Sovereign Dark Factory (local-dark-factory)¶
A 100% sovereign, local-first autonomous AI Software Factory orchestrator designed to coordinate local coding agents in fast, disposable git worktrees, enforce authoritative deterministic verification gates, auto-repair syntax and test failures, preserve tamper-evident patches, and provide human-in-the-loop review.
"Make sure you own your AI. AI in the cloud is not aligned with you; itβs aligned with the company that owns it."
π Core Pillars¶
flowchart TD
A[Task Specification] --> B[Durable SQLite Engine]
B --> C[Git Worktree Sandbox]
C --> D[Local Inference Harness: Ollama / Prism CUDA]
D --> E[Deterministic Verification Gates]
E -- Failed --> F[AST Self-Healing Loop max N retries]
F --> D
E -- Passed --> G[Evidence Locker SHA256 Diff & Telemetry]
G --> H[Human-in-the-Loop Review Gate]
H -- Approve --> I[Target Branch Commit]
H -- Reject --> J[Archived with Operator Notes]
-
:octicons-shield-lock-24: Zero Cloud Tokens Operates 100% locally on workstation hardware (NVIDIA RTX / Apple Silicon). Never makes external cloud API calls ($0.00 token cost).
-
:octicons-cpu-24: Disposable Sandboxes Eliminates heavyweight overhead by utilizing native
git worktreeisolation with directory sanitization and strict path-traversal prevention (repository-level isolation; process sandboxing is tracked inplan.mdΒ§10.5). -
:octicons-check-circle-24: Authoritative Verification Gates Verification is deterministic (compiler, linters, test suites). The LLM cannot vote on whether its own code works.
-
:octicons-lock-24: Gate Anti-Tamper Protection Automatic baseline restoration prevents the model from weakening assertions, commenting out tests, or modifying test runners to fake a pass.
-
:octicons-database-24: Durable SQLite Journaling State transitions, audit logs, and human operator notes survive system crashes and reboots. Complete transactional state recovery.
-
:octicons-sync-24: AST Self-Healing Loop Feeds compiler diagnostics and test assertion failures back to the local model up to configurable retry limits with cumulative telemetry.
β‘ Quick Start¶
# Install local-dark-factory
pip install local-dark-factory
# Check local engines and GPU health
dark-factory doctor
# Run an autonomous task against a repository
dark-factory run \
--repo /path/to/project \
--task "Fix bug in calculation algorithm and ensure all pytest cases pass"
π AI-Native SDLC Architecture¶
This project is built from the ground up according to Anthropic's AI-Native SDLC Playbook:
| Specification Artifact | Role in Factory |
|---|---|
Intent (intent.md) |
Business rationale, sovereignty mandate, and hardware constraints. |
Technical Spec (spec.md) |
Data contracts, architecture choices, and activity contracts. |
Operating Manual (CLAUDE.md) |
Coding standards, safety rules, and test guidelines. |
Agent Guidelines (AGENTS.md) |
Local inference skills routing and zero-cloud-token rule. |
Governance Policy (REVIEW.md) |
Operator approval criteria and diff inspection. |