Problem
Current AI coding tools lack governance. They mix planning, coding, and review into a single reasoning stream with no role separation, no mandatory gates, and no auditability. The result is refactor drift, scope violations, hallucinated changes, and poor traceability.
Solution
Software Fabric is a governance-driven orchestration system for AI coding agents. It transforms LLM-based development from chat-based coding into an industrialized, traceable, repeatable engineering process.
Core design principles:
- CI is authority — gates decide correctness, not the model
- Roles are separated even when the same model is used (Planner / Implementer / Reviewer)
- Artifacts over conversation memory — every run produces immutable outputs
- Policy precedes autonomy — allowlists, diff budgets, and iteration limits are enforced
- Cost control is built-in — local models for planning and review, remote only on demand
System components:
- fabric-api — REST endpoints, job lifecycle management
- fabric-worker — orchestration loop, provider calls, retry logic
- fabric-runner — Docker sandbox for patch application and gate execution
- fabric-policy — allowlist enforcement, diff budget validation, iteration caps
- Providers layer — local (Ollama), OpenAI, Anthropic
Architecture
Each job moves through a defined state machine: CREATED → PLANNED → IMPLEMENTING → VERIFYING → REVIEWING → READY_TO_RELEASE → RELEASED / FAILED. Mandatory gates (pytest, ruff, mypy, pip-audit) block progression on failure. All changes produce structured artifacts: SPEC.md, DIFF.patch, VERIFY.log, REVIEW.md, results.json.
Result
A system that makes AI-assisted software production auditable and controllable — not by limiting what AI can do, but by structuring how it operates under human architectural supervision.