Enterprise-grade, multi-agent AI pipeline for automated code discovery, implementation, testing, and verification across multiple repositories.
This repo contains the .opencode/ pipeline infrastructure — skills, tools, agents, modes, and knowledge that power a multi-agent AI coding pipeline. Drop this into any workspace to get intelligent code automation.
User Input (vague or structured)
│
▼
┌──────────────────────────────┐
│ 1. Requirement Intake │ Transforms vague → structured
└──────────┬───────────────────┘
▼
┌──────────────────────────────┐
│ 2. Pattern Detection │ Classifies → selects pipeline mode
└──────────┬───────────────────┘
▼
┌──────────────────────────────┐
│ 3. Orchestrator Routing │ Backend / Frontend / Both
└──────────┬───────────────────┘
▼
┌──────────────────────────────────────────────────────────────────┐
│ PIPELINE EXECUTION │
│ │
│ full-feature: Discovery → Implementation → Testing → Verification│
│ quick-fix: Implementation → Verification │
│ hotfix: Discovery(lite) → Implementation → Verification │
│ refactor: Discovery → Implementation → Verification │
└──────────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────┐
│ Final Report + Lessons │ Confidence scores + timing + PR
└──────────────────────────────┘
.opencode/
├── agents/ # Agent definitions (6)
│ ├── main-orchestrator.md # Routes, gates, assembles
│ ├── discovery-agent.md # Read-only code analysis
│ ├── implementation-agent.md # Code modifications
│ ├── test-agent.md # Test creation
│ ├── verification-agent.md # Final review & PR prep
│ └── reviewer.md # Code review
├── skills/ # Reusable skill modules (11)
│ ├── requirement-intake/ # Vague → structured requirements
│ ├── pattern-detector/ # Requirement classification
│ ├── context-compressor/ # Inter-agent token optimization
│ ├── pipeline-summary/ # Progress visualization
│ ├── contract-validator/ # Implementation compliance
│ ├── pr-description/ # PR text generation
│ ├── test-scaffold/ # Test boilerplate (~40% speedup)
│ ├── rollback/ # Safe change reversion
│ ├── lessons-learned/ # Outcome recording
│ ├── ado-query/ # Azure DevOps queries
│ └── jenkins-build-status-report/ # CI/CD status
├── tools/ # Shell tools (3)
│ ├── compile-check.sh # Verify compilation
│ ├── test-runner.sh # Run specific tests
│ └── diff-summarizer.sh # Structured diff output
├── modes/ # Pipeline mode definitions (5)
│ ├── full-feature.md # Complete 4-stage pipeline
│ ├── quick-fix.md # Skip discovery + testing
│ ├── hotfix.md # Lite discovery, skip new tests
│ ├── refactor.md # Skip testing, enhanced ref-check
│ └── analyze.md # Read-only analysis mode
├── scripts/
│ └── notify-teams.sh # Teams webhook notifications
├── knowledge/ # Institutional memory
│ ├── pipeline-history.json # All pipeline runs
│ ├── patterns/ # Discovered code patterns
│ ├── pitfalls/ # Common mistakes
│ └── metrics/ # Performance data
├── mcp-servers/
│ └── code-rag/ # Semantic code search (ChromaDB)
├── repos.yaml # Target repository config
├── agent-report-templates.md # Report format standards
└── requirement-template.md # Requirement document template
| Mode | Stages | When to Use |
|---|---|---|
full-feature |
Discovery → Implementation → Testing → Verification | New features, complex changes |
quick-fix |
Implementation → Verification | Typos, labels, config values |
hotfix |
Discovery(lite) → Implementation → Verification | Critical production bugs |
refactor |
Discovery → Implementation → Verification | Structural changes, no behavior change |
Modes automatically escalate when complexity exceeds expectations.
| Agent | Role | Can Write? |
|---|---|---|
| Orchestrator | Routes, gates, assembles reports | Reports only |
| Discovery | Analyzes code, produces implementation contracts | ❌ |
| Implementation | Makes code changes per contract | ✅ |
| Test | Creates/updates tests | ✅ |
| Verification | Reviews changes, runs checks | ❌ |
| Reviewer | Code review feedback | ❌ |
| Skill | Purpose |
|---|---|
requirement-intake |
Transform vague input → structured requirement |
pattern-detector |
Classify requirement → select pipeline mode |
context-compressor |
Compress reports → minimal inter-agent payloads (50-70% token savings) |
pipeline-summary |
Generate progress visualization at gates |
contract-validator |
Validate implementation matches discovery contract |
pr-description |
Generate polished PR descriptions |
test-scaffold |
Pre-generate test boilerplate |
rollback |
Safely revert pipeline changes |
lessons-learned |
Record outcomes for future improvement |
ado-query |
Natural language Azure DevOps queries |
jenkins-build-status |
CI/CD build status and failure analysis |
| Tool | Usage |
|---|---|
compile-check.sh |
.opencode/tools/compile-check.sh <path> <stack> |
test-runner.sh |
.opencode/tools/test-runner.sh <path> <stack> <target> |
diff-summarizer.sh |
.opencode/tools/diff-summarizer.sh <path> [base_ref] |
Locally-running semantic code search using ChromaDB vector embeddings.
| Tool | Description |
|---|---|
search_code |
Natural language code search |
find_related |
Find semantically related files |
get_context |
Expand context around results |
index_stats |
View indexing statistics |
100% local — code never leaves your machine. See .opencode/mcp-servers/code-rag/README.md for setup.
git clone https://github.com/kkuppula/my-opencode-projects.git
cp -r my-opencode-projects/.opencode /path/to/your/workspace/Edit .opencode/repos.yaml:
repos:
backend:
name: MyBackend
path: /path/to/your/backend
stack: Java/Gradle
patterns:
- "API", "service", "controller", "repository"
frontend:
name: MyFrontend
path: /path/to/your/frontend
stack: TypeScript/Angular
patterns:
- "UI", "component", "page", "view"cd .opencode/mcp-servers/code-rag
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
# Edit config.yaml with your repo paths
./run-indexer.shexport TEAMS_WEBHOOK_URL="https://your-org.webhook.office.com/..."- Intelligent mode selection — auto-picks the right pipeline depth
- Multi-repo orchestration — coordinates backend + frontend changes
- Context compression — 50-70% token reduction between stages (see below)
- Confidence scoring — every agent reports 0-100% with rationale
- Human-in-the-loop gates — approval checkpoints between stages
- Institutional learning — records patterns and pitfalls for future runs
- Contract validation — ensures implementation matches discovery
- Safe rollback — reverts changes cleanly on failure
- Teams notifications — posts progress to Microsoft Teams
A key optimization that sits between pipeline stages to reduce token waste by 50-70%.
Without compression, each agent receives the full output of all prior stages — most of which is irrelevant prose, rationale, and verbose code snippets. Tokens compound through the pipeline:
Without compression: With compression:
Discovery Agent: ~15K tokens Discovery Agent: ~6K tokens
Implementation: ~25K tokens Implementation: ~8K tokens
Test Agent: ~30K tokens Test Agent: ~6K tokens
Verification: ~35K tokens Verification: ~6K tokens
──────────────────────────────── ──────────────────────────────
Total: ~100K+ tokens Total: ~26K tokens
Savings: ~70%
Stage N Report (full, human-readable)
│
├── Saved to reports/ (for human gates & debugging)
│
└── Compressed via rule-based extraction
│
▼
Structured JSON Contract (1-3K tokens)
│
└── Passed to Stage N+1 Agent
| Transition | What's Kept | What's Dropped | Savings |
|---|---|---|---|
| Discovery → Implementation | Must Do list, file paths, API contract, patterns | Rationale, alternatives, journey narrative | ~60-75% |
| Implementation → Test | Files changed, behavior added, edge cases | Decisions, contract validation, discovery | ~65-80% |
| Implementation → Verification | Contract checklist, diff ref, commands | Full reports, prose, code-rag results | ~70-80% |
| Test → Verification | Test results, coverage gaps | Test code, mock setup, framework config | ~75% |
❌ Without compression (8K tokens):
"The discovery agent investigated the codebase using code-rag semantic
search. We queried 'user roster course membership' which returned 5
results from the Learn repository. After analyzing MembershipTOPubV1.java
we found that the lastAccessed field has been available since version
3300.9.0 and is stored in the course_users.last_access_date column..."
✅ With compression (2K tokens):
{
"must_do": [
{"action": "Add column header", "file": "course-roster.html", "lines": "289-307"}
],
"api_contract": {"field": "lastAccessDate", "type": "Date|null"},
"files_to_modify": ["course-roster.html"],
"patterns": [{"ref": "course-grades-student.html:209", "note": "bb-date usage"}],
"guard_rails": ["hide-for-small", "No API changes", "No model changes"]
}
| Agent | Max Context | Breakdown |
|---|---|---|
| Discovery | 6K tokens | Requirement (1K) + Instructions (3K) + Repo context (2K) |
| Implementation | 8K tokens | Compressed contract (2K) + Instructions (3K) + Code refs (3K) |
| Test | 6K tokens | Compressed impl (1.5K) + Instructions (3K) + Patterns (1.5K) |
| Verification | 6K tokens | Compressed state (2K) + Instructions (3K) + Diff (1K) |
- No LLM call needed — compression is rule-based extraction, not summarization
- Full reports always preserved — humans see everything at gates; only agents get compressed versions
- Fallback expansion — if an agent reports confidence < 60%, the orchestrator expands the missing section
- Never drop actionable items — Must Do, file paths, and API contracts always pass through
MIT License — Copyright (c) 2026 Kiran Kuppula
See LICENSE for details.