Add documentation

Author: OracleNep

docs/design.md

@@ -0,0 +1,48 @@
+# Design
+
+AgentFlowDesk is built around one opinion:
+
+> AI-agent work should be structured like engineering work, not like disposable chat history.
+
+## Pipeline
+
+```text
+User Task
+  ↓
+Task Brief
+  ↓
+Context Pack Builder
+  ↓
+Agent Export / Agent Command
+  ↓
+Logs + Diff + Checks
+  ↓
+Review Report
+  ↓
+Reusable Skill
+```
+
+## Non-goals
+
+- It is not a model provider.
+- It is not a universal autonomous agent framework.
+- It is not a replacement for GitHub, Jira, Linear, or IDE tools.
+- It does not claim to judge code correctness by itself.
+
+## MVP architecture
+
+```text
+agentflow/
+  cli.py              argparse CLI entrypoint
+  storage.py          local .agentflow JSON store
+  models.py           Task and RunRecord models
+  context_builder.py  context-pack generation
+  exporters.py        AGENTS.md / CLAUDE.md / Cursor / Codex / Gemini exports
+  runner.py           run command wrapper and report generation
+  review.py           markdown review report renderer
+  git_tools.py        git status/diff helpers
+```
+
+## Why local-first
+
+AI-agent usage often touches private repositories and unfinished code. The MVP stores metadata, prompts, logs, and reports locally under `.agentflow/`. Users decide which agent CLI or external model gets called.