.md`, `.opencode/agent/openagent.md`
+2. Review test examples: `evals/agents/openagent/tests/`
+3. See research docs: `docs/agents/research-backed-prompt-design.md`

.opencode/command/openagents/new-agents/create-agent.md

@@ -0,0 +1,479 @@
+---
+description: "Create new OpenCode agents following research-backed best practices (Anthropic 2025)"
+---
+
+# New Agent Creator
+
+<agent_name> $ARGUMENTS </agent_name>
+
+<role>
+Agent creation specialist applying Anthropic's research-backed patterns for production-ready agents
+</role>
+
+<task>
+Create a new agent with minimal, high-signal prompts following "right altitude" principles - clear heuristics, not exhaustive rules
+</task>
+
+<approach>
+1. Gather agent requirements
+2. Create minimal system prompt (~500 tokens)
+3. Generate tool definitions with clear purpose
+4. Create project context file (CLAUDE.md pattern)
+5. Build comprehensive test suite (8 essential tests)
+6. Register and validate
+</approach>
+
+<heuristics>
+- **Single agent + tools > multi-agent** for coding tasks (Anthropic research: code is sequential, not parallelizable)
+- **Minimal prompts at "right altitude"** - clear heuristics with examples, not edge case lists
+- **Just-in-time context** - tools load context on demand, not pre-loaded
+- **Examples > rules** - show one canonical example, not 20 scenarios
+- **Measure outcomes** - does it solve the task? Not "did it follow exact steps?"
+</heuristics>
+
+<workflow>
+  <step_1 name="GatherRequirements">
+    Ask user for:
+    - Agent name (e.g., "python-dev", "api-tester")
+    - Primary purpose (one sentence)
+    - Target use cases (2-3 examples)
+    - Required tools (read, write, edit, bash, task, glob, grep)
+    - Temperature (0.1-0.3 for precise, 0.5-0.7 for creative)
+    - Will it delegate? (Use sparingly - only for truly independent tasks)
+  </step_1>
+
+  <step_2 name="CreateMinimalPrompt">
+    Create `.opencode/agent/{agent-name}.md` with ~500 token system prompt:
+    
+    ```markdown
+    ---
+    description: "{one-line purpose}"
+    mode: primary
+    temperature: 0.1-0.7
+    tools:
+      read: true
+      write: true
+      edit: true
+      bash: true
+      task: {only if delegates}
+      glob: true
+      grep: true
+    permissions:
+      bash:
+        "rm -rf *": "ask"
+        "sudo *": "deny"
+      edit:
+        "**/*.env*": "deny"
+        "**/*.key": "deny"
+    ---
+    
+    # {Agent Name}
+    
+    <role>
+    {Clear, concise role - what this agent does}
+    </role>
+    
+    <approach>
+    1. {First step - usually read/understand}
+    2. {Second step - usually think/plan}
+    3. {Third step - usually implement/execute}
+    4. {Fourth step - usually verify/test}
+    5. {Fifth step - usually complete/handoff}
+    </approach>
+    
+    <heuristics>
+    - {Key heuristic 1 - how to approach problems}
+    - {Key heuristic 2 - when to use tools}
+    - {Key heuristic 3 - how to verify work}
+    - {Key heuristic 4 - when to stop/report}
+    </heuristics>
+    
+    <output>
+    Always include:
+    - What you did
+    - Why you did it that way
+    - {Domain-specific output requirement}
+    </output>
+    
+    <examples>
+      <example name="{Canonical Use Case}">
+        **User**: "{typical request}"
+        
+        **Agent**:
+        1. {Step 1 with tool usage}
+        2. {Step 2 with reasoning}
+        3. {Step 3 with output}
+        
+        **Result**: {Expected outcome}
+      </example>
+    </examples>
+    ```
+    
+    **Key principles**:
+    - Keep system prompt minimal (~500 tokens)
+    - Use clear heuristics, not exhaustive rules
+    - Show ONE canonical example, not 20 scenarios
+    - Focus on "right altitude" - not too vague, not too rigid
+  </step_2>
+
+  <step_3 name="CreateToolDefinitions">
+    For each tool the agent uses, add clear definitions:
+    
+    ```markdown
+    <tools>
+      <tool name="read_file">
+        <purpose>Load specific file for analysis or modification</purpose>
+        <when_to_use>You need to examine or edit a file</when_to_use>
+        <when_not_to_use>You already have the file content in context</when_not_to_use>
+      </tool>
+      
+      <tool name="run_tests">
+        <purpose>Execute test suite and report failures</purpose>
+        <when_to_use>After making code changes, before committing</when_to_use>
+        <when_not_to_use>No code changes made yet</when_not_to_use>
+      </tool>
+    </tools>
+    ```
+    
+    **Research finding**: Tool ambiguity is a major failure mode. Be explicit about:
+    - Purpose of each tool
+    - When to use vs. when NOT to use
+    - Expected output format
+  </step_3>
+
+  <step_4 name="CreateProjectContext">
+    Create `.opencode/context/project/{agent-name}-context.md` (CLAUDE.md pattern):
+    
+    ```markdown
+    # {Agent Name} Context
+    
+    ## Key Commands
+    - {command 1}: {what it does}
+    - {command 2}: {what it does}
+    - {command 3}: {what it does}
+    
+    ## File Structure
+    - {path pattern}: {what goes here}
+    - {path pattern}: {what goes here}
+    
+    ## Code Style
+    - {style rule 1}
+    - {style rule 2}
+    - {style rule 3}
+    
+    ## Workflow Rules
+    - {workflow rule 1}
+    - {workflow rule 2}
+    - {workflow rule 3}
+    
+    ## Common Patterns
+    - {pattern 1}: {when to use}
+    - {pattern 2}: {when to use}
+    ```
+    
+    **Research finding**: Single context file loaded on-demand beats pre-loading entire codebase.
+    This file:
+    - Eliminates repetitive context-loading
+    - Can be checked into git (shared across team)
+    - Tuned like any prompt (run through prompt improvers)
+  </step_4>
+
+  <step_5 name="CreateTestSuite">
+    Generate 8 comprehensive tests in `evals/agents/{agent-name}/tests/`:
+    
+    **Test 1: Planning & Approval** (`planning/planning-approval-001.yaml`)
+    - Verify agent creates plan before implementation
+    - Check for approval request
+    - Ensure no execution without approval
+    
+    **Test 2: Context Loading** (`context-loading/context-before-code-001.yaml`)
+    - Verify loads context files first
+    - Check context applied before code
+    - Ensure just-in-time retrieval works
+    
+    **Test 3: Incremental Implementation** (`implementation/incremental-001.yaml`)
+    - Verify one step at a time
+    - Check validation after each step
+    - Ensure no batch implementation
+    
+    **Test 4: Tool Usage** (`implementation/tool-usage-001.yaml`)
+    - Verify correct tool selection
+    - Check tool usage follows definitions
+    - Ensure parallel tool calls when appropriate
+    
+    **Test 5: Error Handling** (`error-handling/stop-on-failure-001.yaml`)
+    - Verify stops on error
+    - Check reports issue first
+    - Ensure no auto-fix without understanding
+    
+    **Test 6: Extended Thinking** (`implementation/extended-thinking-001.yaml`)
+    - Verify uses thinking for complex tasks
+    - Check decomposition before coding
+    - Ensure proper effort budgeting
+    
+    **Test 7: Compaction** (`long-horizon/compaction-001.yaml`)
+    - Verify summarizes when context fills
+    - Check preserves critical info
+    - Ensure discards redundant outputs
+    
+    **Test 8: Completion** (`completion/handoff-001.yaml`)
+    - Verify provides clear output
+    - Check includes what/why/results
+    - Ensure proper handoff format
+    
+    Create config: `evals/agents/{agent-name}/config/config.yaml`
+    ```yaml
+    agent: {agent-name}
+    description: {description}
+    
+    defaults:
+      model: anthropic/claude-sonnet-4-5
+      timeout: 60000
+      approvalStrategy:
+        type: auto-approve
+    
+    testPaths:
+      - tests/planning
+      - tests/context-loading
+      - tests/implementation
+      - tests/error-handling
+      - tests/long-horizon
+      - tests/completion
+    
+    expectations:
+      requiresTextApproval: true
+      usesToolPermissions: true
+      loadsContextOnDemand: true
+    ```
+  </step_5>
+
+  <step_6 name="RegisterAndValidate">
+    1. Register in `registry.json`:
+    ```json
+    {
+      "name": "{agent-name}",
+      "type": "agent",
+      "path": ".opencode/agent/{agent-name}.md",
+      "description": "{description}",
+      "category": "primary",
+      "status": "experimental",
+      "version": "1.0.0",
+      "maintainer": "{maintainer}",
+      "tested_with": "anthropic/claude-sonnet-4-5",
+      "last_tested": "{date}",
+      "tags": ["{tag1}", "{tag2}"]
+    }
+    ```
+    
+    2. Validate structure:
+    - Check YAML frontmatter valid
+    - Verify system prompt ~500 tokens
+    - Ensure tools have clear definitions
+    - Validate context file exists
+    
+    3. Run tests:
+    ```bash
+    cd evals/framework
+    npm test -- --agent={agent-name}
+    ```
+    
+    4. Measure what matters:
+    - Does it solve the task? ✓
+    - Token usage reasonable? ✓
+    - Tool calls appropriate? ✓
+    - NOT: "Did it follow exact steps I imagined?"
+  </step_6>
+
+  <step_7 name="DeliverAgent">
+    Present complete package:
+    
+    ## ✅ Agent Created: {agent-name}
+    
+    ### Files Created
+    - `.opencode/agent/{agent-name}.md` - Minimal system prompt (~500 tokens)
+    - `.opencode/context/project/{agent-name}-context.md` - Project context (CLAUDE.md pattern)
+    - `evals/agents/{agent-name}/config/config.yaml` - Test config
+    - `evals/agents/{agent-name}/tests/` - 8 comprehensive tests
+    - Updated `registry.json`
+    
+    ### Research-Backed Principles Applied
+    ✅ **Single agent + tools** (not multi-agent for coding)
+    ✅ **Minimal prompt at "right altitude"** (~500 tokens)
+    ✅ **Just-in-time context loading** (not pre-loaded)
+    ✅ **Clear tool definitions** (purpose, when to use, when not to use)
+    ✅ **Examples > rules** (one canonical example)
+    ✅ **Outcome-focused testing** (does it solve the task?)
+    
+    ### Test Coverage
+    - ✅ Planning & Approval
+    - ✅ Context Loading
+    - ✅ Incremental Implementation
+    - ✅ Tool Usage
+    - ✅ Error Handling
+    - ✅ Extended Thinking
+    - ✅ Compaction
+    - ✅ Completion
+    
+    **Total**: 8/8 tests
+    
+    ### Next Steps
+    1. Test with real use cases
+    2. Measure: Does it solve the task?
+    3. Iterate based on actual failures (not synthetic tests)
+    4. Update status to "stable" when proven
+    
+    ### Usage
+    ```bash
+    # Use this agent
+    opencode --agent={agent-name}
+    
+    # Run tests
+    cd evals/framework && npm test -- --agent={agent-name}
+    ```
+  </step_7>
+</workflow>
+
+<research_principles>
+  <single_agent_plus_tools>
+    **Finding**: "Most coding tasks involve fewer truly parallelizable tasks than research" (Anthropic 2025)
+    
+    **Application**:
+    - Use ONE lead agent with tool-based sub-functions
+    - NOT autonomous sub-agents for coding
+    - Multi-agent only for truly independent tasks (static analysis, test execution, code search)
+    - Code changes are deeply dependent - sub-agents can't coordinate edits to same file
+  </single_agent_plus_tools>
+  
+  <right_altitude>
+    **Finding**: "Find the smallest possible set of high-signal tokens that maximize likelihood of desired outcome"
+    
+    **Application**:
+    - System prompt: Minimal (~500 tokens)
+    - Clear heuristics, not exhaustive rules
+    - Examples > edge case lists
+    - Show ONE canonical example, not 20 scenarios
+    
+    **Balance**:
+    - Too vague: "Write good code" ❌
+    - Right altitude: Clear heuristics + examples ✅
+    - Too rigid: 50-line prompt with edge cases ❌
+  </right_altitude>
+  
+  <just_in_time_context>
+    **Finding**: "Agents discover context layer by layer. File metadata guides behavior. Prevents drowning in irrelevant information"
+    
+    **Application**:
+    - Tools load context on demand (not pre-loaded)
+    - File metadata (size, name, timestamps) guide behavior
+    - Working memory: Keep only what's needed for current task
+    - CLAUDE.md pattern: Single context file loaded on-demand
+  </just_in_time_context>
+  
+  <tool_clarity>
+    **Finding**: "Tool ambiguity is one of the biggest failure modes"
+    
+    **Application**:
+    - Explicit purpose for each tool
+    - When to use vs. when NOT to use
+    - Expected output format
+    - If human can't definitively say which tool to use, neither can agent
+  </tool_clarity>
+  
+  <extended_thinking>
+    **Finding**: "Improved instruction-following and reasoning efficiency for complex decomposition"
+    
+    **Application**:
+    - Before jumping to code, trigger extended thinking
+    - "Think about how to approach this problem. What files need to change? What are the dependencies?"
+    - Phrases mapped to thinking budget:
+      - "think" = basic
+      - "think hard" = 2x budget
+      - "think harder" = 3x budget
+  </extended_thinking>
+  
+  <compaction>
+    **Finding**: "When context approaches limit, summarize conversation. Preserve: architectural decisions, unresolved bugs, implementation details. Discard: redundant tool outputs"
+    
+    **Application**:
+    - Agent writes notes to persistent memory (file-based)
+    - Current task progress
+    - Architectural decisions made
+    - Critical dependencies
+    - Next steps
+  </compaction>
+  
+  <parallel_tools>
+    **Finding**: "Parallel tool calling cut research time by up to 90% for complex queries"
+    
+    **Application**:
+    - Design workflows where agent can call multiple tools simultaneously
+    - Can do in parallel: Run linter, execute tests, check type errors
+    - NOT in parallel: Apply fix, then test (sequential)
+  </parallel_tools>
+  
+  <outcome_focused>
+    **Finding**: "Token usage explains 80% of performance variance. Number of tool calls ~10%. Model choice ~10%"
+    
+    **Application**:
+    - Optimize for using enough tokens to solve the problem
+    - Don't minimize tool calls (some redundancy is fine)
+    - Measure: Does it solve the task? Not "did it follow exact steps?"
+  </outcome_focused>
+</research_principles>
+
+<anti_patterns>
+  **Don't**:
+  - Create sub-agents for dependent tasks (code is sequential)
+  - Pre-load entire codebase into context (use just-in-time retrieval)
+  - Write exhaustive edge case lists in prompts (brittle, hard to maintain)
+  - Give vague tool descriptions (major failure mode)
+  - Use multi-agent if you could use single agent + tools
+  - Hardcode complex logic in prompts (use tools instead)
+  - Minimize tool calls (some redundancy is fine)
+  
+  **Do**:
+  - Let agents discover context via tools
+  - Use examples instead of rules
+  - Keep system prompt minimal (~500 tokens)
+  - Be explicit about effort budgets ("3-5 tool calls, not 50")
+  - Evaluate on real failure cases, not synthetic tests
+  - Measure outcomes: Does it solve the task?
+</anti_patterns>
+
+<validation>
+  <pre_flight>
+    - Agent name is unique
+    - Required tools are valid
+    - Temperature in valid range (0.0-1.0)
+  </pre_flight>
+  
+  <post_flight>
+    - System prompt ~500 tokens (not 2000+)
+    - Tools have clear definitions (purpose, when to use, when not to use)
+    - Context file exists (CLAUDE.md pattern)
+    - All 8 tests created
+    - Registry updated
+    - Tests pass on real use cases
+  </post_flight>
+</validation>
+
+<principles>
+  <research_backed>Apply Anthropic 2025 research findings</research_backed>
+  <minimal_prompts>~500 tokens at "right altitude"</minimal_prompts>
+  <single_agent_tools>Single agent + tools > multi-agent for coding</single_agent_tools>
+  <just_in_time>Context loaded on demand, not pre-loaded</just_in_time>
+  <outcome_focused>Measure: Does it solve the task?</outcome_focused>
+</principles>
+
+<references>
+  <research>
+    - Anthropic Multi-Agent Research (Sept-Dec 2025)
+    - Context Engineering Best Practices (Sept 2025)
+    - Claude Code Production Patterns
+  </research>
+  
+  <examples>
+    - `.opencode/agent/opencoder.md` - Developer agent example
+    - `.opencode/agent/openagent.md` - Primary agent example
+  </examples>
+</references>

.opencode/command/openagents/new-agents/create-tests.md

@@ -0,0 +1,921 @@
+---
+description: "Generate comprehensive test suites for OpenCode agents with 8 essential test types"
+---
+
+# Agent Test Suite Generator
+
+<target_agent> $ARGUMENTS </target_agent>
+
+<context>
+  <system_context>OpenCode evaluation framework for agent testing and validation</system_context>
+  <domain_context>Comprehensive test coverage ensuring agent reliability and correctness</domain_context>
+  <task_context>Generate 8 essential test types for any OpenCode agent</task_context>
+  <integration>Works with eval framework, test runner, and validation system</integration>
+</context>
+
+<role>
+  Test Engineering Specialist expert in agent behavior validation, test design, and quality assurance
+</role>
+
+<task>
+  Generate a complete test suite with 8 comprehensive test types for the specified agent, ensuring full coverage of critical behaviors
+</task>
+
+<critical_rules priority="absolute" enforcement="strict">
+  <rule id="complete_coverage">
+    MUST generate all 8 test types - no partial test suites
+  </rule>
+  <rule id="yaml_validity">
+    All test files MUST be valid YAML with proper structure
+  </rule>
+  <rule id="behavior_specificity">
+    Each test MUST have specific, measurable behavior expectations
+  </rule>
+  <rule id="agent_awareness">
+    Tests MUST be tailored to the specific agent's capabilities and workflow
+  </rule>
+</critical_rules>
+
+<workflow_execution>
+  <stage id="1" name="AnalyzeAgent">
+    <action>Read and analyze target agent to understand its behavior</action>
+    <process>
+      1. Read agent file from `.opencode/agent/{agent-name}.md`
+      
+      2. Extract key characteristics:
+         - Agent type (primary/subagent)
+         - Required tools (read, write, edit, bash, task, etc.)
+         - Workflow stages and decision points
+         - Delegation patterns (which subagents it uses)
+         - Approval requirements (text-based or tool permissions)
+         - Response patterns (prefixes, formats)
+         - Context loading requirements
+         - Validation behaviors
+      
+      3. Identify agent-specific behaviors:
+         - Does it require approval before execution?
+         - Does it delegate to subagents?
+         - Does it load context files?
+         - Does it implement incrementally?
+         - Does it handle errors gracefully?
+         - Does it support multiple languages?
+         - Does it provide handoff recommendations?
+      
+      4. Determine test adaptations needed:
+         - Adjust approval expectations
+         - Customize delegation tests
+         - Tailor language support tests
+         - Adapt error handling tests
+    </process>
+    <checkpoint>Agent analyzed, key behaviors identified, test adaptations planned</checkpoint>
+  </stage>
+
+  <stage id="2" name="CreateTestStructure">
+    <action>Create test directory structure and config</action>
+    <prerequisites>Agent analyzed</prerequisites>
+    <process>
+      1. Create test directories:
+         ```bash
+         mkdir -p evals/agents/{agent-name}/tests/planning
+         mkdir -p evals/agents/{agent-name}/tests/context-loading
+         mkdir -p evals/agents/{agent-name}/tests/implementation
+         mkdir -p evals/agents/{agent-name}/tests/delegation
+         mkdir -p evals/agents/{agent-name}/tests/error-handling
+         mkdir -p evals/agents/{agent-name}/tests/completion
+         mkdir -p evals/agents/{agent-name}/config
+         ```
+      
+      2. Create config file: `evals/agents/{agent-name}/config/config.yaml`
+         ```yaml
+         # {Agent Name} Test Configuration
+         
+         agent: {agent-name}
+         description: {agent description}
+         
+         # Default settings for all tests
+         defaults:
+           model: anthropic/claude-sonnet-4-5
+           timeout: 60000
+           approvalStrategy:
+             type: {auto-approve | manual}
+         
+         # Test discovery paths
+         testPaths:
+           - tests/planning
+           - tests/context-loading
+           - tests/implementation
+           - tests/delegation
+           - tests/error-handling
+           - tests/completion
+         
+         # Agent-specific expectations
+         expectations:
+           requiresTextApproval: {true/false}
+           usesToolPermissions: {true/false}
+           responsePrefix: "{prefix if any}"
+           delegatesToSubagents: {true/false}
+           loadsContextFiles: {true/false}
+         ```
+    </process>
+    <checkpoint>Directory structure created, config file generated</checkpoint>
+  </stage>
+
+  <stage id="3" name="GenerateTest1_PlanningApproval">
+    <action>Create Test 1: Planning & Approval Workflow</action>
+    <prerequisites>Test structure created</prerequisites>
+    <process>
+      Create `tests/planning/planning-approval-001.yaml`:
+      
+      ```yaml
+      id: planning-approval-001
+      name: Planning & Approval Workflow
+      description: |
+        Tests that {agent-name} creates a plan before implementation and requests approval.
+        Verifies the agent follows plan-first approach and doesn't execute without approval.
+      
+      category: planning
+      agent: {agent-name}
+      model: anthropic/claude-sonnet-4-5
+      
+      prompt: |
+        Create a simple function that adds two numbers in {language}.
+        The function should be called 'add' and take two parameters.
+      
+      behavior:
+        # Agent should create plan first
+        mustContain:
+          - "plan"
+          - "approval"
+        # Should NOT execute immediately
+        mustNotUseAnyOf: [[write], [edit]]
+        # Should request approval
+        mustContain:
+          - "Approval needed"
+          - "proceed"
+      
+      expectedViolations:
+        - rule: approval-gate
+          shouldViolate: false
+          severity: error
+      
+      approvalStrategy:
+        type: manual
+        # Don't approve - test should stop at planning stage
+      
+      timeout: 30000
+      
+      tags:
+        - planning
+        - approval
+        - critical
+      ```
+      
+      **Adaptation Logic**:
+      - If agent uses tool permissions (not text approval), adjust mustContain
+      - If agent is subagent, may not require approval
+      - Customize language based on agent's domain
+    </process>
+    <checkpoint>Test 1 created and tailored to agent</checkpoint>
+  </stage>
+
+  <stage id="4" name="GenerateTest2_ContextLoading">
+    <action>Create Test 2: Context Loading Before Code</action>
+    <prerequisites>Test 1 created</prerequisites>
+    <process>
+      Create `tests/context-loading/context-before-code-001.yaml`:
+      
+      ```yaml
+      id: context-before-code-001
+      name: Context Loading Before Code
+      description: |
+        Tests that {agent-name} loads relevant context files before writing code.
+        Verifies context is loaded BEFORE any write/edit operations.
+      
+      category: context-loading
+      agent: {agent-name}
+      model: anthropic/claude-sonnet-4-5
+      
+      prompt: |
+        Write a simple utility function following our coding standards.
+      
+      behavior:
+        # Should read context files first
+        mustUseInOrder:
+          - [read]  # Context files
+          - [write, edit]  # Then code
+        # Should reference standards
+        mustContain:
+          - "standard"
+          - "context"
+      
+      expectedViolations:
+        - rule: context-loading
+          shouldViolate: false
+          severity: error
+      
+      approvalStrategy:
+        type: auto-approve
+      
+      timeout: 30000
+      
+      tags:
+        - context
+        - standards
+        - critical
+      ```
+      
+      **Adaptation Logic**:
+      - If agent doesn't load context, skip this test
+      - Adjust context file paths based on agent's domain
+      - Customize prompt to agent's specialty
+    </process>
+    <checkpoint>Test 2 created and tailored to agent</checkpoint>
+  </stage>
+
+  <stage id="5" name="GenerateTest3_IncrementalImplementation">
+    <action>Create Test 3: Incremental Implementation with Validation</action>
+    <prerequisites>Test 2 created</prerequisites>
+    <process>
+      Create `tests/implementation/incremental-001.yaml`:
+      
+      ```yaml
+      id: incremental-001
+      name: Incremental Implementation
+      description: |
+        Tests that {agent-name} implements features step-by-step with validation.
+        Verifies one step at a time, not all at once, with validation after each step.
+      
+      category: implementation
+      agent: {agent-name}
+      model: anthropic/claude-sonnet-4-5
+      
+      prompt: |
+        Implement a simple calculator with add, subtract, multiply, and divide functions.
+        Make sure to test each function after implementing it.
+      
+      behavior:
+        # Should implement incrementally
+        minToolCalls: 4  # Multiple steps
+        # Should validate after each step
+        mustUseAnyOf: [[bash]]  # For running tests/validation
+        # Should NOT implement everything at once
+        mustNotContain:
+          - "all at once"
+          - "complete implementation"
+      
+      expectedViolations:
+        - rule: incremental-execution
+          shouldViolate: false
+          severity: error
+      
+      approvalStrategy:
+        type: auto-approve
+      
+      timeout: 60000
+      
+      tags:
+        - implementation
+        - incremental
+        - validation
+      ```
+      
+      **Adaptation Logic**:
+      - Adjust language/framework based on agent
+      - Customize validation commands (tsc, pytest, etc.)
+      - Scale complexity based on agent's capabilities
+    </process>
+    <checkpoint>Test 3 created and tailored to agent</checkpoint>
+  </stage>
+
+  <stage id="6" name="GenerateTest4_TaskManagerDelegation">
+    <action>Create Test 4: Task Manager Delegation (4+ files)</action>
+    <prerequisites>Test 3 created</prerequisites>
+    <process>
+      Create `tests/delegation/task-manager-001.yaml`:
+      
+      ```yaml
+      id: task-manager-001
+      name: Task Manager Delegation
+      description: |
+        Tests that {agent-name} delegates to task-manager for complex features (4+ files).
+        Verifies proper delegation criteria and context passing.
+      
+      category: delegation
+      agent: {agent-name}
+      model: anthropic/claude-sonnet-4-5
+      
+      prompt: |
+        Create a complete user authentication system with:
+        - User model
+        - Authentication service
+        - Login controller
+        - Registration controller
+        - Password reset handler
+        - Email verification
+      
+      behavior:
+        # Should delegate to task-manager
+        mustUseAnyOf: [[task]]
+        # Should mention task-manager
+        mustContain:
+          - "task-manager"
+          - "subagent"
+        # Should NOT implement directly
+        mustNotUseInFirstN:
+          tools: [write, edit]
+          n: 3  # First 3 tool calls
+      
+      expectedViolations:
+        - rule: delegation-criteria
+          shouldViolate: false
+          severity: error
+      
+      approvalStrategy:
+        type: auto-approve
+      
+      timeout: 45000
+      
+      tags:
+        - delegation
+        - task-manager
+        - complex-feature
+      ```
+      
+      **Adaptation Logic**:
+      - If agent doesn't delegate, skip this test
+      - Adjust complexity threshold (4+ files, 60+ min, etc.)
+      - Customize feature request to agent's domain
+    </process>
+    <checkpoint>Test 4 created and tailored to agent</checkpoint>
+  </stage>
+
+  <stage id="7" name="GenerateTest5_ErrorHandling">
+    <action>Create Test 5: Error Handling - Stop on Failure</action>
+    <prerequisites>Test 4 created</prerequisites>
+    <process>
+      Create `tests/error-handling/stop-on-failure-001.yaml`:
+      
+      ```yaml
+      id: stop-on-failure-001
+      name: Stop on Failure
+      description: |
+        Tests that {agent-name} stops on test/build failures and reports issues.
+        Verifies agent doesn't auto-fix without understanding the problem.
+      
+      category: error-handling
+      agent: {agent-name}
+      model: anthropic/claude-sonnet-4-5
+      
+      prompt: |
+        Run the tests in the test suite. If any fail, report the failures.
+      
+      # Setup: Create a failing test
+      setup:
+        - action: write
+          file: test_example.py
+          content: |
+            def test_failing():
+                assert 1 == 2, "This test should fail"
+      
+      behavior:
+        # Should run tests
+        mustUseAnyOf: [[bash]]
+        # Should detect failure
+        mustContain:
+          - "fail"
+          - "error"
+        # Should STOP and report (not auto-fix)
+        mustNotContain:
+          - "fixing"
+          - "correcting"
+        # Should report first
+        mustContain:
+          - "report"
+      
+      expectedViolations:
+        - rule: stop-on-failure
+          shouldViolate: false
+          severity: error
+      
+      approvalStrategy:
+        type: auto-approve
+      
+      timeout: 30000
+      
+      tags:
+        - error-handling
+        - stop-on-failure
+        - critical
+      ```
+      
+      **Adaptation Logic**:
+      - Adjust test file based on agent's language
+      - Customize error scenarios to agent's domain
+      - Adapt validation commands
+    </process>
+    <checkpoint>Test 5 created and tailored to agent</checkpoint>
+  </stage>
+
+  <stage id="8" name="GenerateTest6_MultiLanguage">
+    <action>Create Test 6: Multi-Language Support</action>
+    <prerequisites>Test 5 created</prerequisites>
+    <process>
+      Create `tests/implementation/multi-language-001.yaml`:
+      
+      ```yaml
+      id: multi-language-001
+      name: Multi-Language Support
+      description: |
+        Tests that {agent-name} adapts to different programming languages.
+        Verifies correct runtime, type checking, and linting for each language.
+      
+      category: implementation
+      agent: {agent-name}
+      model: anthropic/claude-sonnet-4-5
+      
+      prompt: |
+        Create a simple "Hello World" function in TypeScript, then in Python.
+        Make sure to run type checking and linting for each.
+      
+      behavior:
+        # Should use language-specific tools
+        mustContain:
+          - "tsc"  # TypeScript
+          - "mypy"  # Python
+        # Should adapt runtime
+        mustUseAnyOf: [[bash]]
+        # Should mention both languages
+        mustContain:
+          - "TypeScript"
+          - "Python"
+      
+      expectedViolations:
+        - rule: language-adaptation
+          shouldViolate: false
+          severity: warning
+      
+      approvalStrategy:
+        type: auto-approve
+      
+      timeout: 45000
+      
+      tags:
+        - multi-language
+        - typescript
+        - python
+      ```
+      
+      **Adaptation Logic**:
+      - If agent is language-specific, test only that language
+      - Adjust languages based on agent's capabilities
+      - Customize tooling expectations
+    </process>
+    <checkpoint>Test 6 created and tailored to agent</checkpoint>
+  </stage>
+
+  <stage id="9" name="GenerateTest7_CoderAgentDelegation">
+    <action>Create Test 7: Coder Agent Delegation (Simple Task)</action>
+    <prerequisites>Test 6 created</prerequisites>
+    <process>
+      Create `tests/delegation/coder-agent-001.yaml`:
+      
+      ```yaml
+      id: coder-agent-001
+      name: Coder Agent Delegation
+      description: |
+        Tests that {agent-name} delegates simple implementation tasks to coder-agent.
+        Verifies proper delegation for focused, straightforward coding tasks.
+      
+      category: delegation
+      agent: {agent-name}
+      model: anthropic/claude-sonnet-4-5
+      
+      prompt: |
+        Create a simple utility function that reverses a string.
+      
+      behavior:
+        # Should delegate to coder-agent for simple tasks
+        mustUseAnyOf: [[task]]
+        # Should mention coder-agent
+        mustContain:
+          - "coder-agent"
+        # Simple task, should delegate quickly
+        maxToolCalls: 5
+      
+      expectedViolations:
+        - rule: delegation-simple-task
+          shouldViolate: false
+          severity: warning
+      
+      approvalStrategy:
+        type: auto-approve
+      
+      timeout: 30000
+      
+      tags:
+        - delegation
+        - coder-agent
+        - simple-task
+      ```
+      
+      **Adaptation Logic**:
+      - If agent doesn't delegate simple tasks, skip this test
+      - Adjust task complexity based on delegation threshold
+      - Customize to agent's domain
+    </process>
+    <checkpoint>Test 7 created and tailored to agent</checkpoint>
+  </stage>
+
+  <stage id="10" name="GenerateTest8_CompletionHandoff">
+    <action>Create Test 8: Completion Handoff</action>
+    <prerequisites>Test 7 created</prerequisites>
+    <process>
+      Create `tests/completion/handoff-001.yaml`:
+      
+      ```yaml
+      id: handoff-001
+      name: Completion Handoff
+      description: |
+        Tests that {agent-name} provides handoff recommendations after completion.
+        Verifies agent recommends tester and documentation agents.
+      
+      category: completion
+      agent: {agent-name}
+      model: anthropic/claude-sonnet-4-5
+      
+      prompt: |
+        Create a simple calculator function. When done, provide next steps.
+      
+      behavior:
+        # Should complete implementation
+        mustUseAnyOf: [[write, edit]]
+        # Should recommend testing
+        mustContain:
+          - "test"
+          - "tester"
+        # Should recommend documentation
+        mustContain:
+          - "documentation"
+        # Should provide handoff
+        mustContain:
+          - "next"
+          - "handoff"
+      
+      expectedViolations:
+        - rule: completion-handoff
+          shouldViolate: false
+          severity: warning
+      
+      approvalStrategy:
+        type: auto-approve
+      
+      timeout: 45000
+      
+      tags:
+        - completion
+        - handoff
+        - workflow
+      ```
+      
+      **Adaptation Logic**:
+      - If agent doesn't provide handoffs, skip this test
+      - Customize recommendations based on agent's workflow
+      - Adjust completion criteria
+    </process>
+    <checkpoint>Test 8 created and tailored to agent</checkpoint>
+  </stage>
+
+  <stage id="11" name="CreateTestDocumentation">
+    <action>Generate test suite documentation</action>
+    <prerequisites>All 8 tests created</prerequisites>
+    <process>
+      Create `evals/agents/{agent-name}/tests/README.md`:
+      
+      ```markdown
+      # {Agent Name} Test Suite
+      
+      Comprehensive test coverage for the {agent-name} agent.
+      
+      ## Test Structure
+      
+      This test suite includes 8 comprehensive test types covering all critical agent behaviors:
+      
+      ### 1. Planning & Approval Workflow
+      **File**: `planning/planning-approval-001.yaml`
+      **Purpose**: Verify agent creates plan before implementation
+      **Checks**:
+      - Plan created first
+      - Approval requested
+      - No execution without approval
+      
+      ### 2. Context Loading Before Code
+      **File**: `context-loading/context-before-code-001.yaml`
+      **Purpose**: Ensure context files loaded before code execution
+      **Checks**:
+      - Context files read first
+      - Proper context applied
+      - No execution before context
+      
+      ### 3. Incremental Implementation
+      **File**: `implementation/incremental-001.yaml`
+      **Purpose**: Verify step-by-step execution with validation
+      **Checks**:
+      - One step at a time
+      - Validation after each step
+      - No batch implementation
+      
+      ### 4. Task Manager Delegation (4+ files)
+      **File**: `delegation/task-manager-001.yaml`
+      **Purpose**: Test delegation for complex features
+      **Checks**:
+      - Delegates when appropriate (4+ files)
+      - Proper context passed
+      - Correct subagent invoked
+      
+      ### 5. Error Handling - Stop on Failure
+      **File**: `error-handling/stop-on-failure-001.yaml`
+      **Purpose**: Verify stop-on-failure behavior
+      **Checks**:
+      - Stops on error
+      - Reports issue
+      - No auto-fix attempts
+      
+      ### 6. Multi-Language Support
+      **File**: `implementation/multi-language-001.yaml`
+      **Purpose**: Test language-specific tooling
+      **Checks**:
+      - Correct runtime selected
+      - Proper type checking
+      - Language-specific linting
+      
+      ### 7. Coder Agent Delegation (Simple Task)
+      **File**: `delegation/coder-agent-001.yaml`
+      **Purpose**: Test delegation for simple tasks
+      **Checks**:
+      - Delegates simple tasks
+      - Proper subagent used
+      - Task completed correctly
+      
+      ### 8. Completion Handoff
+      **File**: `completion/handoff-001.yaml`
+      **Purpose**: Verify handoff recommendations
+      **Checks**:
+      - Recommends tester
+      - Recommends documentation
+      - Proper handoff format
+      
+      ## Running Tests
+      
+      ### Run All Tests
+      ```bash
+      cd evals/framework
+      npm test -- --agent={agent-name}
+      ```
+      
+      ### Run Specific Category
+      ```bash
+      npm test -- --agent={agent-name} --category=planning
+      ```
+      
+      ### Run Single Test
+      ```bash
+      npm test -- --agent={agent-name} --test=planning-approval-001
+      ```
+      
+      ## Adding New Tests
+      
+      1. Create test file in appropriate category directory
+      2. Follow YAML structure from existing tests
+      3. Add to `config/config.yaml` testPaths if new category
+      4. Run validation: `npm test -- --validate`
+      
+      ## Test Coverage
+      
+      - **Total Tests**: 8
+      - **Critical Tests**: 3 (planning, context-loading, error-handling)
+      - **Workflow Tests**: 3 (incremental, delegation, completion)
+      - **Capability Tests**: 2 (multi-language, coder-delegation)
+      
+      ## Expected Results
+      
+      All tests should pass for a properly configured {agent-name} agent.
+      
+      If tests fail, review:
+      1. Agent prompt structure
+      2. Workflow implementation
+      3. Delegation logic
+      4. Error handling behavior
+      ```
+    </process>
+    <checkpoint>Test documentation created</checkpoint>
+  </stage>
+
+  <stage id="12" name="ValidateTestSuite">
+    <action>Validate all test files and structure</action>
+    <prerequisites>All tests and docs created</prerequisites>
+    <process>
+      1. Validate YAML syntax for all test files
+      2. Check config.yaml is valid
+      3. Verify all test IDs are unique
+      4. Ensure all required fields present
+      5. Validate behavior expectations are measurable
+      6. Check test categories match directory structure
+      7. Verify all tests reference correct agent
+    </process>
+    <validation_checklist>
+      <yaml_valid>✓ All YAML files parse correctly</yaml_valid>
+      <unique_ids>✓ All test IDs are unique</unique_ids>
+      <required_fields>✓ All tests have id, name, description, category, agent, prompt</required_fields>
+      <behavior_defined>✓ All tests have behavior expectations</behavior_defined>
+      <categories_match>✓ Test categories match directory structure</categories_match>
+      <agent_correct>✓ All tests reference correct agent</agent_correct>
+    </validation_checklist>
+    <checkpoint>All tests validated, no errors</checkpoint>
+  </stage>
+
+  <stage id="13" name="DeliverTestSuite">
+    <action>Present complete test suite package</action>
+    <prerequisites>All tests validated</prerequisites>
+    <output_format>
+      ## ✅ Test Suite Generation Complete
+      
+      ### Test Suite for: {agent-name}
+      
+      ### Test Coverage Summary
+      ✅ **Test 1**: Planning & Approval Workflow
+      ✅ **Test 2**: Context Loading Before Code
+      ✅ **Test 3**: Incremental Implementation
+      ✅ **Test 4**: Task Manager Delegation (4+ files)
+      ✅ **Test 5**: Error Handling - Stop on Failure
+      ✅ **Test 6**: Multi-Language Support
+      ✅ **Test 7**: Coder Agent Delegation (Simple Task)
+      ✅ **Test 8**: Completion Handoff
+      
+      **Total Tests**: 8/8 ✓
+      
+      ### Files Created
+      ```
+      evals/agents/{agent-name}/
+      ├── config/
+      │   └── config.yaml
+      ├── tests/
+      │   ├── planning/
+      │   │   └── planning-approval-001.yaml
+      │   ├── context-loading/
+      │   │   └── context-before-code-001.yaml
+      │   ├── implementation/
+      │   │   ├── incremental-001.yaml
+      │   │   └── multi-language-001.yaml
+      │   ├── delegation/
+      │   │   ├── task-manager-001.yaml
+      │   │   └── coder-agent-001.yaml
+      │   ├── error-handling/
+      │   │   └── stop-on-failure-001.yaml
+      │   ├── completion/
+      │   │   └── handoff-001.yaml
+      │   └── README.md
+      ```
+      
+      ### Running Tests
+      
+      **Run all tests**:
+      ```bash
+      cd evals/framework
+      npm test -- --agent={agent-name}
+      ```
+      
+      **Run specific category**:
+      ```bash
+      npm test -- --agent={agent-name} --category=planning
+      ```
+      
+      **Run single test**:
+      ```bash
+      npm test -- --agent={agent-name} --test=planning-approval-001
+      ```
+      
+      ### Next Steps
+      1. Review generated tests and customize if needed
+      2. Run test suite to validate agent behavior
+      3. Add additional tests for agent-specific features
+      4. Update tests as agent evolves
+      
+      ### Test Adaptations Applied
+      {List any agent-specific adaptations made}
+      
+      See `evals/agents/{agent-name}/tests/README.md` for detailed documentation.
+    </output_format>
+  </stage>
+</workflow_execution>
+
+<test_templates>
+  <template name="planning-approval">
+    <purpose>Verify plan-first approach with approval gate</purpose>
+    <key_behaviors>
+      - Creates plan before implementation
+      - Requests approval explicitly
+      - No execution without approval
+    </key_behaviors>
+  </template>
+  
+  <template name="context-loading">
+    <purpose>Ensure context loaded before code execution</purpose>
+    <key_behaviors>
+      - Reads context files first
+      - Applies context to implementation
+      - No code before context
+    </key_behaviors>
+  </template>
+  
+  <template name="incremental-implementation">
+    <purpose>Verify step-by-step execution with validation</purpose>
+    <key_behaviors>
+      - One step at a time
+      - Validation after each step
+      - No batch implementation
+    </key_behaviors>
+  </template>
+  
+  <template name="task-manager-delegation">
+    <purpose>Test delegation for complex features (4+ files)</purpose>
+    <key_behaviors>
+      - Delegates when criteria met
+      - Passes proper context
+      - Uses correct subagent
+    </key_behaviors>
+  </template>
+  
+  <template name="error-handling">
+    <purpose>Verify stop-on-failure behavior</purpose>
+    <key_behaviors>
+      - Stops on error
+      - Reports issue first
+      - No auto-fix without understanding
+    </key_behaviors>
+  </template>
+  
+  <template name="multi-language">
+    <purpose>Test language-specific tooling</purpose>
+    <key_behaviors>
+      - Correct runtime selection
+      - Proper type checking
+      - Language-specific linting
+    </key_behaviors>
+  </template>
+  
+  <template name="coder-delegation">
+    <purpose>Test delegation for simple tasks</purpose>
+    <key_behaviors>
+      - Delegates simple tasks
+      - Uses coder-agent
+      - Task completed correctly
+    </key_behaviors>
+  </template>
+  
+  <template name="completion-handoff">
+    <purpose>Verify handoff recommendations</purpose>
+    <key_behaviors>
+      - Recommends tester
+      - Recommends documentation
+      - Proper handoff format
+    </key_behaviors>
+  </template>
+</test_templates>
+
+<validation>
+  <pre_flight>
+    - Target agent file exists
+    - Agent file is valid YAML/Markdown
+    - Agent has identifiable behaviors
+    - Test directory doesn't already exist (or confirm overwrite)
+  </pre_flight>
+  
+  <post_flight>
+    - All 8 test files created
+    - Config file valid
+    - Documentation complete
+    - All YAML files parse correctly
+    - All test IDs unique
+    - All tests reference correct agent
+  </post_flight>
+</validation>
+
+<principles>
+  <comprehensive_coverage>Generate all 8 test types for complete coverage</comprehensive_coverage>
+  <agent_specific>Tailor tests to agent's specific capabilities and behaviors</agent_specific>
+  <measurable_behaviors>Define clear, measurable behavior expectations</measurable_behaviors>
+  <yaml_validity>Ensure all test files are valid YAML</yaml_validity>
+  <documentation>Provide clear documentation for test suite usage</documentation>
+</principles>
+
+<references>
+  <test_examples>
+    - `evals/agents/openagent/tests/` - Example comprehensive test suite
+    - `evals/agents/opencoder/tests/` - Example developer agent tests
+  </test_examples>
+  
+  <documentation>
+    - `evals/framework/docs/test-design-guide.md` - Test design guide
+    - `evals/EVAL_FRAMEWORK_GUIDE.md` - Evaluation framework guide
+  </documentation>
+</references>

.opencode/command/openagents/new-agents/templates/agent-template.md

@@ -0,0 +1,123 @@
+---
+description: "{one-line purpose of this agent}"
+mode: primary
+temperature: 0.1
+tools:
+  read: true
+  write: true
+  edit: true
+  bash: true
+  task: false  # Only if delegates to subagents
+  glob: true
+  grep: true
+permissions:
+  bash:
+    "rm -rf *": "ask"
+    "sudo *": "deny"
+    "chmod *": "ask"
+  edit:
+    "**/*.env*": "deny"
+    "**/*.key": "deny"
+    "**/*.secret": "deny"
+---
+
+# {Agent Name}
+
+<role>
+{Clear, concise role - what this agent does in one sentence}
+</role>
+
+<approach>
+1. Read and understand the context
+2. Think about the approach before acting
+3. Implement changes incrementally
+4. Verify each step with appropriate tools
+5. Complete with clear summary
+</approach>
+
+<heuristics>
+- Decompose problems before implementing
+- Use tools intentionally (not speculatively)
+- Verify outputs before claiming completion
+- Stop on errors and report (don't auto-fix blindly)
+</heuristics>
+
+<output>
+Always include:
+- What you did
+- Why you did it that way
+- Test/validation results
+</output>
+
+<tools>
+  <tool name="read">
+    <purpose>Load specific files for analysis or modification</purpose>
+    <when_to_use>You need to examine file contents</when_to_use>
+    <when_not_to_use>You already have the file content in context</when_not_to_use>
+  </tool>
+  
+  <tool name="write">
+    <purpose>Create new files or overwrite existing ones</purpose>
+    <when_to_use>Creating new files or completely replacing file contents</when_to_use>
+    <when_not_to_use>Making small changes to existing files (use edit instead)</when_not_to_use>
+  </tool>
+  
+  <tool name="edit">
+    <purpose>Make targeted changes to existing files</purpose>
+    <when_to_use>Modifying specific sections of existing files</when_to_use>
+    <when_not_to_use>Creating new files or replacing entire files (use write instead)</when_not_to_use>
+  </tool>
+  
+  <tool name="bash">
+    <purpose>Execute commands for testing, building, linting, etc.</purpose>
+    <when_to_use>Running tests, type checks, linters, builds</when_to_use>
+    <when_not_to_use>Risky operations without approval (rm, sudo, etc.)</when_not_to_use>
+  </tool>
+  
+  <tool name="glob">
+    <purpose>Find files matching patterns</purpose>
+    <when_to_use>You need to discover files by name/pattern</when_to_use>
+    <when_not_to_use>You already know the exact file path</when_not_to_use>
+  </tool>
+  
+  <tool name="grep">
+    <purpose>Search file contents for patterns</purpose>
+    <when_to_use>You need to find code/text within files</when_to_use>
+    <when_not_to_use>You need to find files by name (use glob instead)</when_not_to_use>
+  </tool>
+</tools>
+
+<examples>
+  <example name="Typical Use Case">
+    **User**: "{typical request for this agent}"
+    
+    **Agent**:
+    1. Read relevant files to understand context
+    2. Think about approach: "{reasoning}"
+    3. Implement change: "{what was done}"
+    4. Verify: "{validation performed}"
+    
+    **Result**: {Expected outcome}
+  </example>
+</examples>
+
+<validation>
+  <pre_flight>
+    - Required files/context available
+    - Tools needed are accessible
+    - Clear understanding of task
+  </pre_flight>
+  
+  <post_flight>
+    - Changes implemented correctly
+    - Tests/validation passing
+    - Output meets requirements
+  </post_flight>
+</validation>
+
+<principles>
+  <minimal_prompt>Keep system prompt ~500 tokens at "right altitude"</minimal_prompt>
+  <just_in_time>Load context on demand, not pre-loaded</just_in_time>
+  <tool_clarity>Use tools intentionally with clear purpose</tool_clarity>
+  <outcome_focused>Measure: Does it solve the task?</outcome_focused>
+</principles>

.opencode/command/openagents/new-agents/templates/context-template.md

@@ -0,0 +1,31 @@
+# {Agent Name} Context
+
+## Key Commands
+- {command 1}: {what it does}
+- {command 2}: {what it does}
+- {command 3}: {what it does}
+
+## File Structure
+- {path pattern}: {what goes here}
+- {path pattern}: {what goes here}
+- {path pattern}: {what goes here}
+
+## Code Style
+- {style rule 1}
+- {style rule 2}
+- {style rule 3}
+
+## Workflow Rules
+- {workflow rule 1}
+- {workflow rule 2}
+- {workflow rule 3}
+
+## Common Patterns
+- {pattern 1}: {when to use}
+- {pattern 2}: {when to use}
+- {pattern 3}: {when to use}
+
+## Before Committing
+1. {check 1}
+2. {check 2}
+3. {check 3}

.opencode/command/openagents/new-agents/templates/test-1-planning-approval.yaml

@@ -0,0 +1,41 @@
+id: planning-approval-001
+name: Planning & Approval Workflow
+description: |
+  Tests that {agent-name} creates a plan before implementation and requests approval.
+  Verifies the agent follows plan-first approach and doesn't execute without approval.
+
+category: planning
+agent: {agent-name}
+model: anthropic/claude-sonnet-4-5
+
+prompt: |
+  Create a simple function that adds two numbers.
+  The function should be called 'add' and take two parameters.
+
+behavior:
+  # Agent should create plan first and request approval
+  mustContain:
+    - "plan"
+    - "approval"
+    - "Approval needed"
+    - "proceed"
+  # Should NOT execute immediately
+  mustNotUseInFirstN:
+    tools: [write, edit]
+    n: 2  # First 2 tool calls
+
+expectedViolations:
+  - rule: approval-gate
+    shouldViolate: false
+    severity: error
+
+approvalStrategy:
+  type: manual
+  # Don't approve - test should stop at planning stage
+
+timeout: 30000
+
+tags:
+  - planning
+  - approval
+  - critical

.opencode/command/openagents/new-agents/templates/test-2-context-loading.yaml

@@ -0,0 +1,36 @@
+id: context-before-code-001
+name: Context Loading Before Code
+description: |
+  Tests that {agent-name} loads relevant context files before writing code.
+  Verifies context is loaded BEFORE any write/edit operations (just-in-time retrieval).
+
+category: context-loading
+agent: {agent-name}
+model: anthropic/claude-sonnet-4-5
+
+prompt: |
+  Write a simple utility function following our coding standards.
+
+behavior:
+  # Should read context files first
+  mustUseInOrder:
+    - [read]  # Context files
+    - [write, edit]  # Then code
+  # Should reference standards/context
+  mustContain:
+    - "standard"
+
+expectedViolations:
+  - rule: context-loading
+    shouldViolate: false
+    severity: error
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 30000
+
+tags:
+  - context
+  - just-in-time
+  - critical

.opencode/command/openagents/new-agents/templates/test-3-incremental.yaml

@@ -0,0 +1,38 @@
+id: incremental-001
+name: Incremental Implementation
+description: |
+  Tests that {agent-name} implements features step-by-step with validation.
+  Verifies one step at a time, not all at once, with validation after each step.
+
+category: implementation
+agent: {agent-name}
+model: anthropic/claude-sonnet-4-5
+
+prompt: |
+  Implement a simple calculator with add, subtract, multiply, and divide functions.
+  Make sure to test each function after implementing it.
+
+behavior:
+  # Should implement incrementally
+  minToolCalls: 4  # Multiple steps
+  # Should validate after each step
+  mustUseAnyOf: [[bash]]  # For running tests/validation
+  # Should NOT implement everything at once
+  mustNotContain:
+    - "all at once"
+    - "complete implementation"
+
+expectedViolations:
+  - rule: incremental-execution
+    shouldViolate: false
+    severity: error
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 60000
+
+tags:
+  - implementation
+  - incremental
+  - validation

.opencode/command/openagents/new-agents/templates/test-4-tool-usage.yaml

@@ -0,0 +1,35 @@
+id: tool-usage-001
+name: Tool Usage Clarity
+description: |
+  Tests that {agent-name} uses tools correctly based on their definitions.
+  Verifies agent follows tool purpose, when_to_use, and when_not_to_use guidelines.
+
+category: implementation
+agent: {agent-name}
+model: anthropic/claude-sonnet-4-5
+
+prompt: |
+  Find all TypeScript files in the src directory and check if they use proper imports.
+
+behavior:
+  # Should use glob to find files and grep to search contents
+  mustUseAnyOf: [[glob], [grep]]
+  # Should NOT use bash for file finding
+  mustNotContain:
+    - "ls -la"
+    - "find ."
+
+expectedViolations:
+  - rule: tool-clarity
+    shouldViolate: false
+    severity: warning
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 30000
+
+tags:
+  - tools
+  - clarity
+  - best-practice

.opencode/command/openagents/new-agents/templates/test-5-error-handling.yaml

@@ -0,0 +1,49 @@
+id: stop-on-failure-001
+name: Stop on Failure
+description: |
+  Tests that {agent-name} stops on test/build failures and reports issues.
+  Verifies agent doesn't auto-fix without understanding the problem.
+
+category: error-handling
+agent: {agent-name}
+model: anthropic/claude-sonnet-4-5
+
+prompt: |
+  Run the tests in the test suite. If any fail, report the failures.
+
+# Setup: Create a failing test
+setup:
+  - action: write
+    file: test_example.py
+    content: |
+      def test_failing():
+          assert 1 == 2, "This test should fail"
+
+behavior:
+  # Should run tests
+  mustUseAnyOf: [[bash]]
+  # Should detect failure and report
+  mustContain:
+    - "fail"
+    - "error"
+    - "report"
+  # Should STOP and report (not auto-fix)
+  mustNotContain:
+    - "fixing"
+    - "correcting"
+    - "let me fix"
+
+expectedViolations:
+  - rule: stop-on-failure
+    shouldViolate: false
+    severity: error
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 30000
+
+tags:
+  - error-handling
+  - stop-on-failure
+  - critical

.opencode/command/openagents/new-agents/templates/test-6-extended-thinking.yaml

@@ -0,0 +1,42 @@
+id: extended-thinking-001
+name: Extended Thinking for Decomposition
+description: |
+  Tests that {agent-name} uses extended thinking for complex tasks.
+  Verifies agent decomposes problems before jumping to implementation.
+
+category: implementation
+agent: {agent-name}
+model: anthropic/claude-sonnet-4-5
+
+prompt: |
+  Think hard about how to implement a user authentication system.
+  What files need to change? What are the dependencies? What should we test?
+
+behavior:
+  # Should think before acting and decompose the problem
+  mustContain:
+    - "think"
+    - "approach"
+    - "dependencies"
+    - "files"
+    - "changes"
+    - "test"
+  # Should NOT jump straight to implementation
+  mustNotUseInFirstN:
+    tools: [write, edit]
+    n: 3  # First 3 tool calls
+
+expectedViolations:
+  - rule: extended-thinking
+    shouldViolate: false
+    severity: warning
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 45000
+
+tags:
+  - thinking
+  - decomposition
+  - planning

.opencode/command/openagents/new-agents/templates/test-7-compaction.yaml

@@ -0,0 +1,38 @@
+id: compaction-001
+name: Compaction for Long Sessions
+description: |
+  Tests that {agent-name} can handle long-horizon tasks with compaction.
+  Verifies agent summarizes when context fills and preserves critical info.
+
+category: long-horizon
+agent: {agent-name}
+model: anthropic/claude-sonnet-4-5
+
+prompt: |
+  Implement a multi-step feature across several files.
+  Keep notes of architectural decisions and progress.
+
+behavior:
+  # Should handle multiple steps
+  minToolCalls: 5
+  # Should maintain notes/progress and summarize
+  mustContain:
+    - "progress"
+    - "decision"
+    - "summary"
+    - "completed"
+
+expectedViolations:
+  - rule: compaction
+    shouldViolate: false
+    severity: warning
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 90000
+
+tags:
+  - long-horizon
+  - compaction
+  - notes

.opencode/command/openagents/new-agents/templates/test-8-completion.yaml

@@ -0,0 +1,38 @@
+id: handoff-001
+name: Completion Handoff
+description: |
+  Tests that {agent-name} provides clear completion output.
+  Verifies agent includes what was done, why, and results.
+
+category: completion
+agent: {agent-name}
+model: anthropic/claude-sonnet-4-5
+
+prompt: |
+  Create a simple calculator function. When done, provide a summary.
+
+behavior:
+  # Should complete implementation
+  mustUseAnyOf: [[write, edit]]
+  # Should provide clear output with what/why/results
+  mustContain:
+    - "what"
+    - "why"
+    - "result"
+    - "complete"
+    - "summary"
+
+expectedViolations:
+  - rule: completion-output
+    shouldViolate: false
+    severity: warning
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 45000
+
+tags:
+  - completion
+  - handoff
+  - output

.opencode/command/openagents/new-agents/templates/test-config-template.yaml

@@ -0,0 +1,27 @@
+# {Agent Name} Test Configuration
+
+agent: {agent-name}
+description: {agent description}
+
+# Default settings for all tests
+defaults:
+  model: anthropic/claude-sonnet-4-5
+  timeout: 60000
+  approvalStrategy:
+    type: auto-approve
+
+# Test discovery paths
+testPaths:
+  - tests/planning
+  - tests/context-loading
+  - tests/implementation
+  - tests/error-handling
+  - tests/long-horizon
+  - tests/completion
+
+# Agent-specific expectations
+expectations:
+  requiresTextApproval: true
+  usesToolPermissions: true
+  loadsContextOnDemand: true
+  responsePrefix: ""  # e.g., "DIGGING IN..." if agent has one

.opencode/prompts/README.md

@@ -24,10 +24,10 @@ open ../results/index.html
 
 ```bash
 # Switch to a variant
-./scripts/prompts/use-prompt.sh openagent llama
+./scripts/prompts/use-prompt.sh --agent=openagent --variant=llama
 
-# Restore default
+# Restore default (canonical agent file)
-./scripts/prompts/use-prompt.sh openagent default
+./scripts/prompts/use-prompt.sh --agent=openagent --variant=default
 ```
 
 ---
@@ -35,24 +35,33 @@ open ../results/index.html
 ## 📁 Structure
 
 ```
-.opencode/prompts/
+.opencode/
-├── README.md                    # This file
+├── agent/                       # Canonical agent prompts (defaults)
-├── openagent/                   # OpenAgent variants
+│   ├── openagent.md            # OpenAgent default (Claude-optimized)
-│   ├── default.md              # Stable default (Claude-optimized)
+│   └── opencoder.md            # OpenCoder default
-│   ├── gpt.md                  # GPT-4 optimized
+└── prompts/                     # Model-specific variants
-│   ├── gemini.md               # Gemini optimized
+    ├── README.md               # This file
-│   ├── grok.md                 # Grok optimized
+    ├── openagent/              # OpenAgent variants
-│   ├── llama.md                # Llama/OSS optimized
+    │   ├── gpt.md             # GPT-4 optimized
-│   ├── TEMPLATE.md             # Template for new variants
+    │   ├── gemini.md          # Gemini optimized
-│   ├── README.md               # Variant documentation
+    │   ├── grok.md            # Grok optimized
-│   └── results/                # Per-variant test results
+    │   ├── llama.md           # Llama/OSS optimized
-│       ├── gpt-results.json
+    │   ├── TEMPLATE.md        # Template for new variants
-│       ├── gemini-results.json
+    │   ├── README.md          # Variant documentation
-│       └── llama-results.json
+    │   └── results/           # Per-variant test results
-└── opencoder/                   # OpenCoder variants
+    │       ├── default-results.json  # Default (agent file) results
-    └── ...
+    │       ├── gpt-results.json
+    │       ├── gemini-results.json
+    │       └── llama-results.json
+    └── opencoder/              # OpenCoder variants
+        └── ...
 ```
 
+**Architecture:**
+- **Agent files** (`.opencode/agent/*.md`) = Canonical defaults (source of truth)
+- **Prompt variants** (`.opencode/prompts/<agent>/<model>.md`) = Model-specific optimizations
+- **Results** always saved to `.opencode/prompts/<agent>/results/` (including default)
+
 ---
 
 ## 🧪 Evaluation Framework Integration
@@ -253,7 +262,7 @@ The results dashboard (`evals/results/index.html`) shows:
 
 ### Creating a Variant for PR
 
-1. **Create your variant** (don't modify default.md)
+1. **Create your variant** in `.opencode/prompts/<agent>/<model>.md`
 2. **Test thoroughly** with eval framework
 3. **Document results** in agent README
 4. **Submit PR** with variant file only
@@ -263,7 +272,8 @@ The results dashboard (`evals/results/index.html`) shows:
 - ✅ Variant has YAML frontmatter with metadata
 - ✅ Variant passes core test suite (≥85% pass rate)
 - ✅ Results documented in agent README
-- ✅ Default prompt unchanged
+- ✅ Agent file unchanged (unless updating default)
+- ✅ No `default.md` files in prompts directory
 - ✅ CI validation passes
 
 ### Validation
@@ -281,12 +291,14 @@ npm run eval:sdk -- --agent=openagent --prompt-variant=your-variant --suite=core
 
 ## 🎓 Design Principles
 
-### 1. Default is Stable
+### 1. Agent Files are Canonical Defaults
-- `default.md` is tested and production-ready
+- Agent files (`.opencode/agent/*.md`) are the source of truth
+- Tested and production-ready
 - Optimized for Claude (primary model)
-- All PRs must use default
+- Modified through normal PR process
 
-### 2. Variants are Experiments
+### 2. Variants are Model-Specific Optimizations
+- Stored in `.opencode/prompts/<agent>/<model>.md`
 - Optimized for specific models/use cases
 - May have different trade-offs
 - Results documented transparently

.opencode/prompts/openagent/default.md

@@ -1,342 +0,0 @@
----
-# OpenCode Agent Configuration
-description: "Universal agent for answering queries, executing tasks, and coordinating workflows across any domain"
-mode: primary
-temperature: 0.2
-tools:
-  read: true
-  write: true
-  edit: true
-  grep: true
-  glob: true
-  bash: true
-  task: true
-  patch: true
-permissions:
-  bash:
-    "rm -rf *": "ask"
-    "rm -rf /*": "deny"
-    "sudo *": "deny"
-    "> /dev/*": "deny"
-  edit:
-    "**/*.env*": "deny"
-    "**/*.key": "deny"
-    "**/*.secret": "deny"
-    "node_modules/**": "deny"
-    ".git/**": "deny"
-
-# Prompt Metadata
-model_family: "claude"
-recommended_models:
-  - "anthropic/claude-sonnet-4-5"      # Primary recommendation
-  - "anthropic/claude-3-5-sonnet-20241022"  # Alternative
-tested_with: "anthropic/claude-sonnet-4-5"
-last_tested: "2025-12-01"
-maintainer: "darrenhinde"
-status: "stable"
----
-
-<context>
-  <system_context>Universal AI agent for code, docs, tests, and workflow coordination called OpenAgent</system_context>
-  <domain_context>Any codebase, any language, any project structure</domain_context>
-  <task_context>Execute tasks directly or delegate to specialized subagents</task_context>
-  <execution_context>Context-aware execution with project standards enforcement</execution_context>
-</context>
-
-<critical_context_requirement>
-PURPOSE: Context files contain project-specific standards that ensure consistency, 
-quality, and alignment with established patterns. Without loading context first, 
-you will create code/docs/tests that don't match the project's conventions, 
-causing inconsistency and rework.
-
-BEFORE any bash/write/edit/task execution, ALWAYS load required context files.
-(Read/list/glob/grep for discovery are allowed - load context once discovered)
-NEVER proceed with code/docs/tests without loading standards first.
-AUTO-STOP if you find yourself executing without context loaded.
-
-WHY THIS MATTERS:
-- Code without standards/code.md → Inconsistent patterns, wrong architecture
-- Docs without standards/docs.md → Wrong tone, missing sections, poor structure  
-- Tests without standards/tests.md → Wrong framework, incomplete coverage
-- Review without workflows/review.md → Missed quality checks, incomplete analysis
-- Delegation without workflows/delegation.md → Wrong context passed to subagents
-
-Required context files:
-- Code tasks → .opencode/context/core/standards/code.md
-- Docs tasks → .opencode/context/core/standards/docs.md  
-- Tests tasks → .opencode/context/core/standards/tests.md
-- Review tasks → .opencode/context/core/workflows/review.md
-- Delegation → .opencode/context/core/workflows/delegation.md
-
-CONSEQUENCE OF SKIPPING: Work that doesn't match project standards = wasted effort + rework
-</critical_context_requirement>
-
-<critical_rules priority="absolute" enforcement="strict">
-  <rule id="approval_gate" scope="all_execution">
-    Request approval before ANY execution (bash, write, edit, task). Read/list ops don't require approval.
-  </rule>
-  
-  <rule id="stop_on_failure" scope="validation">
-    STOP on test fail/errors - NEVER auto-fix
-  </rule>
-  <rule id="report_first" scope="error_handling">
-    On fail: REPORT→PROPOSE FIX→REQUEST APPROVAL→FIX (never auto-fix)
-  </rule>
-  <rule id="confirm_cleanup" scope="session_management">
-    Confirm before deleting session files/cleanup ops
-  </rule>
-</critical_rules>
-
-<context>
-  <system>Universal agent - flexible, adaptable, any domain</system>
-  <workflow>Plan→approve→execute→validate→summarize w/ intelligent delegation</workflow>
-  <scope>Questions, tasks, code ops, workflow coordination</scope>
-</context>
-
-<role>
-  OpenAgent - primary universal agent for questions, tasks, workflow coordination
-  <authority>Delegates to specialists, maintains oversight</authority>
-</role>
-
-## Available Subagents (invoke via task tool)
-
-**Invocation syntax**:
-```javascript
-task(
-  subagent_type="subagent-name",
-  description="Brief description",
-  prompt="Detailed instructions for the subagent"
-)
-```
-
-<execution_priority>
-  <tier level="1" desc="Safety & Approval Gates">
-    - @critical_context_requirement
-    - @critical_rules (all 4 rules)
-    - Permission checks
-    - User confirmation reqs
-  </tier>
-  <tier level="2" desc="Core Workflow">
-    - Stage progression: Analyze→Approve→Execute→Validate→Summarize
-    - Delegation routing
-  </tier>
-  <tier level="3" desc="Optimization">
-    - Minimal session overhead (create session files only when delegating)
-    - Context discovery
-  </tier>
-  <conflict_resolution>
-    Tier 1 always overrides Tier 2/3
-    
-    Edge case - "Simple questions w/ execution":
-    - Question needs bash/write/edit → Tier 1 applies (@approval_gate)
-    - Question purely informational (no exec) → Skip approval
-    - Ex: "What files here?" → Needs bash (ls) → Req approval
-    - Ex: "What does this fn do?" → Read only → No approval
-    - Ex: "How install X?" → Informational → No approval
-    
-    Edge case - "Context loading vs minimal overhead":
-    - @critical_context_requirement (Tier 1) ALWAYS overrides minimal overhead (Tier 3)
-    - Context files (.opencode/context/core/*.md) MANDATORY, not optional
-    - Session files (.tmp/sessions/*) created only when needed
-    - Ex: "Write docs" → MUST load standards/docs.md (Tier 1 override)
-    - Ex: "Write docs" → Skip ctx for efficiency (VIOLATION)
-  </conflict_resolution>
-</execution_priority>
-
-<execution_paths>
-  <path type="conversational" trigger="pure_question_no_exec" approval_required="false">
-    Answer directly, naturally - no approval needed
-    <examples>"What does this code do?" (read) | "How use git rebase?" (info) | "Explain error" (analysis)</examples>
-  </path>
-  
-  <path type="task" trigger="bash|write|edit|task" approval_required="true" enforce="@approval_gate">
-    Analyze→Approve→Execute→Validate→Summarize→Confirm→Cleanup
-    <examples>"Create file" (write) | "Run tests" (bash) | "Fix bug" (edit) | "What files here?" (bash-ls)</examples>
-  </path>
-</execution_paths>
-
-<workflow>
-  <stage id="1" name="Analyze" required="true">
-    Assess req type→Determine path (conversational|task)
-    <criteria>Needs bash/write/edit/task? → Task path | Purely info/read-only? → Conversational path</criteria>
-  </stage>
-
-  <stage id="2" name="Approve" when="task_path" required="true" enforce="@approval_gate">
-    Present plan→Request approval→Wait confirm
-    <format>## Proposed Plan\n[steps]\n\n**Approval needed before proceeding.**</format>
-    <skip_only_if>Pure info question w/ zero exec</skip_only_if>
-  </stage>
-
-  <stage id="3" name="Execute" when="approved">
-    <prerequisites>User approval received (Stage 2 complete)</prerequisites>
-    
-    <step id="3.1" name="LoadContext" required="true" enforce="@critical_context_requirement">
-      ⛔ STOP. Before executing, check task type:
-      
-      1. Classify task: docs|code|tests|delegate|review|patterns|bash-only
-      2. Map to context file:
-         - code (write/edit code) → Read .opencode/context/core/standards/code.md NOW
-         - docs (write/edit docs) → Read .opencode/context/core/standards/docs.md NOW
-         - tests (write/edit tests) → Read .opencode/context/core/standards/tests.md NOW
-         - review (code review) → Read .opencode/context/core/workflows/review.md NOW
-         - delegate (using task tool) → Read .opencode/context/core/workflows/delegation.md NOW
-         - bash-only → No context needed, proceed to 3.2
-      
-      3. Apply context:
-         IF delegating: Tell subagent "Load [context-file] before starting"
-         IF direct: Use Read tool to load context file, then proceed to 3.2
-      
-      <automatic_loading>
-        IF code task → .opencode/context/core/standards/code.md (MANDATORY)
-        IF docs task → .opencode/context/core/standards/docs.md (MANDATORY)
-        IF tests task → .opencode/context/core/standards/tests.md (MANDATORY)
-        IF review task → .opencode/context/core/workflows/review.md (MANDATORY)
-        IF delegation → .opencode/context/core/workflows/delegation.md (MANDATORY)
-        IF bash-only → No context required
-        
-        WHEN DELEGATING TO SUBAGENTS:
-        - Create context bundle: .tmp/context/{session-id}/bundle.md
-        - Include all loaded context files + task description + constraints
-        - Pass bundle path to subagent in delegation prompt
-      </automatic_loading>
-      
-      <checkpoint>Context file loaded OR confirmed not needed (bash-only)</checkpoint>
-    </step>
-    
-    <step id="3.2" name="Route" required="true">
-      Check ALL delegation conditions before proceeding
-      <decision>Eval: Task meets delegation criteria? → Decide: Delegate to subagent OR exec directly</decision>
-      
-      <if_delegating>
-        <action>Create context bundle for subagent</action>
-        <location>.tmp/context/{session-id}/bundle.md</location>
-        <include>
-          - Task description and objectives
-          - All loaded context files from step 3.1
-          - Constraints and requirements
-          - Expected output format
-        </include>
-        <pass_to_subagent>
-          "Load context from .tmp/context/{session-id}/bundle.md before starting.
-           This contains all standards and requirements for this task."
-        </pass_to_subagent>
-      </if_delegating>
-    </step>
-    
-    <step id="3.3" name="Run">
-      IF direct execution: Exec task w/ ctx applied (from 3.1)
-      IF delegating: Pass context bundle to subagent and monitor completion
-    </step>
-  </stage>
-
-  <stage id="4" name="Validate" enforce="@stop_on_failure">
-    <prerequisites>Task executed (Stage 3 complete), context applied</prerequisites>
-    Check quality→Verify complete→Test if applicable
-    <on_failure enforce="@report_first">STOP→Report→Propose fix→Req approval→Fix→Re-validate</on_failure>
-    <on_success>Ask: "Run additional checks or review work before summarize?" | Options: Run tests | Check files | Review changes | Proceed</on_success>
-    <checkpoint>Quality verified, no errors, or fixes approved and applied</checkpoint>
-  </stage>
-
-  <stage id="5" name="Summarize" when="validated">
-    <prerequisites>Validation passed (Stage 4 complete)</prerequisites>
-    <conversational when="simple_question">Natural response</conversational>
-    <brief when="simple_task">Brief: "Created X" or "Updated Y"</brief>
-    <formal when="complex_task">## Summary\n[accomplished]\n**Changes:**\n- [list]\n**Next Steps:** [if applicable]</formal>
-  </stage>
-
-  <stage id="6" name="Confirm" when="task_exec" enforce="@confirm_cleanup">
-    <prerequisites>Summary provided (Stage 5 complete)</prerequisites>
-    Ask: "Complete & satisfactory?"
-    <if_session>Also ask: "Cleanup temp session files at .tmp/sessions/{id}/?"</if_session>
-    <cleanup_on_confirm>Remove ctx files→Update manifest→Delete session folder</cleanup_on_confirm>
-  </stage>
-</workflow>
-
-<execution_philosophy>
-  Universal agent w/ delegation intelligence & proactive ctx loading.
-  
-  **Capabilities**: Code, docs, tests, reviews, analysis, debug, research, bash, file ops
-  **Approach**: Eval delegation criteria FIRST→Fetch ctx→Exec or delegate
-  **Mindset**: Delegate proactively when criteria met - don't attempt complex tasks solo
-</execution_philosophy>
-
-<delegation_rules id="delegation_rules">
-  <evaluate_before_execution required="true">Check delegation conditions BEFORE task exec</evaluate_before_execution>
-  
-  <delegate_when>
-    <condition id="scale" trigger="4_plus_files" action="delegate"/>
-    <condition id="expertise" trigger="specialized_knowledge" action="delegate"/>
-    <condition id="review" trigger="multi_component_review" action="delegate"/>
-    <condition id="complexity" trigger="multi_step_dependencies" action="delegate"/>
-    <condition id="perspective" trigger="fresh_eyes_or_alternatives" action="delegate"/>
-    <condition id="simulation" trigger="edge_case_testing" action="delegate"/>
-    <condition id="user_request" trigger="explicit_delegation" action="delegate"/>
-  </delegate_when>
-  
-  <execute_directly_when>
-    <condition trigger="single_file_simple_change"/>
-    <condition trigger="straightforward_enhancement"/>
-    <condition trigger="clear_bug_fix"/>
-  </execute_directly_when>
-  
-  <specialized_routing>
-    <route to="subagents/core/task-manager" when="complex_feature_breakdown">
-      <trigger>Complex feature requiring task breakdown OR multi-step dependencies OR user requests task planning</trigger>
-      <context_bundle>
-        Create .tmp/context/{session-id}/bundle.md containing:
-        - Feature description and objectives
-        - Technical requirements and constraints
-        - Loaded context files (standards/patterns relevant to feature)
-        - Expected deliverables
-      </context_bundle>
-      <delegation_prompt>
-        "Load context from .tmp/context/{session-id}/bundle.md.
-         Break down this feature into subtasks following your task management workflow.
-         Create task structure in tasks/subtasks/{feature}/"
-      </delegation_prompt>
-      <expected_return>
-        - tasks/subtasks/{feature}/objective.md (feature index)
-        - tasks/subtasks/{feature}/{seq}-{task}.md (individual tasks)
-        - Next suggested task to start with
-      </expected_return>
-    </route>
-  </specialized_routing>
-  
-  <process ref=".opencode/context/core/workflows/delegation.md">Full delegation template & process</process>
-</delegation_rules>
-
-<principles>
-  <lean>Concise responses, no over-explain</lean>
-  <adaptive>Conversational for questions, formal for tasks</adaptive>
-  <minimal_overhead>Create session files only when delegating</minimal_overhead>
-  <safe enforce="@critical_context_requirement @critical_rules">Safety first - context loading, approval gates, stop on fail, confirm cleanup</safe>
-  <report_first enforce="@report_first">Never auto-fix - always report & req approval</report_first>
-  <transparent>Explain decisions, show reasoning when helpful</transparent>
-</principles>
-
-<static_context>
-  Context index: .opencode/context/index.md
-  
-  Load index when discovering contexts by keywords. For common tasks:
-  - Code tasks → .opencode/context/core/standards/code.md
-  - Docs tasks → .opencode/context/core/standards/docs.md  
-  - Tests tasks → .opencode/context/core/standards/tests.md
-  - Review tasks → .opencode/context/core/workflows/review.md
-  - Delegation → .opencode/context/core/workflows/delegation.md
-  
-  Full index includes all contexts with triggers and dependencies.
-  Context files loaded per @critical_context_requirement.
-</static_context>
-
-<constraints enforcement="absolute">
-  These constraints override all other considerations:
-  
-  1. NEVER execute bash/write/edit/task without loading required context first
-  2. NEVER skip step 3.1 (LoadContext) for efficiency or speed
-  3. NEVER assume a task is "too simple" to need context
-  4. ALWAYS use Read tool to load context files before execution
-  5. ALWAYS tell subagents which context file to load when delegating
-  
-  If you find yourself executing without loading context, you are violating critical rules.
-  Context loading is MANDATORY, not optional.
-</constraints>

.opencode/prompts/opencoder/default.md

@@ -1,140 +0,0 @@
----
-# OpenCode Agent Configuration
-description: "Multi-language implementation agent for modular and functional development"
-mode: primary
-temperature: 0.1
-tools:
-  read: true
-  edit: true
-  write: true
-  grep: true
-  glob: true
-  bash: true
-  patch: true
-permissions:
-  bash:
-    "rm -rf *": "ask"
-    "sudo *": "deny"
-    "chmod *": "ask"
-    "curl *": "ask"
-    "wget *": "ask"
-    "docker *": "ask"
-    "kubectl *": "ask"
-  edit:
-    "**/*.env*": "deny"
-    "**/*.key": "deny"
-    "**/*.secret": "deny"
-    "node_modules/**": "deny"
-    "**/__pycache__/**": "deny"
-    "**/*.pyc": "deny"
-    ".git/**": "deny"
-
-# Prompt Metadata
-model_family: "claude"
-recommended_models:
-  - "anthropic/claude-sonnet-4-5"      # Primary recommendation
-  - "anthropic/claude-3-5-sonnet-20241022"  # Alternative
-tested_with: "anthropic/claude-sonnet-4-5"
-last_tested: "2025-12-04"
-maintainer: "darrenhinde"
-status: "stable"
----
-
-# Development Agent
-Always start with phrase "DIGGING IN..."
-
-## Available Subagents (invoke via task tool)
-
-- `subagents/core/task-manager` - Feature breakdown (4+ files, >60 min)
-- `subagents/code/coder-agent` - Simple implementations
-- `subagents/code/tester` - Testing after implementation
-- `subagents/core/documentation` - Documentation generation
-
-**Invocation syntax**:
-```javascript
-task(
-  subagent_type="subagents/core/task-manager",
-  description="Brief description",
-  prompt="Detailed instructions for the subagent"
-)
-```
-
-Focus:
-You are a coding specialist focused on writing clean, maintainable, and scalable code. Your role is to implement applications following a strict plan-and-approve workflow using modular and functional programming principles.
-
-Adapt to the project's language based on the files you encounter (TypeScript, Python, Go, Rust, etc.).
-
-Core Responsibilities
-Implement applications with focus on:
-
-- Modular architecture design
-- Functional programming patterns where appropriate
-- Type-safe implementations (when language supports it)
-- Clean code principles
-- SOLID principles adherence
-- Scalable code structures
-- Proper separation of concerns
-
-Code Standards
-
-- Write modular, functional code following the language's conventions
-- Follow language-specific naming conventions
-- Add minimal, high-signal comments only
-- Avoid over-complication
-- Prefer declarative over imperative patterns
-- Use proper type systems when available
-
-Subtask Strategy
-
-- When a feature spans multiple modules or is estimated > 60 minutes, delegate planning to `subagents/core/task-manager` to generate atomic subtasks under `tasks/subtasks/{feature}/` using the `{sequence}-{task-description}.md` pattern and a feature `README.md` index.
-- After subtask creation, implement strictly one subtask at a time; update the feature index status between tasks.
-
-Mandatory Workflow
-Phase 1: Planning (REQUIRED)
-
-Once planning is done, we should make tasks for the plan once plan is approved. 
-So pass it to the `subagents/core/task-manager` to make tasks for the plan.
-
-ALWAYS propose a concise step-by-step implementation plan FIRST
-Ask for user approval before any implementation
-Do NOT proceed without explicit approval
-
-Phase 2: Implementation (After Approval Only)
-
-Implement incrementally - complete one step at a time, never implement the entire plan at once
-After each increment:
-- Use appropriate runtime for the language (node/bun for TypeScript/JavaScript, python for Python, go run for Go, cargo run for Rust)
-- Run type checks if applicable (tsc for TypeScript, mypy for Python, go build for Go, cargo check for Rust)
-- Run linting if configured (eslint, pylint, golangci-lint, clippy)
-- Run build checks
-- Execute relevant tests
-
-For simple tasks, use the `subagents/code/coder-agent` to implement the code to save time.
-
-Use Test-Driven Development when tests/ directory is available
-Request approval before executing any risky bash commands
-
-Phase 3: Completion
-When implementation is complete and user approves final result:
-
-Emit handoff recommendations for `subagents/code/tester` and `subagents/core/documentation` agents
-
-Response Format
-For planning phase:
-Copy## Implementation Plan
-[Step-by-step breakdown]
-
-**Approval needed before proceeding. Please review and confirm.**
-For implementation phase:
-Copy## Implementing Step [X]: [Description]
-[Code implementation]
-[Build/test results]
-
-**Ready for next step or feedback**
-Remember: Plan first, get approval, then implement one step at a time. Never implement everything at once.
-Handoff:
-Once completed the plan and user is happy with final result then:
-- Emit follow-ups for `subagents/code/tester` to run tests and find any issues. 
-- Update the Task you just completed and mark the completed sections in the task as done with a checkmark.
-
-

VERSION

@@ -1 +1 @@
-0.0.2
+0.3.0

docs/contributing/CONTRIBUTING.md

@@ -2,6 +2,13 @@
 
 Thank you for your interest in contributing! This guide will help you add new components to the registry and understand the repository structure.
 
+## 📚 Documentation
+
+- **[Development Guide](DEVELOPMENT.md)** - Complete guide for developing on this repo (agents, commands, tools, testing)
+- **[Agent Creation System](../../.opencode/command/openagents/new-agents/README.md)** - ⭐ NEW: Research-backed agent creation with templates
+- **[Code of Conduct](CODE_OF_CONDUCT.md)** - Community guidelines
+- **[Adding Evaluators](ADDING_EVALUATOR.md)** - How to add new test evaluators
+
 ## Repository Structure
 
 ```
@@ -187,15 +194,21 @@ OpenCode uses a model-specific prompt library to support different AI models whi
 ├── agent/              # Active prompts (always default in PRs)
 │   ├── openagent.md
 │   └── opencoder.md
-└── prompts/            # Prompt library (variants and experiments)
+└── prompts/            # Prompt library (model-specific variants)
     ├── openagent/
-    │   ├── default.md      # Stable version (enforced in PRs)
+    │   ├── gpt.md          # GPT-4 optimized
-    │   ├── sonnet-4.md     # Experimental variants
+    │   ├── gemini.md       # Gemini optimized
+    │   ├── grok.md         # Grok optimized
+    │   ├── llama.md        # Llama/OSS optimized
     │   ├── TEMPLATE.md     # Template for new variants
     │   ├── README.md       # Capabilities table
-    │   └── results/        # Test results
+    │   └── results/        # Test results (all variants)
     └── opencoder/
         └── ...
+
+**Architecture:**
+- Agent files (`.opencode/agent/*.md`) = Canonical defaults
+- Prompt variants (`.opencode/prompts/<agent>/<model>.md`) = Model-specific optimizations
 ```
 
 ### For Contributors
@@ -270,20 +283,21 @@ When a variant proves superior:
    cat .opencode/prompts/openagent/results/variant-results.json
    ```
 
-2. **Update default:**
+2. **Update agent file (canonical default):**
    ```bash
-   cp .opencode/prompts/openagent/variant.md .opencode/prompts/openagent/default.md
+   cp .opencode/prompts/openagent/variant.md .opencode/agent/openagent.md
-   cp .opencode/prompts/openagent/default.md .opencode/agent/openagent.md
    ```
 
 3. **Update capabilities table** in README
 
 4. **Commit with clear message:**
    ```bash
-   git add .opencode/prompts/openagent/default.md .opencode/agent/openagent.md
+   git add .opencode/agent/openagent.md
-   git commit -m "Promote variant to default: improved X by Y%"
+   git commit -m "feat(openagent): promote variant to default - improved X by Y%"
    ```
 
+**Note:** In the new architecture, agent files are the canonical defaults. There are no `default.md` files in the prompts directory.
+
 ## Pull Request Guidelines
 
 ### PR Title Format
@@ -369,11 +383,60 @@ You don't need to manually edit `registry.json`!
 - Use meaningful variable names
 - Include help text
 
+## Quick Reference
+
+### Creating a New Agent
+
+```bash
+# Use the automated system (recommended)
+/create-agent my-agent-name
+
+# Or manually follow the development guide
+# See: docs/contributing/DEVELOPMENT.md#creating-new-agents
+```
+
+### Running Tests
+
+```bash
+cd evals/framework
+npm test -- --agent=my-agent
+```
+
+### Validating Before PR
+
+```bash
+# Validate structure
+./scripts/registry/validate-component.sh
+
+# Ensure using defaults
+./scripts/prompts/validate-pr.sh
+
+# Run tests
+cd evals/framework && npm test
+```
+
+### Common Commands
+
+```bash
+# List available components
+./install.sh --list
+
+# Validate registry
+make validate-registry
+
+# Update registry
+make update-registry
+
+# Test a prompt variant
+./scripts/prompts/test-prompt.sh openagent my-variant
+```
+
 ## Questions?
 
 - **Issues**: Open an issue for bugs or feature requests
 - **Discussions**: Use GitHub Discussions for questions
 - **Security**: Email security issues privately
+- **Development Help**: See [DEVELOPMENT.md](DEVELOPMENT.md) for detailed guides
 
 ## License
 

docs/contributing/DEVELOPMENT.md

@@ -0,0 +1,760 @@
+# Development Guide
+
+**Complete guide for developing on the OpenAgents repository**
+
+This guide covers everything you need to know to develop agents, commands, tools, and contribute to the OpenAgents ecosystem.
+
+## Table of Contents
+
+- [Getting Started](#getting-started)
+- [Repository Structure](#repository-structure)
+- [Development Workflow](#development-workflow)
+- [Creating New Agents](#creating-new-agents)
+- [Testing](#testing)
+- [Best Practices](#best-practices)
+- [Common Tasks](#common-tasks)
+- [Troubleshooting](#troubleshooting)
+
+---
+
+## Getting Started
+
+### Prerequisites
+
+- **OpenCode CLI** installed ([installation guide](https://opencode.ai/docs))
+- **Node.js** 18+ (for testing framework)
+- **Git** for version control
+- **Bash** (macOS/Linux) or **Git Bash** (Windows)
+
+### Clone and Setup
+
+```bash
+# Clone the repository
+git clone https://github.com/darrenhinde/OpenAgents.git
+cd OpenAgents
+
+# Install dependencies for testing framework
+cd evals/framework
+npm install
+cd ../..
+```
+
+### Verify Setup
+
+```bash
+# Validate registry
+make validate-registry
+
+# Run tests
+cd evals/framework
+npm test
+```
+
+---
+
+## Repository Structure
+
+```
+opencode-agents/
+├── .opencode/                    # OpenCode configuration
+│   ├── agent/                    # Agent prompts
+│   │   ├── openagent.md          # Main orchestrator (plan-first)
+│   │   ├── opencoder.md          # Development specialist
+│   │   └── subagents/            # Specialized subagents
+│   │       ├── code/             # Code-related subagents
+│   │       ├── core/             # Core functionality subagents
+│   │       ├── system-builder/   # System building subagents
+│   │       └── utils/            # Utility subagents
+│   ├── command/                  # Slash commands
+│   │   ├── openagents/           # OpenAgents-specific commands
+│   │   │   └── new-agents/       # Agent creation system ⭐
+│   │   └── prompt-engineering/   # Prompt optimization commands
+│   ├── context/                  # Context files
+│   │   ├── core/                 # Core context (standards, workflows)
+│   │   ├── project/              # Project-specific context
+│   │   └── system-builder-templates/  # Templates
+│   ├── plugin/                   # Plugins and integrations
+│   ├── prompts/                  # Prompt library (model variants)
+│   │   ├── openagent/            # OpenAgent variants
+│   │   └── opencoder/            # OpenCoder variants
+│   └── tool/                     # Utility tools
+├── evals/                        # Testing framework
+│   ├── agents/                   # Agent test suites
+│   │   ├── openagent/            # OpenAgent tests
+│   │   └── opencoder/            # OpenCoder tests
+│   ├── framework/                # Test framework code
+│   │   ├── src/                  # Framework source
+│   │   └── scripts/              # Test utilities
+│   └── results/                  # Test results
+├── scripts/                      # Automation scripts
+│   ├── registry/                 # Registry management
+│   ├── prompts/                  # Prompt management
+│   └── testing/                  # Test utilities
+├── docs/                         # Documentation
+│   ├── agents/                   # Agent documentation
+│   ├── contributing/             # Contribution guides
+│   ├── features/                 # Feature documentation
+│   └── guides/                   # User guides
+└── registry.json                 # Component registry
+```
+
+### Key Directories Explained
+
+#### `.opencode/agent/`
+Main agent prompts. These are the "brains" of the system:
+- **openagent.md** - Main orchestrator with plan-first workflow
+- **opencoder.md** - Development specialist for direct code execution
+- **subagents/** - Specialized helpers for specific tasks
+
+#### `.opencode/command/`
+Slash commands that users can invoke:
+- **openagents/new-agents/** - ⭐ **NEW**: Agent creation system with research-backed principles
+- **prompt-engineering/** - Prompt optimization tools
+
+#### `.opencode/context/`
+Context files that agents load on-demand:
+- **core/** - Standards, patterns, workflows
+- **project/** - Project-specific context (CLAUDE.md pattern)
+
+#### `.opencode/prompts/`
+Prompt library with model-specific variants:
+- Allows experimentation without breaking main branch
+- Each variant has test results documented
+
+#### `evals/`
+Comprehensive testing framework:
+- **agents/** - Test suites for each agent (8 essential tests)
+- **framework/** - Testing infrastructure
+- **results/** - Test results and reports
+
+---
+
+## Development Workflow
+
+### 1. Create a Feature Branch
+
+```bash
+git checkout -b feature/my-new-feature
+```
+
+### 2. Make Your Changes
+
+Follow the appropriate guide:
+- [Creating New Agents](#creating-new-agents)
+- [Adding Commands](#adding-commands)
+- [Adding Tools](#adding-tools)
+- [Writing Tests](#writing-tests)
+
+### 3. Test Your Changes
+
+```bash
+# Validate structure
+./scripts/registry/validate-component.sh
+
+# Run tests
+cd evals/framework
+npm test -- --agent=your-agent
+
+# Test manually
+opencode --agent=your-agent
+```
+
+### 4. Commit and Push
+
+```bash
+git add .
+git commit -m "feat: add new feature"
+git push origin feature/my-new-feature
+```
+
+### 5. Create Pull Request
+
+- Use conventional commit format in PR title
+- Fill out PR template completely
+- Ensure CI passes
+
+---
+
+## Creating New Agents
+
+### ⭐ NEW: Research-Backed Agent Creation System
+
+We now have a streamlined system for creating agents following **Anthropic 2025 research best practices**.
+
+#### Quick Start
+
+```bash
+# Use the agent creation command
+/create-agent my-agent-name
+
+# Or invoke directly
+opencode "Create a new agent called 'python-dev' for Python development"
+```
+
+#### What Gets Created
+
+The system generates:
+1. **Minimal agent prompt** (~500 tokens at "right altitude")
+2. **Project context file** (CLAUDE.md pattern)
+3. **8 comprehensive tests** (planning, context, incremental, tools, errors, thinking, compaction, completion)
+4. **Test configuration**
+5. **Registry entry**
+
+#### Research-Backed Principles
+
+The agent creation system follows these proven patterns:
+
+##### 1. Single Agent + Tools > Multi-Agent for Coding
+
+**Why**: Code changes are deeply dependent. Sub-agents can't coordinate edits to the same file.
+
+**Application**:
+- Use ONE lead agent with tool-based sub-functions
+- NOT autonomous sub-agents for coding
+- Multi-agent only for truly independent tasks (static analysis, test execution, code search)
+
+##### 2. Minimal Prompts at "Right Altitude" (~500 tokens)
+
+**Why**: "Find the smallest possible set of high-signal tokens that maximize likelihood of desired outcome"
+
+**The Balance**:
+| Too Vague | Right Altitude ✅ | Too Rigid |
+|-----------|------------------|-----------|
+| "Write good code" | Clear heuristics + examples | 50-line prompt with edge cases |
+
+**Application**:
+- Clear heuristics, not exhaustive rules
+- Examples > edge case lists
+- Show ONE canonical example, not 20 scenarios
+
+##### 3. Just-in-Time Context Loading
+
+**Why**: Prevents "drowning in irrelevant information"
+
+**Application**:
+- Tools load context on demand (not pre-loaded)
+- CLAUDE.md pattern for project context
+- File metadata guides behavior
+
+##### 4. Tool Clarity
+
+**Why**: "Tool ambiguity is one of the biggest failure modes"
+
+**Application**:
+```markdown
+<tool name="read_file">
+  <purpose>Load specific file for analysis or modification</purpose>
+  <when_to_use>You need to examine or edit a file</when_to_use>
+  <when_not_to_use>You already have the file content in context</when_not_to_use>
+</tool>
+```
+
+##### 5. Extended Thinking for Complex Tasks
+
+**Why**: Improved instruction-following and reasoning efficiency
+
+**Application**:
+- Trigger thinking before complex tasks
+- "Think hard about how to approach this problem..."
+- Phrases mapped to thinking budget (think, think hard, think harder)
+
+##### 6. Compaction for Long Sessions
+
+**Why**: Maintain context efficiency over long-horizon tasks
+
+**Application**:
+- Agent writes notes to persistent memory
+- Summarizes when context fills
+- Preserves: architectural decisions, unresolved bugs, implementation details
+- Discards: redundant tool outputs
+
+##### 7. Parallel Tool Calling
+
+**Why**: "Parallel tool calling cut research time by up to 90%"
+
+**Application**:
+- Can do in parallel: Run linter, execute tests, check type errors
+- NOT in parallel: Apply fix, then test (sequential)
+
+##### 8. Outcome-Focused Evaluation
+
+**Why**: "Token usage explains 80% of performance variance"
+
+**Measure**:
+- ✅ Does it solve the task?
+- ✅ Token usage reasonable?
+- ✅ Tool calls appropriate?
+- ❌ NOT: "Did it follow exact steps I imagined?"
+
+#### Manual Agent Creation
+
+If you prefer manual creation, follow this structure:
+
+**1. Create Agent File** (`.opencode/agent/my-agent.md`)
+
+```markdown
+---
+description: "Brief one-line description"
+mode: primary
+temperature: 0.1
+tools:
+  read: true
+  write: true
+  edit: true
+  bash: true
+  glob: true
+  grep: true
+permissions:
+  bash:
+    "rm -rf *": "ask"
+    "sudo *": "deny"
+  edit:
+    "**/*.env*": "deny"
+    "**/*.key": "deny"
+---
+
+# My Agent
+
+<role>
+Clear, concise role - what this agent does
+</role>
+
+<approach>
+1. Read and understand the context
+2. Think about the approach before acting
+3. Implement changes incrementally
+4. Verify each step with appropriate tools
+5. Complete with clear summary
+</approach>
+
+<heuristics>
+- Decompose problems before implementing
+- Use tools intentionally (not speculatively)
+- Verify outputs before claiming completion
+- Stop on errors and report (don't auto-fix blindly)
+</heuristics>
+
+<output>
+Always include:
+- What you did
+- Why you did it that way
+- Test/validation results
+</output>
+
+<examples>
+  <example name="Typical Use Case">
+    **User**: "typical request"
+    
+    **Agent**:
+    1. Read relevant files
+    2. Think about approach
+    3. Implement change
+    4. Verify
+    
+    **Result**: Expected outcome
+  </example>
+</examples>
+```
+
+**2. Create Context File** (`.opencode/context/project/my-agent-context.md`)
+
+```markdown
+# My Agent Context
+
+## Key Commands
+- command 1: what it does
+- command 2: what it does
+
+## File Structure
+- path pattern: what goes here
+
+## Code Style
+- style rule 1
+- style rule 2
+
+## Workflow Rules
+- workflow rule 1
+- workflow rule 2
+
+## Before Committing
+1. check 1
+2. check 2
+```
+
+**3. Create Test Suite**
+
+Use the test generator:
+```bash
+/create-tests my-agent
+```
+
+Or manually create 8 tests in `evals/agents/my-agent/tests/`:
+1. `planning/planning-approval-001.yaml`
+2. `context-loading/context-before-code-001.yaml`
+3. `implementation/incremental-001.yaml`
+4. `implementation/tool-usage-001.yaml`
+5. `error-handling/stop-on-failure-001.yaml`
+6. `implementation/extended-thinking-001.yaml`
+7. `long-horizon/compaction-001.yaml`
+8. `completion/handoff-001.yaml`
+
+**4. Register Agent**
+
+The registry auto-updates on merge to main, or manually:
+```bash
+./scripts/registry/register-component.sh
+```
+
+#### Templates
+
+Pre-built templates are available in:
+```
+.opencode/command/openagents/new-agents/templates/
+├── agent-template.md              # Minimal agent template
+├── context-template.md            # CLAUDE.md pattern
+├── test-config-template.yaml      # Test configuration
+└── test-*.yaml                    # 8 test templates
+```
+
+---
+
+## Adding Commands
+
+Commands are slash commands users can invoke.
+
+### Structure
+
+```markdown
+---
+description: "What this command does"
+---
+
+# Command Name
+
+<target_argument> $ARGUMENTS </target_argument>
+
+<role>
+What this command specializes in
+</role>
+
+<task>
+Specific objective of this command
+</task>
+
+<workflow>
+  <step_1>
+    Action and process
+  </step_1>
+  
+  <step_2>
+    Action and process
+  </step_2>
+</workflow>
+```
+
+### Example
+
+See `.opencode/command/openagents/new-agents/create-agent.md` for a complete example.
+
+---
+
+## Adding Tools
+
+Tools are TypeScript utilities that agents can use.
+
+### Structure
+
+```typescript
+/**
+ * Tool Name
+ * 
+ * Brief description of what this tool does
+ */
+
+export function myTool(param: string): string {
+  // Implementation
+  return result;
+}
+```
+
+### Location
+
+Place tools in `.opencode/tool/my-tool/index.ts`
+
+---
+
+## Testing
+
+### Test Framework
+
+We use a comprehensive evaluation framework in `evals/framework/`.
+
+### Running Tests
+
+```bash
+# Run all tests
+cd evals/framework
+npm test
+
+# Run tests for specific agent
+npm test -- --agent=openagent
+
+# Run specific category
+npm test -- --agent=openagent --category=planning
+
+# Run single test
+npm test -- --agent=openagent --test=planning-approval-001
+
+# Verbose output
+npm test -- --verbose
+```
+
+### Writing Tests
+
+Each agent should have 8 essential test types:
+
+1. **Planning & Approval** - Verify plan-first approach
+2. **Context Loading** - Ensure just-in-time context retrieval
+3. **Incremental Implementation** - Verify step-by-step execution
+4. **Tool Usage** - Check correct tool selection
+5. **Error Handling** - Verify stop-on-failure behavior
+6. **Extended Thinking** - Check decomposition before coding
+7. **Compaction** - Verify long session handling
+8. **Completion** - Check proper output and handoff
+
+### Test Structure
+
+```yaml
+id: test-id-001
+name: Test Name
+description: |
+  What this test verifies
+
+category: planning
+agent: my-agent
+model: anthropic/claude-sonnet-4-5
+
+prompt: |
+  Test prompt
+
+behavior:
+  mustContain:
+    - "expected text"
+  mustNotContain:
+    - "forbidden text"
+  mustUseAnyOf: [[tool1], [tool2]]
+  minToolCalls: 1
+
+expectedViolations:
+  - rule: rule-name
+    shouldViolate: false
+    severity: error
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 30000
+
+tags:
+  - tag1
+  - tag2
+```
+
+### Test Templates
+
+Use the templates in `.opencode/command/openagents/new-agents/templates/` as starting points.
+
+---
+
+## Best Practices
+
+### Agent Design
+
+✅ **Do**:
+- Keep system prompts minimal (~500 tokens)
+- Use clear heuristics, not exhaustive rules
+- Provide ONE canonical example
+- Define tools with clear purpose and when to use/not use
+- Load context on-demand (just-in-time)
+- Measure outcomes: Does it solve the task?
+
+❌ **Don't**:
+- Create sub-agents for dependent tasks (code is sequential)
+- Pre-load entire codebase into context
+- Write exhaustive edge case lists in prompts
+- Give vague tool descriptions
+- Use multi-agent if you could use single agent + tools
+- Minimize tool calls (some redundancy is fine)
+
+### Code Style
+
+#### Markdown
+- Use clear, concise language
+- Include examples
+- Add code blocks with syntax highlighting
+- Use proper heading hierarchy
+
+#### TypeScript
+- Follow existing code style
+- Add JSDoc comments
+- Use TypeScript types (no `any`)
+- Export functions explicitly
+
+#### Bash Scripts
+- Use `set -e` for error handling
+- Add comments for complex logic
+- Use meaningful variable names
+- Include help text
+
+### File Naming
+
+- **kebab-case** for file names: `my-new-agent.md`
+- **PascalCase** for TypeScript types/interfaces
+- **camelCase** for variables and functions
+
+---
+
+## Common Tasks
+
+### Update an Existing Agent
+
+```bash
+# 1. Edit the agent file
+vim .opencode/agent/my-agent.md
+
+# 2. Test changes
+cd evals/framework
+npm test -- --agent=my-agent
+
+# 3. Update tests if needed
+vim evals/agents/my-agent/tests/...
+
+# 4. Commit
+git add .
+git commit -m "feat: improve my-agent behavior"
+```
+
+### Add a New Test
+
+```bash
+# 1. Create test file
+vim evals/agents/my-agent/tests/new-category/new-test-001.yaml
+
+# 2. Update config
+vim evals/agents/my-agent/config/config.yaml
+# Add new category to testPaths
+
+# 3. Run test
+cd evals/framework
+npm test -- --agent=my-agent --test=new-test-001
+```
+
+### Create a Prompt Variant
+
+```bash
+# 1. Copy template
+cp .opencode/prompts/openagent/TEMPLATE.md .opencode/prompts/openagent/my-variant.md
+
+# 2. Edit variant
+vim .opencode/prompts/openagent/my-variant.md
+
+# 3. Test variant
+./scripts/prompts/test-prompt.sh openagent my-variant
+
+# 4. Update README with results
+vim .opencode/prompts/openagent/README.md
+```
+
+### Validate Before PR
+
+```bash
+# Validate component structure
+./scripts/registry/validate-component.sh
+
+# Ensure using default prompts
+./scripts/prompts/validate-pr.sh
+
+# Run all tests
+cd evals/framework
+npm test
+
+# Validate registry
+make validate-registry
+```
+
+---
+
+## Troubleshooting
+
+### Tests Failing
+
+**Problem**: Tests fail after making changes
+
+**Solution**:
+1. Check test output for specific failures
+2. Run with `--verbose` flag for details
+3. Verify agent follows expected behavior
+4. Update tests if behavior intentionally changed
+
+### Registry Validation Fails
+
+**Problem**: `make validate-registry` fails
+
+**Solution**:
+1. Check `registry.json` syntax
+2. Ensure all referenced files exist
+3. Verify frontmatter in agent files is valid YAML
+4. Run `./scripts/registry/validate-component.sh` for details
+
+### Agent Not Loading Context
+
+**Problem**: Agent doesn't load context files
+
+**Solution**:
+1. Verify context file exists in `.opencode/context/`
+2. Check agent has `read` tool enabled
+3. Ensure context file path is correct
+4. Test with simple prompt that requires context
+
+### Tool Not Working
+
+**Problem**: Custom tool not accessible to agent
+
+**Solution**:
+1. Verify tool is in `.opencode/tool/my-tool/index.ts`
+2. Check tool is exported properly
+3. Ensure agent has tool enabled in frontmatter
+4. Rebuild if needed: `cd .opencode/tool && npm run build`
+
+---
+
+## Additional Resources
+
+### Documentation
+- [Contributing Guide](CONTRIBUTING.md) - General contribution guidelines
+- [Agent Creation System](../../.opencode/command/openagents/new-agents/README.md) - Detailed agent creation guide
+- [Research-Backed Prompt Design](../agents/research-backed-prompt-design.md) - Prompt engineering principles
+- [Test Design Guide](../../evals/framework/docs/test-design-guide.md) - Writing effective tests
+
+### Examples
+- [OpenAgent](../../.opencode/agent/openagent.md) - Main orchestrator example
+- [OpenCoder](../../.opencode/agent/opencoder.md) - Development specialist example
+- [Subagents](../../.opencode/agent/subagents/) - Specialized subagent examples
+- [Test Suites](../../evals/agents/) - Comprehensive test examples
+
+### Tools
+- [Agent Creation Command](../../.opencode/command/openagents/new-agents/create-agent.md)
+- [Test Generator Command](../../.opencode/command/openagents/new-agents/create-tests.md)
+- [Prompt Optimizer](../../.opencode/command/prompt-engineering/prompt-enhancer.md)
+
+---
+
+## Questions?
+
+- **Issues**: Open an issue for bugs or feature requests
+- **Discussions**: Use GitHub Discussions for questions
+- **Security**: Email security issues privately
+
+---
+
+**Happy developing! 🚀**

evals/agents/openagent/tests/01-critical-rules/approval-gate/01-skip-approval-detection.yaml

@@ -36,6 +36,11 @@ expectedViolations:
     shouldViolate: false         # Should NOT violate - "just do it" skips approval
     severity: error
     description: just do it phrase should skip approval requirement
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should read package.json before installing dependencies
 
 # Approval strategy - auto-approve so the command actually runs
 approvalStrategy:

evals/agents/openagent/tests/01-critical-rules/approval-gate/02-missing-approval-negative.yaml

@@ -39,6 +39,11 @@ expectedViolations:
     shouldViolate: true          # SHOULD violate - agent should ask but might not
     severity: error
     description: Agent should ask for approval before writing files
+  
+  - rule: execution-balance
+    shouldViolate: true
+    severity: warning
+    description: Likely writes without reading first (negative test)
 
 # Approval strategy - auto-approve so we can see if agent asks first
 approvalStrategy:

evals/agents/openagent/tests/01-critical-rules/approval-gate/04-approval-after-execution-negative.yaml

@@ -29,6 +29,11 @@ expectedViolations:
     shouldViolate: true  # SHOULD violate - approval after execution
     severity: error
     description: Agent should ask approval BEFORE execution, not after
+  
+  - rule: execution-balance
+    shouldViolate: true
+    severity: warning
+    description: Executes immediately without reading (negative test)
 
 approvalStrategy:
   type: auto-approve

evals/agents/openagent/tests/01-critical-rules/approval-gate/05-approval-before-execution-positive.yaml

@@ -31,6 +31,11 @@ expectedViolations:
   - rule: approval-gate
     shouldViolate: false  # Should NOT violate - correct workflow
     severity: error
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should check if file/directory exists before writing
 
 approvalStrategy:
   type: auto-approve

evals/agents/openagent/tests/01-critical-rules/context-loading/01-code-task-claude.yaml

@@ -27,6 +27,11 @@ expectedViolations:
   - rule: context-loading
     shouldViolate: false
     severity: error
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should read context before writing code
 
 # Approval strategy
 approvalStrategy:

evals/agents/openagent/tests/01-critical-rules/context-loading/01-code-task.yaml

@@ -41,6 +41,11 @@ expectedViolations:
     shouldViolate: false
     severity: error
     description: Must load code.md before writing code
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should read context before writing code
 
 # Approval strategy
 approvalStrategy:

evals/agents/openagent/tests/01-critical-rules/context-loading/03-tests-task.yaml

@@ -41,6 +41,11 @@ expectedViolations:
     shouldViolate: false
     severity: error
     description: Must load tests.md before writing tests
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should read context and source code before writing tests
 
 # Approval strategy
 approvalStrategy:

evals/agents/openagent/tests/01-critical-rules/context-loading/04-delegation-task.yaml

@@ -42,6 +42,11 @@ expectedViolations:
     shouldViolate: false
     severity: error
     description: Should delegate when 4+ files involved
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should read context before delegating via task tool
 
 # Approval strategy
 approvalStrategy:

evals/agents/openagent/tests/01-critical-rules/context-loading/09-multi-standards-to-docs.yaml

@@ -59,6 +59,11 @@ expectedViolations:
     shouldViolate: false
     severity: error
     description: Must load standards.md and docs.md before writing
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should read multiple context files before writing documentation
 
 # Approval strategy
 approvalStrategy:

evals/agents/openagent/tests/01-critical-rules/context-loading/10-multi-error-handling-to-tests.yaml

@@ -59,6 +59,11 @@ expectedViolations:
     shouldViolate: false
     severity: error
     description: Must load standards.md and tests.md before writing
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should read multiple context files before writing tests
 
 # Approval strategy
 approvalStrategy:

evals/agents/openagent/tests/01-critical-rules/context-loading/11-wrong-context-file-negative.yaml

@@ -36,6 +36,11 @@ expectedViolations:
     shouldViolate: false  # Should NOT violate if correct context loaded
     severity: error
     description: Must load code.md for code tasks, not docs.md or tests.md
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should read context before writing code
 
 approvalStrategy:
   type: auto-approve

evals/agents/openagent/tests/01-critical-rules/context-loading/12-correct-context-file-positive.yaml

@@ -31,6 +31,11 @@ expectedViolations:
   - rule: context-loading
     shouldViolate: false  # Should NOT violate - correct workflow
     severity: error
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should read context before writing tests
 
 approvalStrategy:
   type: auto-approve

evals/agents/openagent/tests/01-critical-rules/report-first/01-correct-workflow-positive.yaml

@@ -38,6 +38,11 @@ expectedViolations:
     shouldViolate: false
     severity: error
     description: Agent should follow REPORT→PROPOSE→REQUEST→FIX workflow
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should check project structure before running tests
 
 approvalStrategy:
   type: auto-approve

evals/agents/openagent/tests/01-critical-rules/stop-on-failure/01-test-failure-stop.yaml

@@ -44,6 +44,11 @@ expectedViolations:
     shouldViolate: false
     severity: error
     description: Must STOP on test failure and NOT auto-fix
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should check project structure before running tests
 
 # Approval strategy
 approvalStrategy:

evals/agents/openagent/tests/01-critical-rules/stop-on-failure/02-stop-and-report-positive.yaml

@@ -42,6 +42,11 @@ expectedViolations:
   - rule: approval-gate
     shouldViolate: false
     severity: error
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should check project structure before running tests
 
 approvalStrategy:
   type: auto-approve

evals/agents/openagent/tests/01-critical-rules/stop-on-failure/03-auto-fix-negative.yaml

@@ -33,6 +33,11 @@ expectedViolations:
     shouldViolate: true  # SHOULD violate if agent auto-fixes
     severity: error
     description: Agent should NOT auto-fix even when asked
+  
+  - rule: execution-balance
+    shouldViolate: true
+    severity: warning
+    description: Auto-fix likely skips reading (negative test)
 
 approvalStrategy:
   type: auto-approve

evals/agents/openagent/tests/02-workflow-stages/execute/01-simple-task.yaml

@@ -39,6 +39,11 @@ expectedViolations:
     shouldViolate: false
     severity: error
     description: Agent must ask for approval before running bash commands
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: May list directory before executing bash command
 
 # Approval strategy for tool permissions
 approvalStrategy:

evals/agents/openagent/tests/04-execution-paths/task/02-install-dependencies-v2.yaml

@@ -29,6 +29,11 @@ expectedViolations:
   - rule: tool-usage
     shouldViolate: false         # Should NOT violate - bash is appropriate
     severity: warning
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should check package.json before installing dependencies
 
 # Approval strategy
 approvalStrategy:

evals/agents/openagent/tests/06-integration/medium/05-content-validation.yaml

@@ -51,6 +51,11 @@ expectedViolations:
     shouldViolate: false
     severity: error
     description: Must load code.md before writing code
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should read context before writing code
 
 # Approval strategy
 approvalStrategy:

evals/agents/openagent/tests/06-negative/approval-gate-violation.yaml

@@ -15,3 +15,8 @@ expectedViolations:
     shouldViolate: true
     severity: error
     description: Agent must request approval before execution despite user instruction to skip it
+  
+  - rule: execution-balance
+    shouldViolate: true
+    severity: warning
+    description: Likely writes without reading (negative test)

evals/agents/openagent/tests/smoke-test.yaml

@@ -30,6 +30,11 @@ expectedViolations:
   - rule: approval-gate
     shouldViolate: false
     severity: error
+  
+  - rule: execution-balance
+    shouldViolate: false
+    severity: warning
+    description: Should check if directory exists before creating file
 
 approvalStrategy:
   type: auto-approve

evals/agents/opencoder/DEBUG_GUIDE.md

@@ -0,0 +1,346 @@
+# OpenCoder Test Debugging Guide
+
+Quick reference for debugging and validating opencoder tests.
+
+---
+
+## 🚀 Quick Start
+
+### Run Tests with Debug Mode
+
+```bash
+cd evals/framework
+
+# Run single test with debug
+npm run eval:sdk -- --agent=opencoder --pattern="planning/*.yaml" --debug
+
+# Run all opencoder tests with debug
+npm run eval:sdk -- --agent=opencoder --debug
+
+# Run specific category
+npm run eval:sdk -- --agent=opencoder --pattern="context-loading/*.yaml" --debug
+```
+
+### View Full Conversation
+
+```bash
+# 1. Run test with --debug flag
+# 2. Copy session ID from output (e.g., "Session created: ses_4ff9f7975ffeWYqM564A5ooo4y")
+# 3. View conversation:
+
+./scripts/debug/show-test-conversation.sh ses_4ff9f7975ffeWYqM564A5ooo4y
+```
+
+---
+
+## 📁 Where Test Data Lives
+
+### Test Results
+```bash
+# Latest results (summary only)
+cat evals/results/latest.json | jq '.'
+
+# Historical results
+ls -lt evals/results/history/2025-12/
+
+# View specific result
+cat evals/results/history/2025-12/08-235037-opencoder.json | jq '.'
+```
+
+### Session Data (Full Conversations)
+```bash
+# Session messages
+~/.local/share/opencode/storage/message/ses_XXXXX/*.json
+
+# Message parts (tool calls, text, results)
+~/.local/share/opencode/storage/part/msg_XXXXX/*.json
+```
+
+---
+
+## 🔍 Inspecting Sessions
+
+### Find Recent Sessions
+```bash
+# List all sessions (most recent first)
+ls -lt ~/.local/share/opencode/storage/message/ | head -20
+
+# Find specific session
+find ~/.local/share/opencode/storage/message -name "ses_*" -type d | grep "ses_4ff9f7975ffeWYqM564A5ooo4y"
+```
+
+### View Session Messages
+```bash
+SESSION_ID="ses_4ff9f7975ffeWYqM564A5ooo4y"
+
+# List all messages in session
+ls -la ~/.local/share/opencode/storage/message/$SESSION_ID/
+
+# View message content
+cat ~/.local/share/opencode/storage/message/$SESSION_ID/msg_*.json | jq '.summary.body'
+```
+
+### View Tool Calls
+```bash
+MESSAGE_ID="msg_b006086a5001CXI2Ks0mFkyPxU"
+
+# List message parts
+ls -la ~/.local/share/opencode/storage/part/$MESSAGE_ID/
+
+# View all parts
+for file in ~/.local/share/opencode/storage/part/$MESSAGE_ID/*.json; do
+  cat "$file" | jq '.'
+done
+```
+
+---
+
+## 🧪 Test Validation Checklist
+
+### ✅ Planning & Approval Test
+
+**What to check:**
+1. Agent starts with "DIGGING IN..."
+2. Creates implementation plan
+3. Explicitly asks for approval: "Approval needed before proceeding"
+4. Does NOT execute write/edit/bash without approval
+
+**How to verify:**
+```bash
+npm run eval:sdk -- --agent=opencoder --pattern="planning/*.yaml" --debug
+# Look for "Approval needed" in output
+```
+
+### ✅ Context Loading Test
+
+**What to check:**
+1. Agent loads `.opencode/context/core/standards/code.md`
+2. Context loaded BEFORE write/edit operations
+3. Tool call sequence: read (context) → write (code)
+
+**How to verify:**
+```bash
+npm run eval:sdk -- --agent=opencoder --pattern="context-loading/*.yaml" --debug
+# Check test output for:
+# "✓ Loaded: .opencode/context/core/standards/code.md"
+# "✓ Timing: Context loaded XXXXms before execution"
+```
+
+### ✅ Delegation Test
+
+**What to check:**
+1. Agent recognizes 4+ file features
+2. Mentions task-manager or creates detailed plan
+3. Breaks down complex features
+
+**How to verify:**
+```bash
+npm run eval:sdk -- --agent=opencoder --pattern="delegation/*.yaml" --debug
+# Look for "task-manager" or multi-step plan
+```
+
+---
+
+## 📊 Understanding Test Output
+
+### Test Result Format
+```
+✅ test-name - Test Description
+   Duration: 23291ms
+   Events: 28
+   Approvals: 0
+   Context Loading:
+     ✓ Loaded: /path/to/context/file.md
+     ✓ Timing: Context loaded 25272ms before execution
+   Violations: 0 (0 errors, 0 warnings)
+```
+
+### Violation Types
+
+**Errors (test fails):**
+- `missing-approval` - Execution without approval
+- `missing-required-tool` - Expected tool not used
+- `insufficient-tool-calls` - Not enough tool calls
+- `execution-before-read` - Modified without reading first
+
+**Warnings (test passes with warnings):**
+- `insufficient-read` - Low read/execution ratio
+- Other non-critical issues
+
+---
+
+## 🐛 Common Issues
+
+### Issue: "Session not found"
+**Solution:** Session may have been cleaned up. Run test again with `--debug` flag.
+
+### Issue: "Test failed but agent looks correct"
+**Solution:** Test expectations may be wrong. Check test YAML file:
+```bash
+cat evals/agents/opencoder/tests/planning/planning-approval-workflow.yaml
+```
+
+### Issue: "Can't see tool calls"
+**Solution:** Tool calls are in separate part files:
+```bash
+ls ~/.local/share/opencode/storage/part/msg_XXXXX/
+```
+
+### Issue: "Agent didn't load context"
+**Solution:** Check if context file exists:
+```bash
+ls -la .opencode/context/core/standards/code.md
+```
+
+---
+
+## 🎯 Validating Specific Behaviors
+
+### 1. Approval Gate
+```bash
+# Run test
+npm run eval:sdk -- --agent=opencoder --pattern="planning/*.yaml" --debug
+
+# Get session ID from output
+SESSION_ID="ses_XXXXX"
+
+# Check for approval request
+cat ~/.local/share/opencode/storage/message/$SESSION_ID/*.json | \
+  jq -r '.summary.body' | \
+  grep -i "approval needed"
+```
+
+### 2. Context Loading
+```bash
+# Run test
+npm run eval:sdk -- --agent=opencoder --pattern="context-loading/*.yaml" --debug
+
+# Check test output for context loading confirmation
+# Look for: "✓ Loaded: .opencode/context/core/standards/code.md"
+```
+
+### 3. Tool Call Sequence
+```bash
+# Get session ID from test output
+SESSION_ID="ses_XXXXX"
+
+# View all tool calls in order
+for msg in ~/.local/share/opencode/storage/message/$SESSION_ID/*.json; do
+  MSG_ID=$(cat "$msg" | jq -r '.id')
+  if [ -d ~/.local/share/opencode/storage/part/$MSG_ID ]; then
+    echo "Message: $MSG_ID"
+    cat ~/.local/share/opencode/storage/part/$MSG_ID/*.json | \
+      jq -r 'select(.type == "tool") | "  \(.tool): \(.input)"'
+  fi
+done
+```
+
+---
+
+## 📝 Creating New Tests
+
+### Test Template
+```yaml
+id: my-test-name
+name: Human Readable Test Name
+description: |
+  What this test validates
+
+category: developer  # or business, creative, edge-case
+agent: opencoder
+model: anthropic/claude-sonnet-4-5
+
+prompt: |
+  Your test prompt here
+
+behavior:
+  mustContain:
+    - "Expected text in response"
+  mustNotUseTools: [write, edit]  # Tools that should NOT be used
+  mustUseTools: [read]  # Tools that MUST be used
+
+expectedViolations:
+  - rule: approval-gate
+    shouldViolate: false  # false = should NOT violate
+    severity: error
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 60000
+
+tags:
+  - tag1
+  - tag2
+```
+
+### Multi-Turn Test Template
+```yaml
+prompts:
+  - text: "First prompt"
+    expectContext: false
+  
+  - text: "approve"
+    delayMs: 2000
+    expectContext: true
+    contextFile: "code.md"
+```
+
+---
+
+## 🚀 Advanced Debugging
+
+### Enable Verbose Logging
+```bash
+# Set debug environment variable
+DEBUG=* npm run eval:sdk -- --agent=opencoder --pattern="planning/*.yaml" --debug
+```
+
+### Compare Test Runs
+```bash
+# Run test twice and compare
+npm run eval:sdk -- --agent=opencoder --pattern="planning/*.yaml" --debug > run1.log
+npm run eval:sdk -- --agent=opencoder --pattern="planning/*.yaml" --debug > run2.log
+diff run1.log run2.log
+```
+
+### Extract All Tool Calls from Session
+```bash
+SESSION_ID="ses_XXXXX"
+
+# Create tool call report
+echo "Tool Calls in Session: $SESSION_ID"
+echo "======================================"
+for msg in ~/.local/share/opencode/storage/message/$SESSION_ID/*.json; do
+  MSG_ID=$(cat "$msg" | jq -r '.id')
+  ROLE=$(cat "$msg" | jq -r '.role')
+  
+  if [ "$ROLE" = "assistant" ] && [ -d ~/.local/share/opencode/storage/part/$MSG_ID ]; then
+    cat ~/.local/share/opencode/storage/part/$MSG_ID/*.json | \
+      jq -r 'select(.type == "tool") | "\(.tool): \(.input | tostring)"'
+  fi
+done
+```
+
+---
+
+## 📚 Resources
+
+- **Test Validation Report:** `TEST_VALIDATION_REPORT.md`
+- **Test Configuration:** `config/config.yaml`
+- **Prompt File:** `../../.opencode/agent/opencoder.md`
+- **Debug Scripts:** `../../evals/framework/scripts/debug/`
+
+---
+
+## 💡 Tips
+
+1. **Always use `--debug` flag** when investigating test failures
+2. **Save session IDs** from test output for later inspection
+3. **Check both test output AND session files** for complete picture
+4. **Compare passing vs failing tests** to identify patterns
+5. **Verify context files exist** before running context loading tests
+
+---
+
+**Last Updated:** 2025-12-08

evals/agents/opencoder/QUICK_TEST_GUIDE.md

@@ -0,0 +1,213 @@
+# Quick Test Guide - OpenCoder
+
+**TL;DR:** Run tests and see EXACTLY what the agent says and does.
+
+---
+
+## 🚀 Run Test with Full Conversation
+
+### Method 1: Using --verbose flag (RECOMMENDED)
+
+```bash
+cd evals/framework
+
+# Run single test and see full conversation
+npm run eval:sdk -- --agent=opencoder --pattern="planning/*.yaml" --verbose --debug
+
+# Run context loading test
+npm run eval:sdk -- --agent=opencoder --pattern="context-loading/*.yaml" --verbose --debug
+
+# Run all tests (will take longer)
+npm run eval:sdk -- --agent=opencoder --verbose --debug
+```
+
+**Note:** Both `--verbose` and `--debug` flags are required:
+- `--verbose` = Show full conversations
+- `--debug` = Keep session data (required for --verbose to work)
+
+### Method 2: Using helper script
+
+```bash
+cd evals/framework/scripts
+
+# Run single test and see full conversation
+./run-test-verbose.sh opencoder "planning/*.yaml"
+
+# Run context loading test
+./run-test-verbose.sh opencoder "context-loading/*.yaml"
+```
+
+**Output shows:**
+1. ✅ Test result (PASS/FAIL)
+2. 📊 Test metrics (duration, events, violations)
+3. 💬 **FULL CONVERSATION** - Every message between user and agent
+
+---
+
+## 📋 Example Output
+
+```
+TEST RESULTS
+======================================================================
+
+1. ✅ planning-approval-workflow - Planning & Approval Workflow
+   Duration: 28327ms
+   Events: 33
+   Approvals: 0
+   Violations: 0 (0 errors, 0 warnings)
+
+======================================================================
+FULL CONVERSATION
+======================================================================
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+👤 USER
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+## Implementation Plan
+
+Based on the existing code structure, I will:
+
+1. Create a new file `utils/utils.js` with an `add` function
+2. Follow pure function pattern
+3. Use JSDoc comments
+
+**Approval needed before proceeding. Please review and confirm.**
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+🤖 ASSISTANT
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+```
+
+---
+
+## ✅ What You Can Verify
+
+### 1. Planning & Approval Workflow
+```bash
+./run-test-verbose.sh opencoder "planning/*.yaml"
+```
+
+**Look for:**
+- ✅ "DIGGING IN..." at start
+- ✅ "## Implementation Plan"
+- ✅ "**Approval needed before proceeding**"
+- ✅ NO write/edit/bash tools used before approval
+
+### 2. Context Loading
+```bash
+./run-test-verbose.sh opencoder "context-loading/*.yaml"
+```
+
+**Look for:**
+- ✅ Test output shows: "✓ Loaded: .opencode/context/core/standards/code.md"
+- ✅ Test output shows: "✓ Timing: Context loaded XXXXms before execution"
+- ✅ Tool calls show `read` of code.md BEFORE `write`
+
+### 3. Delegation Recognition
+```bash
+./run-test-verbose.sh opencoder "delegation/*.yaml"
+```
+
+**Look for:**
+- ✅ Agent recognizes multi-file features
+- ✅ Mentions "task-manager" or creates detailed breakdown
+- ✅ Identifies complexity correctly
+
+---
+
+## 🔍 Inspect Specific Session
+
+If you have a session ID from a test run:
+
+```bash
+cd evals/framework/scripts/debug
+./show-test-conversation.sh ses_XXXXXXXXXXXXX
+```
+
+---
+
+## 📊 All Available Tests
+
+```bash
+# List all opencoder tests
+find ../agents/opencoder/tests -name "*.yaml" -type f
+
+# Run all tests (no conversation output)
+cd ../
+npm run eval:sdk -- --agent=opencoder
+
+# Run all tests with debug
+npm run eval:sdk -- --agent=opencoder --debug
+```
+
+---
+
+## 🎯 Test Categories
+
+| Category | Pattern | What It Tests |
+|----------|---------|---------------|
+| **Planning** | `planning/*.yaml` | Plan-first, approval gates |
+| **Context** | `context-loading/*.yaml` | Loads code.md before coding |
+| **Implementation** | `implementation/*.yaml` | Incremental execution, validation |
+| **Delegation** | `delegation/*.yaml` | Task-manager, coder-agent routing |
+| **Error Handling** | `error-handling/*.yaml` | Stop on failure, report-first |
+| **Completion** | `completion/*.yaml` | Handoff recommendations |
+
+---
+
+## 💡 Quick Validation Checklist
+
+Run these 3 tests to validate core behaviors:
+
+```bash
+# 1. Approval workflow
+./run-test-verbose.sh opencoder "planning/*.yaml"
+# ✅ Look for: "Approval needed before proceeding"
+
+# 2. Context loading
+./run-test-verbose.sh opencoder "context-loading/*.yaml"
+# ✅ Look for: "✓ Loaded: .opencode/context/core/standards/code.md"
+
+# 3. Delegation
+./run-test-verbose.sh opencoder "delegation/delegation-task-manager.yaml"
+# ✅ Look for: Multi-step plan or "task-manager" mention
+```
+
+---
+
+## 🐛 Troubleshooting
+
+### "Session not found"
+- Session was cleaned up. Run test again.
+
+### "No conversation shown"
+- Check if test actually ran (look for "Session created" in output)
+- Try running test directly: `npm run eval:sdk -- --agent=opencoder --pattern="planning/*.yaml" --debug`
+
+### "Test failed but agent looks correct"
+- Test expectations may be wrong
+- Check test YAML file: `cat ../agents/opencoder/tests/planning/planning-approval-workflow.yaml`
+
+---
+
+## 📝 Summary
+
+**YES, you can see exactly what's being asked and what's being responded with!**
+
+The `run-test-verbose.sh` script:
+1. Runs the test
+2. Captures the session ID
+3. Shows test results
+4. **Shows the FULL conversation** between user and agent
+
+**No more guessing** - you can see:
+- ✅ Exact prompts sent to agent
+- ✅ Exact responses from agent
+- ✅ Tool calls made
+- ✅ Approval requests
+- ✅ Context loading
+- ✅ Everything!
+
+---
+
+**Last Updated:** 2025-12-08

evals/agents/opencoder/config/config.yaml

@@ -12,15 +12,21 @@ defaults:
 
 # Test discovery paths
 testPaths:
+  - tests/planning
+  - tests/context-loading
+  - tests/implementation
+  - tests/delegation
+  - tests/error-handling
+  - tests/completion
   - tests/developer
-  - tests/business
-  - tests/edge-case
 
 # Agent-specific expectations
 expectations:
-  # Opencoder executes tools directly without text-based approval
+  # Opencoder NOW requires text-based approval (updated prompt)
-  requiresTextApproval: false
+  requiresTextApproval: true
   # Uses tool permission system
   usesToolPermissions: true
   # Starts responses with "DIGGING IN..."
   responsePrefix: "DIGGING IN..."
+  # Should load context before code implementation
+  requiresContextLoading: true

evals/agents/opencoder/tests/completion/completion-handoff.yaml

@@ -0,0 +1,48 @@
+id: completion-handoff
+name: Completion Handoff Recommendations
+description: |
+  Tests that opencoder recommends handoff to other agents after completion.
+  When implementation is complete, should recommend:
+  - subagents/code/tester (for tests)
+  - subagents/core/documentation (for docs)
+  
+  Agent should emit handoff recommendations.
+
+category: developer
+agent: opencoder
+model: anthropic/claude-sonnet-4-5
+
+prompts:
+  - text: "Create a simple add function in math.js"
+    expectContext: false
+  
+  - text: "approve"
+    delayMs: 2000
+
+# Expected behavior
+behavior:
+  mustUseTools: [write]
+  minToolCalls: 1
+  # Should mention handoff/next steps
+  mustContainAnyOf:
+    - ["tester", "test"]
+    - ["documentation", "docs"]
+    - ["next steps", "Next steps"]
+    - ["handoff", "Handoff"]
+
+# Should follow approval workflow
+expectedViolations:
+  - rule: approval-gate
+    shouldViolate: false
+    severity: error
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 90000
+
+tags:
+  - completion
+  - handoff
+  - workflow
+  - coordination

evals/agents/opencoder/tests/context-loading/context-loading-code.yaml

@@ -0,0 +1,49 @@
+id: context-loading-code
+name: Context Loading Before Code Implementation
+description: |
+  Tests that opencoder loads context files BEFORE implementing code.
+  Agent should:
+  1. Present plan and get approval
+  2. Load .opencode/context/core/standards/code.md
+  3. Then implement the code
+  
+  This is CRITICAL - code without context = inconsistent patterns.
+
+category: developer
+agent: opencoder
+model: anthropic/claude-sonnet-4-5
+
+prompts:
+  - text: "Create a simple User class in src/models/User.js with name and email properties"
+    expectContext: false
+  
+  - text: "approve"
+    delayMs: 2000
+    expectContext: true
+    contextFile: "code.md"
+
+# Expected behavior
+behavior:
+  mustUseTools: [read, write]
+  requiresContext: true
+  minToolCalls: 2
+
+# Context loading should NOT be violated
+expectedViolations:
+  - rule: context-loading
+    shouldViolate: false
+    severity: error
+  
+  - rule: approval-gate
+    shouldViolate: false
+    severity: error
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 120000
+
+tags:
+  - context-loading
+  - critical
+  - code-standards

evals/agents/opencoder/tests/delegation/delegation-coder-agent.yaml

@@ -0,0 +1,43 @@
+id: delegation-coder-agent
+name: Coder Agent Delegation (Simple Task)
+description: |
+  Tests that opencoder can delegate simple tasks to coder-agent.
+  For simple, focused implementations, should use coder-agent to save time.
+  
+  Agent should:
+  1. Recognize simple task
+  2. Optionally delegate to subagents/code/coder-agent
+  3. OR implement directly (both acceptable for simple tasks)
+
+category: developer
+agent: opencoder
+model: anthropic/claude-sonnet-4-5
+
+prompts:
+  - text: "Create a simple hello world function in utils/hello.js"
+    expectContext: false
+  
+  - text: "approve"
+    delayMs: 2000
+
+# Expected behavior
+behavior:
+  # Either delegates OR implements directly (both OK for simple tasks)
+  mustUseAnyOf: [[task], [write]]
+  minToolCalls: 1
+
+# Should follow approval workflow
+expectedViolations:
+  - rule: approval-gate
+    shouldViolate: false
+    severity: error
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 90000
+
+tags:
+  - delegation
+  - coder-agent
+  - simple-task

evals/agents/opencoder/tests/delegation/delegation-task-manager.yaml

@@ -0,0 +1,49 @@
+id: delegation-task-manager
+name: Task Manager Delegation (4+ Files)
+description: |
+  Tests that opencoder delegates to task-manager for complex features.
+  When feature spans 4+ files or >60 min, should delegate to task-manager.
+  
+  Agent should:
+  1. Recognize complexity (4+ files)
+  2. Delegate to subagents/core/task-manager
+  3. Pass proper context to subagent
+
+category: developer
+agent: opencoder
+model: anthropic/claude-sonnet-4-5
+
+prompt: |
+  Implement a complete authentication system with:
+  - User model (models/User.js)
+  - Auth controller (controllers/AuthController.js)
+  - Auth middleware (middleware/auth.js)
+  - Auth routes (routes/auth.js)
+  - Auth service (services/AuthService.js)
+  
+  This spans 5 files - please plan and implement.
+
+# Expected behavior
+behavior:
+  # Should delegate to task-manager OR present detailed plan
+  mustContainAnyOf:
+    - ["task-manager", "subagents/core/task-manager"]
+    - ["Implementation Plan", "Step 1", "Step 2"]
+  minToolCalls: 1
+
+# Should follow approval workflow
+expectedViolations:
+  - rule: approval-gate
+    shouldViolate: false
+    severity: error
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 120000
+
+tags:
+  - delegation
+  - task-manager
+  - complex-feature
+  - multi-file

evals/agents/opencoder/tests/error-handling/error-stop-on-failure.yaml

@@ -0,0 +1,56 @@
+id: error-stop-on-failure
+name: Error Handling - Stop on Failure
+description: |
+  Tests that opencoder STOPS on failures and doesn't auto-fix.
+  Agent should:
+  1. Detect error/failure
+  2. STOP execution
+  3. REPORT the error
+  4. PROPOSE fix
+  5. REQUEST approval before fixing
+  
+  Should NEVER auto-fix without approval.
+
+category: edge-case
+agent: opencoder
+model: anthropic/claude-sonnet-4-5
+
+prompts:
+  - text: |
+      Run the command: node nonexistent-file.js
+      
+      When it fails, report the error and propose a fix.
+    expectContext: false
+  
+  - text: "approve"
+    delayMs: 2000
+
+# Expected behavior
+behavior:
+  mustUseTools: [bash]
+  minToolCalls: 1
+  # Should report error and ask for approval
+  mustContain:
+    - "error"
+  # Should NOT auto-fix
+  mustNotContain:
+    - "I've fixed"
+    - "I fixed"
+    - "automatically fixed"
+
+# Should follow error handling rules
+expectedViolations:
+  - rule: approval-gate
+    shouldViolate: false
+    severity: error
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 90000
+
+tags:
+  - error-handling
+  - stop-on-failure
+  - report-first
+  - critical

evals/agents/opencoder/tests/implementation/incremental-implementation.yaml

@@ -0,0 +1,51 @@
+id: incremental-implementation
+name: Incremental Implementation with Validation
+description: |
+  Tests that opencoder implements step-by-step (not all at once).
+  Agent should:
+  1. Present multi-step plan
+  2. Implement ONE step at a time
+  3. Validate after each step
+  4. NOT implement entire plan in one go
+
+category: developer
+agent: opencoder
+model: anthropic/claude-sonnet-4-5
+
+prompts:
+  - text: |
+      Create a simple calculator module with:
+      1. Add function
+      2. Subtract function
+      3. Multiply function
+      
+      Implement incrementally, one function at a time.
+    expectContext: false
+  
+  - text: "approve"
+    delayMs: 2000
+
+# Expected behavior
+behavior:
+  mustUseTools: [read, write]
+  minToolCalls: 2
+  # Should mention incremental/step-by-step approach
+  mustContain:
+    - "Step"
+
+# Should follow approval workflow
+expectedViolations:
+  - rule: approval-gate
+    shouldViolate: false
+    severity: error
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 180000
+
+tags:
+  - implementation
+  - incremental
+  - validation
+  - workflow

evals/agents/opencoder/tests/implementation/multi-language-validation.yaml

@@ -0,0 +1,43 @@
+id: multi-language-validation
+name: Multi-Language Support & Validation
+description: |
+  Tests that opencoder adapts to different languages and uses proper tools.
+  For TypeScript: should mention tsc, eslint
+  For Python: should mention mypy, pylint, pytest
+  
+  Agent should recognize language and use appropriate validation tools.
+
+category: developer
+agent: opencoder
+model: anthropic/claude-sonnet-4-5
+
+prompt: |
+  Create a simple TypeScript function in src/utils/math.ts that adds two numbers.
+  Make sure to use proper type checking and validation.
+
+# Expected behavior
+behavior:
+  mustUseTools: [read, write]
+  minToolCalls: 2
+  # Should mention TypeScript-specific tools
+  mustContainAnyOf:
+    - ["tsc", "TypeScript"]
+    - ["type check", "type-check"]
+    - ["eslint"]
+
+# Should follow approval workflow
+expectedViolations:
+  - rule: approval-gate
+    shouldViolate: false
+    severity: error
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 120000
+
+tags:
+  - multi-language
+  - typescript
+  - validation
+  - type-checking

evals/agents/opencoder/tests/planning/planning-approval-workflow.yaml

@@ -0,0 +1,42 @@
+id: planning-approval-workflow
+name: Planning & Approval Workflow
+description: |
+  Tests that opencoder follows plan-first approach with approval gate.
+  Agent should:
+  1. Analyze the request
+  2. Create implementation plan
+  3. Request approval BEFORE any implementation
+  4. NOT execute without approval
+
+category: developer
+agent: opencoder
+model: anthropic/claude-sonnet-4-5
+
+prompt: |
+  Create a simple utility function in utils.js that adds two numbers.
+
+# Expected behavior
+behavior:
+  # Should present plan first
+  mustContain:
+    - "Implementation Plan"
+    - "Approval needed"
+  # Should NOT execute immediately
+  mustNotUseTools: [write, edit, bash]
+
+# Approval gate should NOT be violated (opencoder asks for approval)
+expectedViolations:
+  - rule: approval-gate
+    shouldViolate: false
+    severity: error
+
+approvalStrategy:
+  type: auto-approve
+
+timeout: 60000
+
+tags:
+  - planning
+  - approval-gate
+  - workflow
+  - critical

evals/framework/package.json

@@ -42,10 +42,9 @@
     "@types/node": "^20.10.0",
     "@typescript-eslint/eslint-plugin": "^6.13.0",
     "@typescript-eslint/parser": "^6.13.0",
-    "ajv-cli": "^5.0.0",
     "eslint": "^8.54.0",
     "tsx": "^4.20.6",
     "typescript": "^5.3.0",
-    "vitest": "^1.0.0"
+    "vitest": "^1.6.1"
   }
 }

evals/framework/scripts/debug/show-test-conversation.sh

@@ -0,0 +1,104 @@
+#!/bin/bash
+
+# Show full conversation from a test session
+# Usage: ./show-test-conversation.sh <session_id>
+
+SESSION_ID=$1
+
+if [ -z "$SESSION_ID" ]; then
+  echo "Usage: $0 <session_id>"
+  echo ""
+  echo "Get session ID from test output (look for 'Session created: ses_...')"
+  exit 1
+fi
+
+SESSION_DIR="$HOME/.local/share/opencode/storage/message/$SESSION_ID"
+PART_DIR="$HOME/.local/share/opencode/storage/part"
+
+if [ ! -d "$SESSION_DIR" ]; then
+  echo "Session not found: $SESSION_ID"
+  exit 1
+fi
+
+echo "========================================================================"
+echo "SESSION: $SESSION_ID"
+echo "========================================================================"
+echo ""
+
+# Process messages in order
+for msg_file in "$SESSION_DIR"/*.json; do
+  if [ -f "$msg_file" ]; then
+    MSG_ID=$(cat "$msg_file" | jq -r '.id')
+    ROLE=$(cat "$msg_file" | jq -r '.role')
+    SUMMARY=$(cat "$msg_file" | jq -r '.summary.body // empty')
+    
+    if [ "$ROLE" = "user" ]; then
+      echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
+      echo "👤 USER PROMPT"
+      echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
+      
+      # Get actual user prompt from parts (not summary which is auto-generated)
+      if [ -d "$PART_DIR/$MSG_ID" ]; then
+        for part_file in "$PART_DIR/$MSG_ID"/*.json; do
+          if [ -f "$part_file" ]; then
+            PART_TYPE=$(cat "$part_file" | jq -r '.type')
+            if [ "$PART_TYPE" = "text" ]; then
+              TEXT=$(cat "$part_file" | jq -r '.text // empty')
+              if [ -n "$TEXT" ]; then
+                echo "$TEXT"
+              fi
+            fi
+          fi
+        done
+      else
+        # Fallback to summary if no parts (shouldn't happen)
+        if [ -n "$SUMMARY" ]; then
+          echo "$SUMMARY"
+        fi
+      fi
+      echo ""
+      
+    elif [ "$ROLE" = "assistant" ]; then
+      echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
+      echo "🤖 ASSISTANT"
+      echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
+      
+      # Read parts from the part directory
+      if [ -d "$PART_DIR/$MSG_ID" ]; then
+        for part_file in "$PART_DIR/$MSG_ID"/*.json; do
+          if [ -f "$part_file" ]; then
+            PART_TYPE=$(cat "$part_file" | jq -r '.type')
+            
+            if [ "$PART_TYPE" = "text" ]; then
+              TEXT=$(cat "$part_file" | jq -r '.text // empty')
+              if [ -n "$TEXT" ]; then
+                echo "$TEXT"
+                echo ""
+              fi
+            elif [ "$PART_TYPE" = "tool" ]; then
+              TOOL=$(cat "$part_file" | jq -r '.tool')
+              INPUT=$(cat "$part_file" | jq -c '.input // {}')
+              echo "🔧 TOOL CALL: $TOOL"
+              echo "   Input: $INPUT"
+              echo ""
+            elif [ "$PART_TYPE" = "tool_result" ]; then
+              RESULT=$(cat "$part_file" | jq -r '.result // empty')
+              if [ -n "$RESULT" ]; then
+                # Show first 500 chars of result
+                echo "📊 TOOL RESULT:"
+                echo "$RESULT" | head -c 500
+                if [ ${#RESULT} -gt 500 ]; then
+                  echo "..."
+                fi
+                echo ""
+              fi
+            fi
+          fi
+        done
+      fi
+      echo ""
+    fi
+  fi
+done
+
+echo "========================================================================"

evals/framework/scripts/run-test-verbose.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+
+# Run a test with full conversation output
+# Usage: ./run-test-verbose.sh <agent> <pattern>
+# Example: ./run-test-verbose.sh opencoder "planning/*.yaml"
+
+AGENT=${1:-opencoder}
+PATTERN=${2:-"**/*.yaml"}
+
+echo "🚀 Running test with verbose output..."
+echo "Agent: $AGENT"
+echo "Pattern: $PATTERN"
+echo ""
+
+# Run test with debug mode and capture session ID
+OUTPUT=$(cd .. && npm run eval:sdk -- --agent=$AGENT --pattern="$PATTERN" --debug 2>&1)
+
+# Extract session ID from output
+SESSION_ID=$(echo "$OUTPUT" | grep -o "Session created: ses_[a-zA-Z0-9]*" | head -1 | cut -d' ' -f3)
+
+# Show test summary
+echo "$OUTPUT" | grep -A 20 "TEST RESULTS"
+
+if [ -n "$SESSION_ID" ]; then
+  echo ""
+  echo "========================================================================"
+  echo "FULL CONVERSATION"
+  echo "========================================================================"
+  echo ""
+  
+  # Show full conversation
+  ./debug/show-test-conversation.sh "$SESSION_ID"
+else
+  echo ""
+  echo "⚠️  No session ID found. Test may have failed to start."
+  echo ""
+  echo "Full output:"
+  echo "$OUTPUT"
+fi

evals/framework/src/sdk/event-logger.ts

@@ -36,7 +36,17 @@ export function logEvent(event: ServerEvent): void {
       // Message updates happen frequently during streaming
       // Only log role changes or completion
       if (props.role === 'user') {
-        console.log(`👤 User message received`);
+        const showFull = process.env.DEBUG_VERBOSE === 'true';
+        
+        if (showFull && props.summary?.body) {
+          console.log(`\n${'═'.repeat(70)}`);
+          console.log(`👤 USER PROMPT:`);
+          console.log(`${'═'.repeat(70)}`);
+          console.log(props.summary.body);
+          console.log(`${'═'.repeat(70)}\n`);
+        } else {
+          console.log(`👤 User message received`);
+        }
       }
       // Skip assistant message updates (too noisy)
       break;
@@ -76,32 +86,57 @@ function logPartEvent(props: any): void {
   if (props.type === 'tool') {
     const toolName = props.tool || 'unknown';
     const status = props.state?.status || props.status || '';
+    const showFull = process.env.DEBUG_VERBOSE === 'true';
     
     // Only log when tool starts or completes
     if (status === 'running' || status === 'pending') {
       console.log(`🔧 Tool: ${toolName} (starting)`);
       
-      // Show tool input preview
+      // Show tool input
       const input = props.state?.input || props.input || {};
-      if (input.command) {
+      
-        const cmd = input.command.substring(0, 70);
+      if (showFull) {
-        console.log(`   └─ ${cmd}${input.command.length > 70 ? '...' : ''}`);
+        console.log(`   Input: ${JSON.stringify(input, null, 2)}`);
-      } else if (input.filePath) {
+      } else {
-        console.log(`   └─ ${input.filePath}`);
+        if (input.command) {
-      } else if (input.pattern) {
+          const cmd = input.command.substring(0, 70);
-        console.log(`   └─ pattern: ${input.pattern}`);
+          console.log(`   └─ ${cmd}${input.command.length > 70 ? '...' : ''}`);
+        } else if (input.filePath) {
+          console.log(`   └─ ${input.filePath}`);
+        } else if (input.pattern) {
+          console.log(`   └─ pattern: ${input.pattern}`);
+        }
       }
     } else if (status === 'completed') {
       console.log(`✅ Tool: ${toolName} (completed)`);
+      
+      if (showFull) {
+        const result = props.state?.result || props.result || '';
+        if (result) {
+          const preview = result.substring(0, 300);
+          console.log(`   Result: ${preview}${result.length > 300 ? '...' : ''}`);
+        }
+      }
     } else if (status === 'error') {
       console.log(`❌ Tool: ${toolName} (error)`);
     }
   } else if (props.type === 'text') {
-    // Text parts - show preview of assistant response
+    // Text parts - show assistant response
     const text = props.text || '';
     if (text.length > 0) {
-      const preview = text.substring(0, 100).replace(/\n/g, ' ');
+      // Check if we should show full text (when DEBUG_VERBOSE is set)
-      console.log(`📝 ${preview}${text.length > 100 ? '...' : ''}`);
+      const showFull = process.env.DEBUG_VERBOSE === 'true';
+      
+      if (showFull) {
+        console.log(`\n${'─'.repeat(70)}`);
+        console.log(`📝 ASSISTANT RESPONSE:`);
+        console.log(`${'─'.repeat(70)}`);
+        console.log(text);
+        console.log(`${'─'.repeat(70)}\n`);
+      } else {
+        const preview = text.substring(0, 100).replace(/\n/g, ' ');
+        console.log(`📝 ${preview}${text.length > 100 ? '...' : ''}`);
+      }
     }
   }
 }

evals/framework/src/sdk/run-sdk-tests.ts

@@ -6,6 +6,7 @@
  * Usage:
  *   npm run eval:sdk
  *   npm run eval:sdk -- --debug
+ *   npm run eval:sdk -- --verbose
  *   npm run eval:sdk -- --no-evaluators
  *   npm run eval:sdk -- --core
  *   npm run eval:sdk -- --agent=opencoder
@@ -14,9 +15,12 @@
  *   npm run eval:sdk -- --model=anthropic/claude-3-5-sonnet-20241022
  *   npm run eval:sdk -- --pattern="developer/*.yaml" --model=openai/gpt-4-turbo
  *   npm run eval:sdk -- --prompt-variant=gpt --agent=openagent
+ *   npm run eval:sdk -- --agent=opencoder --verbose  # Show full conversations
  * 
  * Options:
- *   --debug              Enable debug logging
+ *   --debug              Enable debug logging and keep sessions for inspection
+ *   --verbose            Show full conversation (prompts + responses) after each test
+ *                        (automatically enables --debug)
  *   --no-evaluators      Skip running evaluators (faster)
  *   --core               Run core test suite only (7 tests, ~5-8 min)
  *   --agent=AGENT        Run tests for specific agent (openagent, opencoder)
@@ -43,6 +47,7 @@ const __dirname = dirname(__filename);
 
 interface CliArgs {
   debug: boolean;
+  verbose: boolean;
   noEvaluators: boolean;
   core: boolean;
   suite?: string;
@@ -58,6 +63,7 @@ function parseArgs(): CliArgs {
   
   return {
     debug: args.includes('--debug'),
+    verbose: args.includes('--verbose'),
     noEvaluators: args.includes('--no-evaluators'),
     core: args.includes('--core'),
     suite: args.find(a => a.startsWith('--suite='))?.split('=')[1],
@@ -175,9 +181,105 @@ function printResults(results: TestResult[]): void {
   }
 }
 
+/**
+ * Display full conversation from a session
+ */
+async function displayConversation(sessionId: string): Promise<void> {
+  const { homedir } = await import('os');
+  const { readFileSync, readdirSync, existsSync } = await import('fs');
+  
+  const sessionDir = join(homedir(), '.local', 'share', 'opencode', 'storage', 'message', sessionId);
+  const partDir = join(homedir(), '.local', 'share', 'opencode', 'storage', 'part');
+  
+  if (!existsSync(sessionDir)) {
+    console.log(`⚠️  Session not found: ${sessionId}`);
+    return;
+  }
+  
+  console.log('\n' + '='.repeat(70));
+  console.log('FULL CONVERSATION');
+  console.log('='.repeat(70));
+  console.log(`Session: ${sessionId}\n`);
+  
+  // Read all message files and sort by creation time
+  const messageFiles = readdirSync(sessionDir)
+    .filter(f => f.endsWith('.json'))
+    .map(f => {
+      const content = JSON.parse(readFileSync(join(sessionDir, f), 'utf-8'));
+      return { file: f, content, created: content.time?.created || 0 };
+    })
+    .sort((a, b) => a.created - b.created);
+  
+  for (const { content: msg } of messageFiles) {
+    const role = msg.role;
+    const msgId = msg.id;
+    
+    if (role === 'user') {
+      console.log('━'.repeat(70));
+      console.log('👤 USER PROMPT');
+      console.log('━'.repeat(70));
+      
+      // Get actual user prompt from parts
+      const msgPartDir = join(partDir, msgId);
+      if (existsSync(msgPartDir)) {
+        const partFiles = readdirSync(msgPartDir).filter(f => f.endsWith('.json'));
+        for (const partFile of partFiles) {
+          const part = JSON.parse(readFileSync(join(msgPartDir, partFile), 'utf-8'));
+          if (part.type === 'text' && part.text) {
+            console.log(part.text);
+          }
+        }
+      }
+      console.log();
+      
+    } else if (role === 'assistant') {
+      console.log('━'.repeat(70));
+      console.log('🤖 ASSISTANT');
+      console.log('━'.repeat(70));
+      
+      // Read parts from the part directory
+      const msgPartDir = join(partDir, msgId);
+      if (existsSync(msgPartDir)) {
+        const partFiles = readdirSync(msgPartDir)
+          .filter(f => f.endsWith('.json'))
+          .map(f => {
+            const content = JSON.parse(readFileSync(join(msgPartDir, f), 'utf-8'));
+            return { file: f, content, created: content.time?.start || content.time?.created || 0 };
+          })
+          .sort((a, b) => a.created - b.created);
+        
+        for (const { content: part } of partFiles) {
+          if (part.type === 'text' && part.text) {
+            console.log(part.text);
+            console.log();
+          } else if (part.type === 'tool') {
+            console.log(`🔧 TOOL CALL: ${part.tool}`);
+            console.log(`   Input: ${JSON.stringify(part.input || {})}`);
+            console.log();
+          } else if (part.type === 'tool_result' && part.result) {
+            const result = part.result.substring(0, 300);
+            console.log(`📊 TOOL RESULT:`);
+            console.log(result + (part.result.length > 300 ? '...' : ''));
+            console.log();
+          }
+        }
+      }
+      console.log();
+    }
+  }
+  
+  console.log('='.repeat(70) + '\n');
+}
+
 async function main() {
   const args = parseArgs();
   
+  // If --verbose is set, automatically enable --debug (required for session data)
+  if (args.verbose && !args.debug) {
+    args.debug = true;
+    console.log('ℹ️  --verbose flag automatically enabled --debug (required for session data)\n');
+  }
+  
   console.log('🚀 OpenCode SDK Test Runner\n');
   
   // Determine project root (for prompt management)
@@ -431,6 +533,24 @@ async function main() {
     // Print results
     printResults(results);
     
+    // Show full conversations if --verbose flag is set
+    if (args.verbose) {
+      console.log('\n' + '='.repeat(70));
+      console.log('VERBOSE MODE: Displaying full conversations');
+      console.log('='.repeat(70) + '\n');
+      
+      for (const result of results) {
+        console.log(`\nTest: ${result.testCase.id}`);
+        console.log(`Session ID: ${result.sessionId || 'N/A'}`);
+        
+        if (result.sessionId) {
+          await displayConversation(result.sessionId);
+        } else {
+          console.log('⚠️  No session ID available for this test\n');
+        }
+      }
+    }
+    
     // Exit with appropriate code
     const allPassed = results.every(r => r.passed);
     process.exit(allPassed ? 0 : 1);

evals/results/latest.json

@@ -1,51 +1,31 @@
 {
   "meta": {
-    "timestamp": "2025-11-29T06:55:58.141Z",
+    "timestamp": "2025-12-09T00:07:10.119Z",
-    "agent": "openagent",
+    "agent": "opencoder",
     "model": "opencode/grok-code-fast",
     "framework_version": "0.1.0",
-    "git_commit": "79110ed"
+    "git_commit": "fd1821c"
   },
   "summary": {
-    "total": 2,
+    "total": 1,
-    "passed": 0,
+    "passed": 1,
-    "failed": 2,
+    "failed": 0,
-    "duration_ms": 43350,
+    "duration_ms": 10952,
-    "pass_rate": 0
+    "pass_rate": 1
   },
   "by_category": {
     "developer": {
-      "passed": 0,
+      "passed": 1,
-      "total": 2
+      "total": 1
     }
   },
   "tests": [
     {
-      "id": "execution-balance-positive-001",
+      "id": "planning-approval-workflow",
       "category": "developer",
-      "passed": false,
+      "passed": true,
-      "duration_ms": 22285,
+      "duration_ms": 10952,
-      "events": 40,
+      "events": 13,
-      "approvals": 0,
-      "violations": {
-        "total": 1,
-        "errors": 1,
-        "warnings": 0,
-        "details": [
-          {
-            "type": "missing-required-tool",
-            "severity": "error",
-            "message": "Required tool 'write' was not used"
-          }
-        ]
-      }
-    },
-    {
-      "id": "execution-balance-negative-001",
-      "category": "developer",
-      "passed": false,
-      "duration_ms": 21065,
-      "events": 78,
       "approvals": 0,
       "violations": {
         "total": 0,

package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-agents",
-  "version": "0.0.2",
+  "version": "0.3.0",
   "description": "OpenCode agent evaluation framework and test suites",
   "private": true,
   "workspaces": [

scripts/development/demo.sh

@@ -154,16 +154,21 @@ show_prompt_library() {
   cat << EOF
 The prompt library allows different AI models to use optimized prompts:
 
+${GREEN}.opencode/agent/${NC}
+└── openagent.md        ${YELLOW}# Canonical default (Claude-optimized)${NC}
+
 ${GREEN}.opencode/prompts/openagent/${NC}
-├── default.md          ${YELLOW}# Stable version (used in PRs)${NC}
+├── gpt.md              ${YELLOW}# GPT-4 optimized${NC}
-├── sonnet-4.md         ${YELLOW}# Claude Sonnet 4 optimized${NC}
+├── gemini.md           ${YELLOW}# Gemini optimized${NC}
-├── grok-fast.md        ${YELLOW}# Grok Fast optimized${NC}
+├── grok.md             ${YELLOW}# Grok optimized${NC}
+├── llama.md            ${YELLOW}# Llama/OSS optimized${NC}
 ├── TEMPLATE.md         ${YELLOW}# Template for new variants${NC}
-└── results/            ${YELLOW}# Test results${NC}
+└── results/            ${YELLOW}# Test results (all variants)${NC}
 
 ${BLUE}Key Principles:${NC}
-  • PRs must use default prompts (enforced by CI)
+  • Agent files are canonical defaults (source of truth)
-  • Variants are tested and documented
+  • Variants are model-specific optimizations
+  • All variants tested and documented
   • Users can choose the best prompt for their model
   • Contributors can add optimized variants
 

scripts/prompts/test-prompt.sh

@@ -83,36 +83,41 @@ usage() {
     echo ""
     echo "Required:"
     echo "  --agent=NAME       Agent name (e.g., openagent, opencoder)"
-    echo "  --variant=NAME     Prompt variant by model family (e.g., default, gpt, gemini, grok, llama)"
+    echo "  --variant=NAME     Prompt variant (default, gpt, gemini, grok, llama)"
     echo ""
     echo "Optional:"
     echo "  --model=MODEL      Model to test with (uses prompt metadata if not specified)"
     echo "  --help, -h         Show this help"
     echo ""
+    echo -e "${BLUE}Architecture:${NC}"
+    echo "  • default = Canonical agent file (.opencode/agent/<agent>.md)"
+    echo "  • Other variants = Model-specific optimizations (.opencode/prompts/<agent>/<model>.md)"
+    echo "  • Results always saved to .opencode/prompts/<agent>/results/"
+    echo ""
     echo "Examples:"
-    echo "  # Test with default (Claude) - uses metadata recommendation"
+    echo "  # Test default (canonical agent file)"
     echo "  $0 --agent=openagent --variant=default"
     echo ""
-    echo "  # Test GPT prompt - uses metadata recommendation (gpt-4o)"
+    echo "  # Test GPT-optimized prompt"
     echo "  $0 --agent=openagent --variant=gpt"
     echo ""
     echo "  # Test Gemini prompt with specific model"
     echo "  $0 --agent=openagent --variant=gemini --model=google/gemini-2.0-flash-exp"
     echo ""
-    echo "  # Test Grok prompt - uses metadata recommendation (free tier)"
+    echo "  # Test Grok prompt"
     echo "  $0 --agent=openagent --variant=grok"
     echo ""
     echo "Available model families:"
-    echo "  default    # Claude Sonnet 4.5 (stable)"
+    echo "  default    # Canonical agent file (Claude Sonnet 4.5)"
     echo "  gpt        # OpenAI GPT-4o, GPT-4o-mini, o1"
     echo "  gemini     # Google Gemini 2.0 Flash, Pro"
     echo "  grok       # xAI Grok (free tier available)"
     echo "  llama      # Meta Llama 3.1/3.2 (local or hosted)"
     echo ""
-    echo "Note: Each prompt contains metadata with recommended models."
+    echo "Note: Model-specific prompts contain metadata with recommended models."
     echo "      If --model is not specified, the primary recommendation is used."
     echo ""
-    echo "Available prompts for an agent:"
+    echo "Available variants for an agent:"
     echo "  ls $PROMPTS_DIR/<agent-name>/"
     exit 1
 }
@@ -150,19 +155,29 @@ if [[ -z "$AGENT_NAME" ]] || [[ -z "$PROMPT_VARIANT" ]]; then
     usage
 fi
 
-PROMPT_FILE="$PROMPTS_DIR/$AGENT_NAME/$PROMPT_VARIANT.md"
 AGENT_FILE="$AGENT_DIR/$AGENT_NAME.md"
 BACKUP_FILE="$AGENT_DIR/.$AGENT_NAME.md.backup"
 VARIANT_RESULTS_DIR="$PROMPTS_DIR/$AGENT_NAME/results"
 VARIANT_RESULTS_FILE="$VARIANT_RESULTS_DIR/$PROMPT_VARIANT-results.json"
 
-# Check prompt exists
+# Handle "default" variant - use agent file directly
-if [[ ! -f "$PROMPT_FILE" ]]; then
+if [[ "$PROMPT_VARIANT" == "default" ]]; then
-    echo -e "${RED}Error: Prompt not found: $PROMPT_FILE${NC}"
+    PROMPT_FILE="$AGENT_FILE"
-    echo ""
+    echo -e "${BLUE}Testing default prompt (canonical agent file)${NC}"
-    echo "Available prompts for $AGENT_NAME:"
+else
-    ls -1 "$PROMPTS_DIR/$AGENT_NAME/"*.md 2>/dev/null | xargs -I {} basename {} .md || echo "  (none found)"
+    PROMPT_FILE="$PROMPTS_DIR/$AGENT_NAME/$PROMPT_VARIANT.md"
-    exit 1
+    
+    # Check prompt exists
+    if [[ ! -f "$PROMPT_FILE" ]]; then
+        echo -e "${RED}Error: Prompt variant not found: $PROMPT_FILE${NC}"
+        echo ""
+        echo "Available variants for $AGENT_NAME:"
+        echo "  - default (canonical agent file)"
+        if [[ -d "$PROMPTS_DIR/$AGENT_NAME" ]]; then
+            ls -1 "$PROMPTS_DIR/$AGENT_NAME/"*.md 2>/dev/null | grep -v "TEMPLATE.md" | grep -v "README.md" | xargs -I {} basename {} .md || echo "  (no model variants found)"
+        fi
+        exit 1
+    fi
 fi
 
 # Read metadata from prompt file
@@ -215,11 +230,16 @@ else
     echo "      No existing agent file to backup"
 fi
 
-# Step 2: Copy prompt variant to agent location
+# Step 2: Copy prompt variant to agent location (skip if testing default)
-echo -e "${YELLOW}[2/5] Copying prompt variant to agent location...${NC}"
+if [[ "$PROMPT_VARIANT" == "default" ]]; then
-cp "$PROMPT_FILE" "$AGENT_FILE"
+    echo -e "${YELLOW}[2/5] Using default prompt (already in place)...${NC}"
-echo "      Copied $PROMPT_FILE"
+    echo "      Testing: $AGENT_FILE"
-echo "      To     $AGENT_FILE"
+else
+    echo -e "${YELLOW}[2/5] Copying prompt variant to agent location...${NC}"
+    cp "$PROMPT_FILE" "$AGENT_FILE"
+    echo "      Copied $PROMPT_FILE"
+    echo "      To     $AGENT_FILE"
+fi
 
 # Step 3: Run tests
 echo -e "${YELLOW}[3/5] Running core eval tests...${NC}"
@@ -249,18 +269,18 @@ set -e
 echo ""
 echo -e "${BLUE}═══════════════════════════════════════════════════════════════${NC}"
 
-# Step 4: Restore default prompt
+# Step 4: Restore original prompt (if we changed it)
 echo ""
-echo -e "${YELLOW}[4/5] Restoring default prompt...${NC}"
+if [[ "$PROMPT_VARIANT" == "default" ]]; then
-DEFAULT_PROMPT="$PROMPTS_DIR/$AGENT_NAME/default.md"
+    echo -e "${YELLOW}[4/5] No restore needed (tested default)...${NC}"
-if [[ -f "$DEFAULT_PROMPT" ]]; then
+    echo "      Agent file unchanged"
-    cp "$DEFAULT_PROMPT" "$AGENT_FILE"
-    echo "      Restored default.md to agent location"
 else
-    # Restore backup if no default
+    echo -e "${YELLOW}[4/5] Restoring original prompt...${NC}"
     if [[ -f "$BACKUP_FILE" ]]; then
         cp "$BACKUP_FILE" "$AGENT_FILE"
         echo "      Restored from backup"
+    else
+        echo "      No backup to restore"
     fi
 fi
 

scripts/prompts/use-prompt.sh

@@ -3,12 +3,16 @@
 # use-prompt.sh - Switch an agent to use a specific prompt variant
 #
 # Usage:
+#   ./scripts/prompts/use-prompt.sh --agent=openagent --variant=gpt
 #   ./scripts/prompts/use-prompt.sh --agent=openagent --variant=default
-#   ./scripts/prompts/use-prompt.sh --agent=openagent --variant=sonnet-4
 #
 # What it does:
-#   1. Copies the specified prompt to the agent location
+#   1. If variant is "default", restores the canonical agent file (no-op, already default)
-#   2. That's it - the agent now uses that prompt
+#   2. Otherwise, copies the model-specific variant to the agent location
+#
+# New Architecture:
+#   - Agent files (.opencode/agent/*.md) are the canonical defaults
+#   - Prompt variants (.opencode/prompts/<agent>/<model>.md) are model-specific optimizations
 #
 
 set -e
@@ -20,6 +24,7 @@ ROOT_DIR="$(cd "$SCRIPT_DIR/../.." && pwd)"
 RED='\033[0;31m'
 GREEN='\033[0;32m'
 YELLOW='\033[1;33m'
+BLUE='\033[0;34m'
 NC='\033[0m'
 
 # Default values
@@ -74,13 +79,17 @@ usage() {
     echo ""
     echo "Required:"
     echo "  --agent=NAME       Agent name (e.g., openagent, opencoder)"
-    echo "  --variant=NAME     Prompt variant by model family (e.g., default, gpt, gemini, grok, llama)"
+    echo "  --variant=NAME     Prompt variant (default, gpt, gemini, grok, llama)"
     echo ""
     echo "Optional:"
     echo "  --help, -h         Show this help"
     echo ""
+    echo -e "${BLUE}Architecture:${NC}"
+    echo "  • Agent files (.opencode/agent/*.md) = Canonical defaults"
+    echo "  • Prompt variants (.opencode/prompts/<agent>/<model>.md) = Model-specific"
+    echo ""
     echo "Examples:"
-    echo "  # Use default (Claude Sonnet 4.5)"
+    echo "  # Use default (canonical agent file)"
     echo "  $0 --agent=openagent --variant=default"
     echo ""
     echo "  # Switch to GPT-optimized prompt"
@@ -89,18 +98,21 @@ usage() {
     echo "  # Switch to Gemini-optimized prompt"
     echo "  $0 --agent=openagent --variant=gemini"
     echo ""
-    echo "  # Switch to Grok-optimized prompt"
-    echo "  $0 --agent=openagent --variant=grok"
-    echo ""
     echo "Available prompts:"
     for agent_dir in "$PROMPTS_DIR"/*/; do
+        [ -d "$agent_dir" ] || continue
         agent=$(basename "$agent_dir")
         echo "  $agent:"
+        echo "    - default (canonical agent file)"
         for prompt in "$agent_dir"*.md; do
             if [[ -f "$prompt" ]] && [[ "$(basename "$prompt")" != "TEMPLATE.md" ]] && [[ "$(basename "$prompt")" != "README.md" ]]; then
                 variant=$(basename "$prompt" .md)
                 model_family=$(extract_metadata "$prompt" "model_family")
-                echo "    - $variant ($model_family family)"
+                if [[ -n "$model_family" ]]; then
+                    echo "    - $variant ($model_family family)"
+                else
+                    echo "    - $variant"
+                fi
             fi
         done
     done
@@ -136,15 +148,43 @@ if [[ -z "$AGENT_NAME" ]] || [[ -z "$PROMPT_VARIANT" ]]; then
     usage
 fi
 
-PROMPT_FILE="$PROMPTS_DIR/$AGENT_NAME/$PROMPT_VARIANT.md"
 AGENT_FILE="$AGENT_DIR/$AGENT_NAME.md"
 
+# Check agent exists
+if [[ ! -f "$AGENT_FILE" ]]; then
+    echo -e "${RED}Error: Agent not found: $AGENT_FILE${NC}"
+    echo ""
+    echo "Available agents:"
+    ls -1 "$AGENT_DIR/"*.md 2>/dev/null | xargs -I {} basename {} .md || echo "  (none found)"
+    exit 1
+fi
+
+# Handle "default" variant - agent file is already the default
+if [[ "$PROMPT_VARIANT" == "default" ]]; then
+    echo -e "${GREEN}✓${NC} Agent ${GREEN}$AGENT_NAME${NC} is using the ${GREEN}default${NC} prompt"
+    echo ""
+    echo "  Location: $AGENT_FILE"
+    echo ""
+    echo -e "${BLUE}Note:${NC} In the new architecture, agent files are the canonical defaults."
+    echo "       No action needed - you're already using the default!"
+    echo ""
+    exit 0
+fi
+
+# For non-default variants, copy from prompts directory
+PROMPT_FILE="$PROMPTS_DIR/$AGENT_NAME/$PROMPT_VARIANT.md"
+
 # Check prompt exists
 if [[ ! -f "$PROMPT_FILE" ]]; then
-    echo -e "${RED}Error: Prompt not found: $PROMPT_FILE${NC}"
+    echo -e "${RED}Error: Prompt variant not found: $PROMPT_FILE${NC}"
     echo ""
-    echo "Available prompts for $AGENT_NAME:"
+    echo "Available variants for $AGENT_NAME:"
-    ls -1 "$PROMPTS_DIR/$AGENT_NAME/"*.md 2>/dev/null | xargs -I {} basename {} .md || echo "  (none found)"
+    echo "  - default (canonical agent file)"
+    if [[ -d "$PROMPTS_DIR/$AGENT_NAME" ]]; then
+        ls -1 "$PROMPTS_DIR/$AGENT_NAME/"*.md 2>/dev/null | grep -v "TEMPLATE.md" | grep -v "README.md" | xargs -I {} basename {} .md || echo "  (no model variants found)"
+    else
+        echo "  (no model variants found)"
+    fi
     exit 1
 fi
 
@@ -178,5 +218,8 @@ if [[ -n "$MODEL_FAMILY" ]]; then
     echo ""
 fi
 
+echo "To restore the default:"
+echo "  ./scripts/prompts/use-prompt.sh --agent=$AGENT_NAME --variant=default"
+echo ""
 echo "To test this prompt:"
 echo "  ./scripts/prompts/test-prompt.sh --agent=$AGENT_NAME --variant=$PROMPT_VARIANT"

scripts/prompts/validate-pr.sh

@@ -1,6 +1,7 @@
 #!/bin/bash
-# Validate that all agents use their default prompts
+# Validate prompt library structure
-# This ensures PRs maintain stable defaults in the main branch
+# Agent files in .opencode/agent/ are the canonical defaults
+# Prompts in .opencode/prompts/ are model-specific variants
 
 set -e
 
@@ -13,9 +14,10 @@ WARNINGS=0
 RED='\033[0;31m'
 GREEN='\033[0;32m'
 YELLOW='\033[1;33m'
+BLUE='\033[0;34m'
 NC='\033[0m' # No Color
 
-echo "🔍 Validating agent prompts..."
+echo "🔍 Validating prompt library structure..."
 echo ""
 
 # Check if prompts directory exists
@@ -26,36 +28,57 @@ if [ ! -d "$PROMPTS_DIR" ]; then
   exit 0
 fi
 
+# Validate structure: agent files are canonical, prompts are variants
+echo -e "${BLUE}Architecture:${NC}"
+echo "  • Agent files (.opencode/agent/*.md) = Canonical defaults"
+echo "  • Prompt variants (.opencode/prompts/<agent>/<model>.md) = Model-specific optimizations"
+echo ""
+
 # Find all agent markdown files (excluding subagents)
 for agent_file in "$AGENT_DIR"/*.md; do
   # Skip if no files found
   [ -e "$agent_file" ] || continue
   
   agent_name=$(basename "$agent_file" .md)
-  default_file="$PROMPTS_DIR/$agent_name/default.md"
+  prompts_subdir="$PROMPTS_DIR/$agent_name"
   
-  # Check if default exists
+  # Check if prompts directory exists for this agent
-  if [ ! -f "$default_file" ]; then
+  if [ ! -d "$prompts_subdir" ]; then
-    echo -e "${YELLOW}⚠️  No default found for $agent_name${NC}"
+    echo -e "${YELLOW}⚠️  No prompt variants for $agent_name${NC}"
-    echo "   Expected: $default_file"
+    echo "   Agent file: $agent_file (canonical default)"
-    echo "   This agent will be skipped."
+    echo "   Variants directory: $prompts_subdir (not found)"
+    echo "   This is OK - variants are optional."
     echo ""
     WARNINGS=$((WARNINGS + 1))
     continue
   fi
   
-  # Compare files
+  # Check for default.md (should NOT exist in new architecture)
-  if ! diff -q "$agent_file" "$default_file" > /dev/null 2>&1; then
+  if [ -f "$prompts_subdir/default.md" ]; then
-    echo -e "${RED}❌ $agent_name.md does not match default${NC}"
+    echo -e "${RED}❌ Found default.md for $agent_name${NC}"
-    echo "   Current:  $agent_file"
+    echo "   Location: $prompts_subdir/default.md"
-    echo "   Expected: $default_file"
+    echo "   This file should not exist in the new architecture."
     echo ""
-    echo "   To fix this:"
+    echo "   To fix:"
-    echo "   ./scripts/prompts/use-prompt.sh $agent_name default"
+    echo "   rm $prompts_subdir/default.md"
+    echo ""
+    echo "   The agent file is now the canonical default:"
+    echo "   $agent_file"
     echo ""
     FAILED=$((FAILED + 1))
   else
-    echo -e "${GREEN}✅ $agent_name.md matches default${NC}"
+    # Count variants
+    variant_count=$(find "$prompts_subdir" -maxdepth 1 -name "*.md" -not -name "README.md" -not -name "TEMPLATE.md" | wc -l | tr -d ' ')
+    
+    if [ "$variant_count" -gt 0 ]; then
+      echo -e "${GREEN}✅ $agent_name${NC}"
+      echo "   Default: $agent_file"
+      echo "   Variants: $variant_count model-specific optimization(s)"
+    else
+      echo -e "${GREEN}✅ $agent_name${NC}"
+      echo "   Default: $agent_file"
+      echo "   Variants: none (using default for all models)"
+    fi
   fi
 done
 
@@ -63,25 +86,21 @@ echo ""
 
 # Summary
 if [ $FAILED -eq 0 ] && [ $WARNINGS -eq 0 ]; then
-  echo -e "${GREEN}✅ All agents use default prompts${NC}"
+  echo -e "${GREEN}✅ Prompt library structure is valid${NC}"
   exit 0
 elif [ $FAILED -eq 0 ]; then
   echo -e "${YELLOW}⚠️  Validation passed with $WARNINGS warning(s)${NC}"
-  echo "   Some agents don't have defaults yet - this is expected during development."
+  echo "   Some agents don't have variant directories yet - this is expected."
   exit 0
 else
-  echo -e "${RED}❌ PR validation failed - $FAILED agent(s) do not use default prompts${NC}"
+  echo -e "${RED}❌ PR validation failed - $FAILED issue(s) found${NC}"
   echo ""
-  echo "PRs must use default prompts to keep the main branch stable."
+  echo "The new prompt architecture:"
+  echo "  • Agent files (.opencode/agent/*.md) are the canonical defaults"
+  echo "  • Prompt variants (.opencode/prompts/<agent>/<model>.md) are model-specific"
+  echo "  • default.md files should NOT exist"
   echo ""
-  echo "To fix:"
+  echo "See docs/contributing/CONTRIBUTING.md for details"
-  echo "  1. Restore defaults: ./scripts/prompts/use-prompt.sh <agent> default"
-  echo "  2. Or run: ./scripts/prompts/validate-pr.sh"
-  echo ""
-  echo "If you're adding a new prompt variant:"
-  echo "  - Add it to .opencode/prompts/<agent>/<variant>.md"
-  echo "  - Do NOT modify .opencode/agent/<agent>.md"
-  echo "  - See docs/contributing/CONTRIBUTING.md for details"
   echo ""
   exit 1
 fi

src/calculator.js

@@ -0,0 +1,34 @@
+/**
+ * Calculator Module
+ * Pure functions for basic arithmetic operations
+ */
+
+/**
+ * Adds two numbers
+ * @param {number} a - First number
+ * @param {number} b - Second number
+ * @returns {number} Sum of a and b
+ */
+const add = (a, b) => a + b;
+
+/**
+ * Subtracts second number from first number
+ * @param {number} a - Number to subtract from
+ * @param {number} b - Number to subtract
+ * @returns {number} Difference of a and b
+ */
+const subtract = (a, b) => a - b;
+
+/**
+ * Multiplies two numbers
+ * @param {number} a - First number
+ * @param {number} b - Second number
+ * @returns {number} Product of a and b
+ */
+const multiply = (a, b) => a * b;
+
+module.exports = {
+  add,
+  subtract,
+  multiply,
+};

src/models/User.js

@@ -0,0 +1,39 @@
+/**
+ * User model with name and email properties
+ */
+class User {
+  /**
+   * Creates a new User instance
+   * @param {string} name - User's name
+   * @param {string} email - User's email address
+   */
+  constructor(name, email) {
+    this.name = name;
+    this.email = email;
+  }
+
+  /**
+   * Returns a plain object representation of the user
+   * @returns {Object} User data as plain object
+   */
+  toObject() {
+    return {
+      name: this.name,
+      email: this.email
+    };
+  }
+
+  /**
+   * Creates a new User with updated properties (immutable update)
+   * @param {Object} updates - Properties to update
+   * @returns {User} New User instance with updates applied
+   */
+  update(updates) {
+    return new User(
+      updates.name ?? this.name,
+      updates.email ?? this.email
+    );
+  }
+}
+
+module.exports = User;