AI agents degrade on long tasks — context rots, reviews get rubber-stamped, verification is an honor system. Wazir is the operating model that's missing.

$ /wazir Build a REST API with authentication

[clarify]     3 questions asked, answers collected
[specify]     47 acceptance criteria written
[spec-gate]   APPROVED
[plan]        6 implementation tasks
[plan-gate]   APPROVED
[execute]     6/6 tasks complete, 43 tests passing
[verify]      0 lint errors, proof artifact generated
[review]      2 findings, both resolved
[learn]       3 learnings captured

Pipeline complete. 3/3 gates passed.

Quick Start

Requires Claude Code and Node.js 20+.

npm install -g @wazir-dev/cli

Then in Claude Code:

/plugin marketplace add MohamedAbdallah-14/Wazir
/plugin install wazir

Run a task:

/wazir Build a REST API for task management with authentication

Control the depth: /wazir quick ... for fast fixes, /wazir deep ... for full pipeline, /wazir audit ... for dedicated audits.

What Makes This Different

clarify → specify → [gate] → design → [gate] → plan → [gate] → execute → verify → review → learn

• Mandatory research phase. Before any code is written, a researcher fetches live API docs, changelogs, and prior art. Not optional — a pipeline phase.

• Adversarial review. The reviewer is never the author. Three gates reject work back until quality passes.

• AI writing detection and removal. Strips AI patterns from specs, comments, and commit messages. Output reads like a human wrote it.

• Published compliance data. Every run scores itself across five dimensions. Numbers in SQLite, not marketing.

• 324 expertise modules, deterministic composition. Modules compose based on project context across 13 domains. Same input, same composition, every time.

• Export compiler: one source, four host-native packages. Write the process once, compile to Claude, Codex, Gemini, and Cursor.

• Fresh agent per step. Each pipeline phase gets a clean context window. No carryover. No contamination between research, implementation, and review.

Wazir vs. Alternatives

| Dimension | Wazir | Claude Code (bare) | Superpowers | Spec-Kit | Raw Prompting | |---|:---:|:---:|:---:|:---:|:---:| | Enforced delivery pipeline | 14 phases, 3 gates | — | — | Spec-first | — | | Mandatory pre-coding research | ✓ | — | — | — | — | | Adversarial review (reviewer ≠ author) | ✓ | — | — | — | — | | Expertise composition per task | 324 modules | — | ~15 skills | Extensions | — | | Published compliance measurement | ✓ | — | — | — | — | | AI writing detection + removal | ✓ | — | — | — | — | | Export compiler (one source → host-native) | 4 hosts | — | 5 hosts (per-host install) | 3 hosts | — |

Wazir is not competing with these tools — it learned from all of them. See Acknowledgments.

Documentation

| Section | What You'll Find | |---|---| | Architecture | System design, component interactions, context tiers | | Roles & Workflows | 10 roles, 14 phases, gate mechanics | | Composition Engine | How 324 modules are assembled per task | | Pipeline Vision | Every design decision with research citations | | Research Index | 122 research files across 14 categories |

Wazir (وزير): Arabic for advisor.

Active development. The pipeline works. Rough edges remain.

| Project | What Wazir Learned | |---|---| | superpowers | Skill system architecture, bootstrap injection pattern | | spec-kit | Specification-driven development patterns |

Full acknowledgments · Contributing · Security · MIT License