An automated issue-fixing system that watches Jira tickets labeled autofix and dispatches AI agents to fix bugs, review code, and manage the full lifecycle from ticket to merged PR.

Runtime: OpenCode (agent runtime) + OpenShell (sandbox isolation). Model: Claude Sonnet 4.6 (default). Also supports open models via Ollama/LiteMaaS. Status: E2E verified locally, in OpenShell sandbox, and on OpenShift 4.21 cluster.

flowchart LR
    A["🎫 Jira<br>autofix"] --> B["👁 Watcher"]

    subgraph INV["INVESTIGATE (Phases 0-4)"]
        C["📂 Clone +<br>Root Cause"] --> D["📝 Plan +<br>3-Agent Audit"]
    end

    B --> C
    D --> E

    E["🧑 GATE 1<br>Human Plan<br>Review"]:::gate

    subgraph IMPL["IMPLEMENT (Phases 5-11)"]
        F["⚙️ Code Fix +<br>Tests +<br>Blocklist"] --> G["📤 Create PR +<br>Jira Telemetry"]
    end

    E -->|approved| F

    subgraph REV["REVIEW"]
        H["🔎 3-Lens<br>Correctness<br>Security<br>Quality"]
        H -->|findings| I["🔧 Review Fix"]
        I -->|"< 3 cycles"| H
        H -->|clean| J["✅ Done"]
    end

    G --> H

    J --> K["🧑 GATE 2<br>Human PR<br>Review"]:::gate
    K -->|approved| L["🚀 Merged"]:::merged

    I -->|"3 cycles"| M["⚠️ Escalate"]:::fail
    M -.->|"bot-retry<br>(max 2x)"| B

    classDef gate fill:#fff3e0,stroke:#f57c00,stroke-width:2px
    classDef fail fill:#ffebee,stroke:#c62828,stroke-width:2px
    classDef merged fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px
    style INV fill:#e3f2fd,stroke:#1565c0
    style IMPL fill:#e8f5e9,stroke:#2e7d32
    style REV fill:#f3e5f5,stroke:#7b1fa2

1. A user creates a Jira ticket with the autofix label and includes the repository URL
2. The Watcher polls Jira, picks up the ticket, dispatches the Investigation Agent
3. The Investigation Agent clones the repo, investigates, writes a fix plan, runs 3 audit sub-agents
4. A human reviews and approves the plan
5. The Implementation Agent implements the fix, runs tests, creates a PR
6. The Review Agent reviews the PR (correctness, security, quality)
7. The Review-Fix Agent addresses findings (max 3 cycles)
8. A human approves and merges the PR

[Describe the bug — what's broken, steps to reproduce, expected behavior]

## Agent Configuration
**Repository**: https://github.com/org/repo          (REQUIRED)
**Branch**: main                                      (optional)
**Commit**: abc1234def                                (optional)
**Skills**:                                           (optional)
  - https://raw.githubusercontent.com/org/repo/main/.claude/skills/conventions.md
**Knowledge Repo**: https://github.com/org/team-docs  (optional)

Full config reference: docs/Architecture.md → config.env section.

Note: Open models (30-35B) can often identify root causes correctly but struggle with the multi-phase implementation pipeline. The bottleneck is instruction following and tool-call reliability, not reasoning capability.

For full setup instructions, see the Model Configuration Guide.

.opencode/
├── agents/           # Agent definitions (fix-investigate, fix-implement, review, review-fix, 3 audit)
├── skills/           # Skill files (issue-investigate, issue-implement, issue-review, review-fix)
├── plugins/          # Safety hooks (block-destructive.js)
└── settings.json     # Pre-allowed permissions for unattended agents
orchestrator/
├── watcher.py        # Jira polling, label state machine, 9 phases
├── dispatcher.py     # Agent dispatch with OpenShell sandbox support
├── jira_client.py    # REST API client for Jira (v3 ADF parsing)
├── config.py         # Config from env vars + projects.json
└── models.py         # Data models (Ticket, CycleStats)
policies/             # OpenShell sandbox policies (filesystem + network)
manifests/            # K8s manifests (namespace, RBAC, PVC, secrets, deployment)
docs/                 # Deployment guides and architecture
eval/                 # Model evaluation results
Containerfile         # UBI9 image with OpenCode, OpenShell, toolchain
opencode.json         # OpenCode config — MCP servers, instructions
AGENTS.md             # Project rules loaded into agent context

Initial skill patterns inspired by the AAP SDLC Harness (bugfix-workflow, code-review, git-workflow, jira-integration, ai-attribution). Skills have since been rewritten for OpenCode with structured playbooks, audit sub-agents, and MCP-based Jira integration.