governed-workflowSAFE

A zero-trust orchestration layer for Claude Code that enforces phase progression, scope constraints, and human approval gates on agentic coding sessions. Instead of letting the agent self-certify its work, every phase transition is validated server-side — research must be proven, file edits are scope-locked, and critical checkpoints require explicit human approval through the admin panel.

The system ships as a set of Claude Code extensions: agent definitions, hook scripts, skills, and a Flask-based admin panel with an MCP server.

How It Works

The workflow moves through assessment, research, planning, execution, review, reflection, and delivery. Four transitions are user gates that halt progress until a human approves or rejects via the admin panel.

flowchart TB
Start([0: Init])

subgraph P1["Phase 1 — Pre-planning"]
direction LR
A["1.0: AssessmentPost research questions"] --> R["1.1: ResearchAnswer questions with findings"] --> RP["1.2: Research Proving"] --> IA["1.3: Impact Analysis"] --> G1{{"1.4: Preparation ReviewUSER GATE"}}
G1 -.->|Reject| R
end

subgraph P2["Phase 2 — Planning"]
direction LR
PL["2.0: PlanningPropose criteria"] --> G2{{"2.1: Plan ReviewUSER GATEApproving plan accepts all criteria"}}
G2 -.->|Reject| PL
end

subgraph P3["Phase 3 — Execution (per sub-phase, loops)"]
direction LR
IM["3.N.0: Implementation"] --> V["3.N.1: Validation"]
V -->|Clean| G3{{"3.N.3: Code ReviewUSER GATE"}}
V -->|Issues found| F["3.N.2: Fixes"]
F -->|Re-validate| V
G3 -->|Approve| C["3.N.4: Commit"]
G3 -.->|Reject| F
C -.->|More sub-phases| IM
C -->|Last sub-phase| CV{"Criteriavalidated?"}
CV -.->|Fail| F
end

subgraph P4["Phase 4 — Final Review"]
direction LR
BR["4.0: Blind Code Review"] --> AV["4.1: Address &