[WIP] Add refactoring Agent

agents/refactoring-agent/.env.example

@@ -0,0 +1,5 @@
+GITHUB_TOKEN=your_github_personal_access_token_here
+REPO_OWNER=github_username_or_org
+REPO_NAME=repository_name
+GEMINI_API_KEY=your_gemini_api_key_here
+BASE_BRANCH=main

agents/refactoring-agent/.gitignore

@@ -0,0 +1,8 @@
+dist/
+node_modules/
+state.json
+.env
+.DS_Store
+.pnpm-store/
+.npm-cache/
+.yarn-cache/

agents/refactoring-agent/README.md

@@ -0,0 +1,56 @@
+# Refactoring Agent
+
+An autonomous Level-1 agent that reviews merged GitHub PRs since its last run, synthesises the diffs with local skill documentation, and opens a draft PR with prioritised refactoring suggestions.
+
+## Architecture
+
+```
+src/
+├── index.ts          # Agent orchestrator (entry point)
+├── config.ts         # Environment variable loading
+├── state.ts          # Reads/writes state.json (lastRunAt)
+├── skills.ts         # Loads markdown files from skills/ into system prompt
+└── tools/
+    ├── list-prs.ts         # makeListPrsTool — lists merged PRs since last run
+    ├── get-pr-diff.ts      # makeGetPrDiffTool — fetches unified diff
+    └── create-draft-pr.ts  # makeCreateDraftPrTool — creates the output draft PR
+prompts/system.md     # System prompt template ({SKILLS} and {REPO} are replaced at runtime)
+skills/               # Drop .md files here to give the agent domain context
+state.json            # Gitignored — persists lastRunAt timestamp
+```
+
+**Model:** `gemini-2.5-flash` via `@mariozechner/pi-ai` + `@mariozechner/pi-agent-core`
+
+## Setup
+
+```bash
+cp .env.example .env
+# Fill in GITHUB_TOKEN, REPO_OWNER, REPO_NAME, GEMINI_API_KEY
+npm install
+```
+
+## Running
+
+```bash
+npm run dev        # Quick run via tsx (no build needed)
+npm run build      # Compile to dist/
+npm start          # Run compiled output
+```
+
+## Adding Skills
+
+Drop any `.md` file into the `skills/` directory. These are read at startup and injected into the system prompt, giving the agent project-specific domain context (coding conventions, architectural patterns, etc.).
+
+## How It Works
+
+1. Loads `state.json` to get `lastRunAt` (defaults to 7 days ago on first run)
+2. Loads skill `.md` files and injects them into the system prompt
+3. Starts the agent loop with three tools: `list_prs`, `get_pr_diff`, `create_draft_pr`  
+4. The LLM lists PRs → fetches diffs → synthesises → creates a draft PR
+5. Saves updated `lastRunAt` to `state.json`
+
+## Tests
+
+```bash
+npm test
+```

agents/refactoring-agent/docs/plans/2026-03-06-make-a-task-list-to-implement-the-agent-described-in-plan.md

@@ -0,0 +1,54 @@
+---
+name: Refactoring Agent — Implementation Plan
+goal: Build a Level-1 autonomous agent using `@mariozechner/pi-agent-core` + `@mariozechner/pi-ai` that fetches GitHub PR diffs since its last run, synthesizes them with local skill docs, and opens a draft PR with prioritised refactoring suggestions.
+created: 2026-03-06T16:35:45+01:00
+progress: 0/0
+---
+
+# Refactoring Agent — Implementation Plan
+
+**Goal:** Build a Level-1 autonomous agent using `@mariozechner/pi-agent-core` + `@mariozechner/pi-ai` that fetches GitHub PR diffs since its last run, synthesizes them with local skill docs, and opens a draft PR with prioritised refactoring suggestions.
+
+**Architecture:** A standalone TypeScript project under `agent-hours/agents/refactoring-agent/`. The `Agent` class from `@mariozechner/pi-agent-core` is instantiated in `src/index.ts` with three `AgentTool` objects wrapping `@octokit/rest` (`list_prs`, `get_pr_diff`, `create_draft_pr`). A `state.json` file (gitignored) tracks `lastRunAt`. The system prompt (from `prompts/system.md`) is dynamically prepended with skill markdown files loaded via `src/skills.ts` — the domain context informs the LLM without modifying pi-mono internals.
+
+---
+
+## 8 Tasks at a Glance
+
+### Task 1: Project scaffolding
+**Key Output:** `package.json`, `tsconfig.json`, `.gitignore`
+
+### Task 2: Config & state management
+**Key Output:** `src/config.ts`, `src/state.ts`
+
+### Task 3: GitHub API tools
+**Key Output:** `src/tools/{list-prs,get-pr-diff,create-draft-pr}.ts`
+
+### Task 4: Skills context loader
+**Key Output:** `src/skills.ts`
+
+### Task 5: System prompt
+**Key Output:** `prompts/system.md`
+
+### Task 6: Agent orchestrator
+**Key Output:** `src/index.ts` — wires all pieces together
+
+### Task 7: README & submission
+**Key Output:** `README.md`, `SUBMISSION.md`
+
+### Task 8: Unit tests
+**Key Output:** `src/tools/list-prs.test.ts`, `vitest.config.ts`
+
+---
+
+## Key Design Decisions
+
+- **`AgentTool` factory functions** (e.g. `makeListPrsTool(octokit, owner, repo, since)`) close over the Octokit client and `since` timestamp — the `Agent` sees clean, parameterless tools while the software layer handles state.
+- **Skill loading at startup** — `loadSkills(dir)` reads all `.md` files from the configured skills directory and prepends them to the system prompt. This is the "knowledge injection" step that distinguishes this agent from a generic LLM call.
+- **`state.json`** holds only `{ lastRunAt: string | null }` — minimal, easily inspectable, gitignored.
+- **Error contract**: tools `throw` on GitHub API failures; pi-mono catches and reports these to the LLM as `isError: true` tool results, which naturally triggers graceful stopping per the operational boundaries in `plan.md`.
+- **`google/gemini-3-flash`** is used as the model via `getModel("google", "gemini-3-flash")` — fast and cheap for a text-only synthesis task.
+<!-- dispatch-status: task-1 = running -->
+<!-- dispatch-status: task-2 = running -->
+<!-- dispatch-status: task-3 = running -->
+<!-- dispatch-status: task-4 = running -->

agents/refactoring-agent/eslint.config.js

@@ -0,0 +1,53 @@
+// @ts-check
+import tseslint from "typescript-eslint";
+
+export default tseslint.config(
+    // Base: recommended + strict + stylistic type-checked rules
+    ...tseslint.configs.strictTypeChecked,
+    ...tseslint.configs.stylisticTypeChecked,
+
+    {
+        languageOptions: {
+            parserOptions: {
+                project: "tsconfig.eslint.json",
+                tsconfigRootDir: import.meta.dirname,
+            },
+        },
+        rules: {
+            // ── Correctness ───────────────────────────────────────────────
+            "@typescript-eslint/no-explicit-any": "error",
+            "@typescript-eslint/no-unsafe-assignment": "error",
+            "@typescript-eslint/no-unsafe-call": "error",
+            "@typescript-eslint/no-unsafe-member-access": "error",
+            "@typescript-eslint/no-unsafe-return": "error",
+            "@typescript-eslint/no-floating-promises": "error",
+
+            // Allow _-prefixed params (required by AgentTool interface signature)
+            "@typescript-eslint/no-unused-vars": ["error", { argsIgnorePattern: "^_" }],
+
+            // ── Clarity ───────────────────────────────────────────────────
+            "@typescript-eslint/explicit-function-return-type": ["error", { allowExpressions: true }],
+            "@typescript-eslint/consistent-type-imports": ["error", { prefer: "type-imports" }],
+            "@typescript-eslint/no-import-type-side-effects": "error",
+
+            // ── Safety ────────────────────────────────────────────────────
+            "@typescript-eslint/no-non-null-assertion": "error",
+            "no-console": "off", // console logging is intentional in CLI tools
+            "eqeqeq": ["error", "always"],
+        },
+    },
+
+    // Relax rules for test files
+    {
+        files: ["**/*.test.ts"],
+        rules: {
+            "@typescript-eslint/no-unsafe-assignment": "off",
+            "@typescript-eslint/no-unsafe-call": "off",
+            "@typescript-eslint/no-unsafe-member-access": "off",
+            "@typescript-eslint/no-explicit-any": "off",
+        },
+    },
+
+    // Ignore compiled output and the eslint config itself
+    { ignores: ["dist/**", "eslint.config.js"] },
+);