chore: import upstream snapshot with attribution

2026-07-13 12:24:16 +08:00
commit 4b92209caa
317 changed files with 67848 additions and 0 deletions
@@ -0,0 +1,69 @@
+# Contributing to SkillOpt
+
+Thank you for your interest in contributing to SkillOpt! This guide covers how to get started.
+
+## Development Setup
+
+```bash
+git clone https://github.com/microsoft/SkillOpt.git
+cd SkillOpt
+pip install -e ".[dev]"
+```
+
+## Ways to Contribute
+
+### 🐛 Bug Reports
+
+Open an issue with:
+- Steps to reproduce
+- Expected vs actual behavior
+- Config file used (sanitize API keys)
+- Python version and OS
+
+### 🔧 New Benchmark
+
+See [Add a New Benchmark](guide/new-benchmark.md) for the implementation guide.
+
+**Checklist:**
+- [ ] Data loader in `skillopt/envs/<benchmark>/loader.py`
+- [ ] Environment adapter in `skillopt/envs/<benchmark>/env.py`
+- [ ] Config file in `configs/<benchmark>/default.yaml`
+- [ ] Registration in `skillopt/envs/__init__.py`
+- [ ] Documentation page in `docs/`
+
+### 🤖 New Model Backend
+
+See [Add a New Model Backend](guide/new-backend.md) for the implementation guide.
+
+**Checklist:**
+- [ ] Backend in `skillopt/model/<backend>.py`
+- [ ] Registration in `skillopt/model/__init__.py`
+- [ ] API key entry in `.env.example`
+- [ ] Documentation update
+
+### 📝 Documentation
+
+Documentation is built with MkDocs Material:
+
+```bash
+pip install -e ".[docs]"
+mkdocs serve  # Preview at http://localhost:8000
+```
+
+## Code Style
+
+- Follow existing patterns in the codebase
+- Use type hints for function signatures
+- Keep docstrings concise
+
+## Pull Request Process
+
+1. Fork the repository
+2. Create a feature branch: `git checkout -b feature/my-benchmark`
+3. Make your changes
+4. Test with an existing benchmark config
+5. Submit a PR with a clear description
+
+## License
+
+By contributing, you agree that your contributions will be licensed under the MIT License.
@@ -0,0 +1,139 @@
+# Configuration Guide
+
+SkillOpt uses YAML configuration files with a hierarchical override system.
+
+## Config Structure
+
+```
+configs/
+├── _base_/
+│   └── default.yaml          # Global defaults
+├── searchqa/
+│   └── default.yaml          # SearchQA overrides
+├── docvqa/
+│   └── default.yaml          # DocVQA overrides
+└── alfworld/
+    └── default.yaml          # ALFWorld overrides
+```
+
+Benchmark configs inherit from `_base_/default.yaml` and override specific values.
+
+## Key Parameters
+
+### Model
+
+```yaml
+model:
+  backend: azure_openai          # azure_openai | openai_chat | claude_code_exec | qwen
+  optimizer: gpt-5.5               # Optimizer model (for reflection)
+  target: gpt-5.5               # Target model (for rollout)
+```
+
+### Training
+
+```yaml
+train:
+  num_epochs: 4                  # Number of training epochs
+  batch_size: 40                 # Tasks per step (batch size)
+  accumulation: 1                # Gradient accumulation
+  seed: 42
+```
+
+### Gradient (Reflection)
+
+```yaml
+gradient:
+  minibatch_size: 8              # Reflect minibatch size
+  analyst_workers: 16            # Parallel reflection workers
+  max_analyst_rounds: 3          # Max rounds of analyst reflection
+  failure_only: false            # Only reflect on failures
+```
+
+### Optimizer
+
+```yaml
+optimizer:
+  learning_rate: 4               # Max edits per step (edit budget)
+  min_learning_rate: 2           # Min edits for decay schedulers
+  lr_scheduler: cosine           # constant | linear | cosine | autonomous
+  use_slow_update: true          # Momentum-like blending at epoch boundary
+  slow_update_samples: 20        # Samples for slow update evaluation
+  use_meta_skill: true           # Cross-epoch strategy memory
+```
+
+### Skill-Aware Reflection (optional, off by default)
+
+EmbodiSkill-style failure routing: the failure analyst classifies each
+failure pattern as **SKILL_DEFECT** (the rule is wrong or missing → normal
+gated body edit) or **EXECUTION_LAPSE** (a valid rule exists but was not
+followed → a short reminder appended to a protected appendix region inside
+the skill that step-level edits can never modify).
+
+```yaml
+optimizer:
+  use_skill_aware_reflection: false    # Master switch (default off = baseline-identical)
+  skill_aware_appendix_source: both    # both | failure_only (paper-faithful S_app)
+  skill_aware_consolidate_threshold: 0 # >0: LLM-compact the appendix past N notes (experimental)
+```
+
+Notes:
+
+- The switch is resolved process-wide from the config
+  (`configure_skill_aware_reflection`), so it applies to every benchmark
+  with no per-adapter wiring.
+- `failure_only` restricts appendix notes to the failure analyst, matching
+  the original S_app formulation; `both` additionally lets the success
+  analyst re-emphasize existing rules.
+- Appendix notes bypass the validation gate by design and accumulate with
+  order-preserving dedup; lapse-only steps (no body edits) still flush
+  their notes.
+- Not supported together with `skill_update_mode=rewrite_from_suggestions`
+  or the full-rewrite modes: whole-document rewrites can drop the appendix
+  region.
+
+### Evaluation
+
+```yaml
+evaluation:
+  use_gate: true                 # Validation gating (accept/reject updates)
+  eval_test: true                # Run test evaluation after training
+```
+
+### Environment (Data)
+
+```yaml
+env:
+  name: searchqa                 # Benchmark name
+  split_mode: ratio              # ratio | split_dir
+  split_ratio: "2:1:7"           # train:val:test ratio
+  data_path: ""                  # Path to dataset
+  exec_timeout: 120              # Per-task timeout (seconds)
+```
+
+## CLI Overrides
+
+Override any config value from the command line:
+
+```bash
+python scripts/train.py \
+  --config configs/searchqa/default.yaml \
+  optimizer.learning_rate=16 \
+  optimizer.lr_scheduler=linear \
+  gradient.analyst_workers=8
+```
+
+## Environment Variables
+
+Model credentials are loaded from environment variables:
+
+| Variable | Backend | Description |
+|---|---|---|
+| `AZURE_OPENAI_ENDPOINT` | azure_openai | Azure resource endpoint |
+| `AZURE_OPENAI_API_KEY` | azure_openai | Azure API key |
+| `OPENAI_API_KEY` | openai | OpenAI API key |
+| `ANTHROPIC_API_KEY` | claude | Anthropic API key |
+| `QWEN_API_BASE` | qwen | Local Qwen vLLM endpoint |
+
+## Full Reference
+
+See [Configuration Reference](../reference/config.md) for the complete parameter list.
@@ -0,0 +1,51 @@
+# Deep Learning ↔ SkillOpt Analogy
+
+SkillOpt is designed around a core insight: **optimizing natural-language prompts follows the same structure as training neural networks**. This page maps every DL concept to its SkillOpt counterpart.
+
+## Complete Mapping
+
+| Deep Learning | SkillOpt | Description |
+|---|---|---|
+| **Model weights** | Skill document (Markdown) | The thing being optimized |
+| **Forward pass** | Rollout | Target executes tasks using current skill |
+| **Loss function** | Task evaluator | Scores task execution quality |
+| **Backpropagation** | Reflect | Optimizer analyzes failures → edit patches |
+| **Gradients** | Edit patches | Proposed changes to the skill |
+| **Gradient aggregation** | Patch aggregation | Merge similar edits |
+| **Gradient clipping** | Edit selection | Cap max edits per step |
+| **Learning rate** | `learning_rate` | Max number of edits applied per step |
+| **LR scheduler** | `lr_scheduler` | Decay schedule: cosine, linear, constant |
+| **SGD step** | Skill update | Apply selected patches to document |
+| **Validation set** | Selection split | Gate checks improvement before accepting |
+| **Early stopping** | Gate patience | Reject updates that don't improve |
+| **Training step** | Step | One rollout → reflect → update cycle |
+| **Epoch** | Epoch | Full pass with slow update + meta memory |
+| **Momentum** | Slow update | Longitudinal comparison at epoch boundary |
+| **Meta-learning** | Meta skill | Cross-epoch optimizer strategy memory |
+| **Batch size** | `batch_size` | Tasks sampled per rollout |
+| **Data parallelism** | `analyst_workers` | Parallel reflection workers |
+| **Training set** | Train split | Items used for rollout |
+| **Test set** | Test split | Held-out final evaluation |
+| **Warm-up** | (implicit) | High LR early steps explore broadly |
+| **Checkpointing** | Skill snapshots | Saved after each accepted step |
+| **Transfer learning** | Seed skill / cross-benchmark init | Start from pre-trained skill |
+
+## Why This Analogy Matters
+
+1. **Familiar mental model**: ML practitioners immediately understand how to tune SkillOpt
+2. **Principled hyperparameter search**: Grid search over `learning_rate` × `lr_scheduler` works just like in DL
+3. **Proven mechanisms**: Gating ≈ validation-based selection, patience ≈ early stopping, slow update ≈ momentum — all with strong theoretical motivation
+
+## Hyperparameter Transfer Rules
+
+From our experiments, these DL intuitions transfer well:
+
+!!! success "What transfers"
+    - **Cosine schedule > constant** — same as in DL, cosine annealing helps convergence
+    - **Moderate LR (4-16) > very high/low** — too few edits = slow learning, too many = noisy
+    - **Slow update helps** — longitudinal comparison prevents catastrophic forgetting across epochs
+    - **Meta skill memory improves reflection** — optimizer benefits from cross-epoch strategy notes
+
+!!! warning "What doesn't transfer"
+    - **Batch size ≠ better** — larger rollout batches have diminishing returns due to API costs
+    - **More epochs ≠ better** — skills converge faster than neural networks (2-4 epochs usually enough)
@@ -0,0 +1,110 @@
+# Your First Experiment
+
+This guide walks through running a complete SkillOpt training on SearchQA.
+
+## 1. Choose a Benchmark
+
+SkillOpt includes ready-to-use configs for several benchmarks:
+
+| Benchmark | Difficulty | Typical Runtime |
+|---|---|---|
+| SearchQA | ⭐ Easy | ~30 min |
+| DocVQA | ⭐⭐ Medium | ~2 hours |
+| ALFWorld | ⭐⭐⭐ Hard | ~3 hours |
+
+We'll use **SearchQA** as it's the fastest to complete.
+
+## 2. Configure
+
+Review the config file:
+
+```bash
+cat configs/searchqa/default.yaml
+```
+
+Key parameters (deep learning analogy in parentheses):
+
+```yaml
+train:
+  num_epochs: 4           # (epochs)
+  batch_size: 40          # (batch size)
+
+optimizer:
+  learning_rate: 4        # (max edits per step)
+  lr_scheduler: cosine    # (learning rate schedule)
+  use_slow_update: true   # (momentum at epoch boundary)
+  use_meta_skill: true    # (cross-epoch optimizer memory)
+
+gradient:
+  analyst_workers: 16     # (parallel reflection workers)
+
+evaluation:
+  use_gate: true          # (validation gating)
+```
+
+## 3. Train
+
+```bash
+python scripts/train.py --config configs/searchqa/default.yaml
+```
+
+You'll see output like:
+
+```
+[Step 1/8] Rollout: 20 items, 4 workers...
+[Step 1/8] Score: 0.65 → Reflect...
+[Step 1/8] 6 edit patches generated
+[Step 1/8] Selected 4 edits (lr=8, cosine → 7.7)
+[Step 1/8] Gate: val score 0.68 > 0.65 ✓ ACCEPT
+[Step 2/8] ...
+```
+
+## 4. Monitor
+
+Training outputs are saved to `outputs/<benchmark>/<run_id>/`:
+
+```
+outputs/searchqa/2024-01-15_10-30-00/
+├── steps/
+│   ├── step_0001/
+│   │   ├── candidate_skill.md
+│   │   ├── step_record.json
+│   │   └── trajectory_digest.json
+│   └── step_0002/
+├── slow_update/
+│   └── epoch_02/
+├── meta_skill/
+│   └── epoch_02/
+├── skills/
+│   └── step_0001.md
+├── best_skill.md
+├── history.json
+└── config.yaml
+```
+
+## 5. Evaluate
+
+Evaluate the best skill on the test split:
+
+```bash
+python scripts/eval_only.py \
+  --config configs/searchqa/default.yaml \
+  --skill outputs/searchqa/<run_id>/skills/best_skill.md
+```
+
+## WebUI
+
+Prefer a graphical interface? Launch the WebUI:
+
+```bash
+pip install -e ".[webui]"
+python -m skillopt_webui.app
+```
+
+Then open `http://localhost:7860` in your browser to configure parameters and launch training.
+
+## Next Steps
+
+- [Understand the training loop](training-loop.md)
+- [Configuration reference](../reference/config.md)
+- [Add a new benchmark](new-benchmark.md)
@@ -0,0 +1,89 @@
+# Installation
+
+## Requirements
+
+- Python ≥ 3.10
+- At least one model API key (Azure OpenAI, OpenAI, Anthropic, or local Qwen)
+
+## Quick Install
+
+```bash
+git clone https://github.com/microsoft/SkillOpt.git
+cd SkillOpt
+pip install -e .
+```
+
+## Optional Dependencies
+
+Install extras for specific benchmarks or backends:
+
+=== "ALFWorld"
+
+    ```bash
+    pip install -e ".[alfworld]"
+    ```
+
+=== "Claude Backend"
+
+    ```bash
+    pip install -e ".[claude]"
+    ```
+
+=== "Qwen (Local)"
+
+    ```bash
+    pip install -e ".[qwen]"
+    ```
+
+=== "WebUI"
+
+    ```bash
+    pip install -e ".[webui]"
+    ```
+
+=== "Development"
+
+    ```bash
+    pip install -e ".[dev]"
+    ```
+
+=== "All"
+
+    ```bash
+    pip install -e ".[alfworld,claude,qwen,webui,dev]"
+    ```
+
+## Environment Variables
+
+Copy the example `.env` file and fill in your credentials:
+
+```bash
+cp .env.example .env
+```
+
+Edit `.env` with your API keys:
+
+```ini
+# Azure OpenAI (default backend)
+AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
+AZURE_OPENAI_API_KEY=your-key
+
+# Or use OpenAI directly
+OPENAI_API_KEY=sk-...
+
+# Or Anthropic Claude
+ANTHROPIC_API_KEY=sk-ant-...
+```
+
+!!! tip
+    You only need credentials for the backend you plan to use. Azure OpenAI is the default.
+
+## Verify Installation
+
+```bash
+python -c "import skillopt; print('SkillOpt ready!')"
+```
+
+## Next Steps
+
+→ [Run your first experiment](first-experiment.md)
@@ -0,0 +1,143 @@
+# Local Environment Smoke Tests
+
+This guide describes a lightweight pattern for testing a custom SkillOpt environment before connecting it to expensive model calls or a full benchmark dataset.
+
+The goal is to validate the training loop plumbing first:
+
+- config loading
+- adapter construction
+- dataloader splits
+- rollout output shape
+- reflection patch shape
+- merge/rank/update control flow
+- artifact creation under `out_root`
+
+Once those are stable, you can switch the same environment to real model calls and larger evaluation splits.
+
+## 1. Add a tiny fixture split
+
+Start with a handful of deterministic examples that cover the expected pass/fail cases for your environment. Keep them small enough that a single training step can run locally.
+
+A minimal fixture item usually needs:
+
+```json
+{
+  "id": "example-1",
+  "split": "train",
+  "question": "...",
+  "expected": "..."
+}
+```
+
+Use the split names your adapter maps to SkillOpt phases:
+
+- `train` for optimization rollouts
+- `val` or `valid_seen` for selection/gating
+- `test` or `valid_unseen` for final evaluation
+
+## 2. Support an offline mock mode
+
+Add a configuration flag such as `mock: true` to your adapter. In mock mode, `rollout()` should return deterministic responses without calling external model APIs.
+
+This lets you verify the SkillOpt loop with a fast command such as:
+
+```bash
+python scripts/train.py \
+  --config configs/myenv/tiny_mock.yaml
+```
+
+Mock mode should still write the same artifacts as a real run, for example:
+
+- `responses.json`
+- `rollout_results.json`
+- `ranked_edits.json`
+- `candidate_skill.md`
+- `summary.json`
+
+## 3. Keep the smoke config tiny
+
+A CI-friendly smoke config should run a single small step:
+
+```yaml
+train:
+  num_epochs: 1
+  train_size: 3
+  batch_size: 3
+
+gradient:
+  minibatch_size: 1
+  merge_batch_size: 2
+  analyst_workers: 1
+  max_analyst_rounds: 1
+
+optimizer:
+  learning_rate: 1
+  min_learning_rate: 1
+  lr_scheduler: constant
+  skill_update_mode: patch
+  use_slow_update: false
+
+evaluation:
+  use_gate: true
+  sel_env_num: 2
+  test_env_num: 2
+  eval_test: false
+
+env:
+  name: myenv
+  out_root: outputs/myenv_tiny_mock
+  mock: true
+```
+
+Prefer a mock config that runs without credentials. That makes it useful for contributors and CI.
+
+## 4. Validate optimizer JSON before returning it
+
+If your environment or extension asks an LLM to merge or rank skill edits, validate the returned JSON before passing it back into SkillOpt. This avoids silent fallbacks from empty, malformed, or out-of-range responses.
+
+Useful checks for edit payloads:
+
+- response is a JSON object
+- `edits` is a non-empty list
+- every edit is an object
+- every edit has an allowed operation
+- required fields such as `content` or `target` are present for that operation
+
+Useful checks for ranking payloads:
+
+- `selected_indices` exists
+- indices are integers
+- indices are unique
+- indices are within the candidate edit range
+- selected count does not exceed the edit budget
+
+On failure, retry with a compact prompt that includes the schema error. If retries fail, raise an explicit error instead of silently accepting malformed output.
+
+## 5. Run progressively stronger checks
+
+A good development sequence is:
+
+```bash
+python -m py_compile scripts/train.py skillopt/envs/myenv/adapter.py
+python scripts/train.py --config configs/myenv/tiny_mock.yaml
+python scripts/train.py --config configs/myenv/tiny.yaml
+```
+
+For the real tiny run, verify that:
+
+- the run completes
+- `summary.json` is written
+- `ranked_edits.json` contains the expected ranking metadata
+- any optimizer bridge log marks the response schema as valid
+- no generated files are written outside `out_root`
+
+## 6. Keep custom environments isolated
+
+When adding a custom environment to the registry, avoid side effects for existing benchmarks:
+
+- lazy-import optional dependencies
+- install environment-specific hooks only when `cfg["env"]` matches your environment
+- keep mock behavior behind an explicit config flag
+- write generated artifacts only under `out_root`
+
+This makes it easier to review and test a custom integration without affecting the built-in benchmarks.
@@ -0,0 +1,130 @@
+# Add a New Model Backend
+
+SkillOpt supports multiple LLM backends. This guide shows how to add your own.
+
+## Backend Architecture
+
+```
+skillopt/model/
+├── base.py           # Abstract base class
+├── azure_openai.py   # Azure OpenAI backend
+├── openai_model.py   # Direct OpenAI backend
+├── claude.py         # Anthropic Claude backend
+├── qwen.py           # Local Qwen (vLLM) backend
+└── your_backend.py   # Your new backend
+```
+
+## Step 1: Create the Backend
+
+Create `skillopt/model/your_backend.py`:
+
+```python
+from skillopt.model.base import ModelBackend, ModelResponse
+
+class YourBackend(ModelBackend):
+    """Your custom model backend."""
+    
+    def __init__(self, cfg: dict):
+        super().__init__(cfg)
+        self.model_name = cfg.get('model_name', 'your-default-model')
+        self.api_key = os.environ.get('YOUR_API_KEY', '')
+        self.client = self._init_client()
+    
+    def _init_client(self):
+        """Initialize API client."""
+        # TODO: Set up your API client
+        pass
+    
+    async def generate(
+        self,
+        messages: list[dict],
+        temperature: float = 0.7,
+        max_tokens: int = 4096,
+        **kwargs
+    ) -> ModelResponse:
+        """
+        Generate a completion.
+        
+        Args:
+            messages: Chat messages [{"role": "...", "content": "..."}]
+            temperature: Sampling temperature
+            max_tokens: Maximum tokens in response
+            
+        Returns:
+            ModelResponse with content, usage, and metadata
+        """
+        response = await self.client.chat(
+            model=self.model_name,
+            messages=messages,
+            temperature=temperature,
+            max_tokens=max_tokens,
+        )
+        
+        return ModelResponse(
+            content=response.text,
+            usage={
+                'prompt_tokens': response.usage.input,
+                'completion_tokens': response.usage.output,
+            },
+            model=self.model_name,
+        )
+    
+    async def generate_with_tools(
+        self,
+        messages: list[dict],
+        tools: list[dict],
+        **kwargs
+    ) -> ModelResponse:
+        """Generate with tool/function calling support."""
+        # Optional: implement if your model supports tool use
+        raise NotImplementedError("Tool use not supported")
+```
+
+## Step 2: Register the Backend
+
+Add to `skillopt/model/__init__.py`:
+
+```python
+from .your_backend import YourBackend
+
+BACKEND_REGISTRY = {
+    # ... existing backends ...
+    'your_backend': YourBackend,
+}
+```
+
+## Step 3: Configure
+
+Use your backend in any config:
+
+```yaml
+model:
+  backend: your_backend
+  model_name: your-model-id
+  temperature: 0.7
+  max_tokens: 4096
+```
+
+Set credentials via environment variable:
+
+```bash
+export YOUR_API_KEY="your-key"
+```
+
+## Required Interface
+
+Your backend must implement these methods:
+
+| Method | Required | Description |
+|---|---|---|
+| `generate()` | ✅ | Basic text generation |
+| `generate_with_tools()` | Optional | Tool/function calling |
+| `count_tokens()` | Optional | Token counting for context management |
+
+## Tips
+
+!!! tip
+    - Test your backend with `python -c "from skillopt.model.your_backend import YourBackend"` first
+    - Use `async` methods for all API calls — SkillOpt uses asyncio throughout
+    - Implement retry logic with exponential backoff for production use
+    - Add your API key to `.env.example` when submitting a PR
@@ -0,0 +1,371 @@
+# Add a New Benchmark
+
+Extend SkillOpt with your own benchmark in ~200 lines of code. We will use
+a tiny worked example, `docfaithful`, that scores a target model on
+how faithfully it answers questions grounded in a small reference doc.
+
+> **Working reference.** The easiest way to copy-cargo-cult a new env is
+> to read [`skillopt/envs/officeqa/`](https://github.com/microsoft/SkillOpt/tree/main/skillopt/envs/officeqa).
+> Everything below is the same shape, simplified.
+
+## What you need to build
+
+To add a benchmark you implement four things:
+
+1. **A `SplitDataLoader` subclass** — knows how to load train / val / test
+   item dicts from disk.
+2. **A rollout helper** — runs the target model on a batch of items
+   under the current skill and scores each prediction.
+3. **An `EnvAdapter` subclass** — wires the loader + rollout helper into
+   SkillOpt's lifecycle (`build_*_env`, `rollout`, `reflect`,
+   `get_task_types`).
+4. **A YAML config** — references your env name plus the standard
+   train / optimizer / gradient knobs.
+
+Then one line in `scripts/train.py`'s `_register_builtins()` makes it
+discoverable.
+
+---
+
+## Step 1 — Create the package
+
+```bash
+mkdir -p skillopt/envs/docfaithful
+touch skillopt/envs/docfaithful/__init__.py
+```
+
+## Step 2 — Implement the data loader
+
+`skillopt/envs/docfaithful/dataloader.py`:
+
+```python
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+from skillopt.datasets.base import SplitDataLoader
+
+
+def _normalize(raw: dict) -> dict:
+    """Make sure every item has an ``id``. Other keys are env-specific."""
+    return {
+        "id": str(raw["uid"]),
+        "question": raw["question"],
+        "ground_truth": raw["answer"],
+        "reference_text": raw.get("reference", ""),
+        "task_type": raw.get("category", "docfaithful"),
+    }
+
+
+class DocFaithfulDataLoader(SplitDataLoader):
+    """Load DocFaithful items from JSON files inside each split dir."""
+
+    def load_split_items(self, split_path: str) -> list[dict]:
+        # split_path is e.g. data/docfaithful_split/train/
+        json_files = sorted(Path(split_path).glob("*.json"))
+        if not json_files:
+            raise FileNotFoundError(f"No .json file found in {split_path}")
+        with json_files[0].open(encoding="utf-8") as f:
+            raw = json.load(f)
+        return [_normalize(item) for item in raw]
+```
+
+Only `load_split_items()` is mandatory. If you also want to support
+`split_mode="ratio"` (auto-split a single raw file into train/val/test),
+override `load_raw_items(data_path)` as well — see
+`skillopt/datasets/base.py` docstrings.
+
+## Step 3 — Write the rollout helper
+
+`skillopt/envs/docfaithful/rollout.py`:
+
+```python
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+
+from skillopt.model import chat_target
+
+
+def _score(prediction: str, ground_truth: str) -> tuple[int, float]:
+    """Trivial exact-match scorer. Replace with F1 / ROUGE / LLM-judge."""
+    p = (prediction or "").strip().lower()
+    g = (ground_truth or "").strip().lower()
+    hard = int(p == g and bool(g))
+    soft = 1.0 if hard else 0.0
+    return hard, soft
+
+
+def _rollout_one(item: dict, skill_content: str,
+                 *, max_completion_tokens: int) -> dict:
+    system = skill_content
+    user = (
+        f"Question: {item['question']}\n\n"
+        f"Reference:\n{item.get('reference_text', '')}\n\n"
+        "Answer:"
+    )
+    prediction, _usage = chat_target(
+        system=system,
+        user=user,
+        max_completion_tokens=max_completion_tokens,
+    )
+    hard, soft = _score(prediction, item.get("ground_truth", ""))
+    return {
+        "id": str(item["id"]),
+        "hard": hard,
+        "soft": soft,
+        "predicted_answer": prediction,
+        "question": item.get("question", ""),
+        "reference_text": item.get("reference_text", ""),
+        "task_type": item.get("task_type", "docfaithful"),
+    }
+
+
+def run_batch(*, items: list[dict], skill_content: str, out_root: str,
+              workers: int = 4, max_completion_tokens: int = 4096) -> list[dict]:
+    """Run a batch of episodes sequentially or with a thread pool."""
+    os.makedirs(out_root, exist_ok=True)
+    # For brevity we go sequentially — swap in concurrent.futures.ThreadPoolExecutor
+    # when network / model latency dominates.
+    results = [
+        _rollout_one(item, skill_content,
+                     max_completion_tokens=max_completion_tokens)
+        for item in items
+    ]
+    Path(out_root, "rollouts.json").write_text(
+        json.dumps(results, ensure_ascii=False, indent=2)
+    )
+    return results
+```
+
+Two design points worth flagging:
+
+- **Scoring lives here, not in `EnvAdapter`.** There is no `evaluate()`
+  method on the ABC. Whatever signal you put in `hard` (0/1, or a float
+  in [0, 1] for smoothed reward) and `soft` (float in [0, 1]) is what
+  the optimizer reads.
+- **Use `skillopt.model.chat_target`**, not raw OpenAI/Claude calls.
+  That routes through whichever **chat** target backend the user
+  configured (`openai_chat` / `claude_chat` / `qwen_chat` /
+  `minimax_chat`) without your adapter caring. Exec-style backends
+  (`codex_exec`, `claude_code_exec`) need env-specific rollout code —
+  see `skillopt/envs/swebench/` for an example.
+
+## Step 4 — Implement the environment adapter
+
+`skillopt/envs/docfaithful/adapter.py`:
+
+```python
+from __future__ import annotations
+
+from skillopt.datasets.base import BatchSpec
+from skillopt.envs.base import EnvAdapter
+from skillopt.envs.docfaithful.dataloader import DocFaithfulDataLoader
+from skillopt.envs.docfaithful.rollout import run_batch
+
+
+class DocFaithfulAdapter(EnvAdapter):
+    """SkillOpt adapter for the DocFaithful benchmark."""
+
+    def __init__(
+        self,
+        split_dir: str = "",
+        data_path: str = "",
+        split_mode: str = "split_dir",
+        split_ratio: str = "2:1:7",
+        split_seed: int = 42,
+        split_output_dir: str = "",
+        workers: int = 4,
+        analyst_workers: int = 4,
+        failure_only: bool = False,
+        minibatch_size: int = 8,
+        edit_budget: int = 4,
+        seed: int = 42,
+        limit: int = 0,
+        max_completion_tokens: int = 4096,
+    ) -> None:
+        self.workers = workers
+        self.analyst_workers = analyst_workers
+        self.failure_only = failure_only
+        self.minibatch_size = minibatch_size
+        self.edit_budget = edit_budget
+        self.max_completion_tokens = int(max_completion_tokens)
+        self.dataloader = DocFaithfulDataLoader(
+            split_dir=split_dir,
+            data_path=data_path,
+            split_mode=split_mode,
+            split_ratio=split_ratio,
+            split_seed=split_seed,
+            split_output_dir=split_output_dir,
+            seed=seed,
+            limit=limit,
+        )
+
+    # ── Lifecycle ───────────────────────────────────────────────────────
+
+    def setup(self, cfg: dict) -> None:
+        super().setup(cfg)
+        self.dataloader.setup(cfg)
+
+    def get_dataloader(self):
+        return self.dataloader
+
+    # ── Env construction ────────────────────────────────────────────────
+
+    def build_env_from_batch(self, batch: BatchSpec, **kwargs):
+        # For dataset-backed envs the "manager" is just the items list.
+        return list(batch.payload or [])
+
+    def build_train_env(self, batch_size: int, seed: int, **kwargs):
+        batch = self.dataloader.build_train_batch(
+            batch_size=batch_size, seed=seed, **kwargs
+        )
+        return self.build_env_from_batch(batch, **kwargs)
+
+    def build_eval_env(self, env_num: int, split: str, seed: int, **kwargs):
+        batch = self.dataloader.build_eval_batch(
+            env_num=env_num, split=split, seed=seed, **kwargs
+        )
+        return self.build_env_from_batch(batch, **kwargs)
+
+    # ── The rollout method (reflect is inherited) ───────────────────────
+
+    def rollout(self, env_manager, skill_content: str,
+                out_dir: str, **kwargs) -> list[dict]:
+        items: list[dict] = env_manager
+        return run_batch(
+            items=items,
+            skill_content=skill_content,
+            out_root=out_dir,
+            workers=self.workers,
+            max_completion_tokens=self.max_completion_tokens,
+        )
+
+    # reflect() is inherited from EnvAdapter — it delegates to
+    # run_minibatch_reflect with your analyst_error_* / analyst_success_*
+    # prompts. Override it only if you need custom reflection logic.
+
+    def get_task_types(self) -> list[str]:
+        seen: list[str] = []
+        for item in (
+            self.dataloader.train_items
+            + self.dataloader.val_items
+            + self.dataloader.test_items
+        ):
+            tt = str(item.get("task_type") or "docfaithful")
+            if tt not in seen:
+                seen.append(tt)
+        return seen or ["docfaithful"]
+```
+
+### What the rollout actually does
+
+Look back at `run_batch` from Step 3 — it sends each `item["question"]`
+to the target model with `skill_content` as the system prompt, scores
+the answer against `item["ground_truth"]`, and returns a list of dicts:
+
+```python
+[
+    {"id": "ex_001", "hard": 1, "soft": 0.92,
+     "predicted_answer": "...", "question": "...",
+     "reference_text": item["reference_text"]},
+    {"id": "ex_002", "hard": 0, "soft": 0.13, "fail_reason": "...", ...},
+    ...
+]
+```
+
+The trainer only requires `id`, `hard`, `soft`. The rest is preserved on
+`RolloutResult.extras` (see `skillopt/types.py`) and is what your
+`reflect()` consumes via `run_minibatch_reflect`.
+
+## Step 5 — Register the adapter
+
+Edit [`scripts/train.py`](https://github.com/microsoft/SkillOpt/blob/main/scripts/train.py)
+and add to `_register_builtins()`:
+
+```python
+    try:
+        from skillopt.envs.docfaithful.adapter import DocFaithfulAdapter
+        _ENV_REGISTRY["docfaithful"] = DocFaithfulAdapter
+    except ImportError:
+        pass  # docfaithful deps not installed — skip
+```
+
+There is **no `BENCHMARK_REGISTRY` dict in `skillopt/envs/__init__.py`** —
+the registry lives in `scripts/train.py` and is populated lazily so that
+optional deps don't break `--help`.
+
+## Step 6 — Create the YAML config
+
+`configs/docfaithful/default.yaml`:
+
+```yaml
+_base_: ../_base_/default.yaml      # NOTE: string, not list
+
+model:
+  reasoning_effort: medium
+
+train:
+  batch_size: 16
+  accumulation: 1
+  num_epochs: 4
+
+gradient:
+  minibatch_size: 8
+  merge_batch_size: 8
+
+optimizer:
+  learning_rate: 4
+
+env:
+  name: docfaithful
+  # Optional: a seed skill document. Create this file (or any markdown
+  # file) yourself before the first run, or omit the key to let SkillOpt
+  # start from an empty skill.
+  skill_init: skillopt/envs/docfaithful/skills/initial.md
+  split_mode: split_dir
+  split_dir: data/docfaithful_split
+  workers: 4
+  max_completion_tokens: 4096
+  limit: 0
+```
+
+> ⚠️ `_base_` is currently parsed as a **string path**, not a list. Write
+> `_base_: ../_base_/default.yaml`, not `_base_: ['../_base_/default.yaml']`.
+> See [`skillopt/config.py`](https://github.com/microsoft/SkillOpt/blob/main/skillopt/config.py)
+> if you want to add list-form inheritance.
+
+## Step 7 — Run
+
+```bash
+# If you set skill_init above, create the seed skill first:
+#   mkdir -p skillopt/envs/docfaithful/skills
+#   echo "# DocFaithful initial skill" > skillopt/envs/docfaithful/skills/initial.md
+
+python scripts/train.py --config configs/docfaithful/default.yaml
+```
+
+If you get `ValueError: Unknown environment 'docfaithful'. Available: [...]`,
+you forgot Step 5.
+
+If you get `TypeError: Can't instantiate abstract class DocFaithfulAdapter`,
+you forgot to implement one of the four abstract methods on `EnvAdapter`:
+`build_train_env`, `build_eval_env`, `rollout`, `get_task_types`.
+
+## Tips
+
+- Start with `train.batch_size: 4` and `limit: 10` while debugging.
+- The `evaluate` half lives **inside your `rollout`**, not as a separate
+  method — there is no `evaluate()` in the `EnvAdapter` ABC. Score the
+  prediction in `run_batch` and put the score on each result dict's
+  `hard` / `soft`.
+- Noisy scoring kills the optimizer. Spend time on `run_batch`'s scoring
+  before you spend time on prompts.
+- If your benchmark needs heavy optional deps (selenium, vllm, ...),
+  wrap the registration block with `try / except ImportError` (Step 5)
+  so people without those deps can still `--help`.
+- Copy `skillopt/envs/_template/` as a starting skeleton — it now
+  implements the real abstract methods.
@@ -0,0 +1,78 @@
+# Skill Document
+
+A **skill document** is a Markdown file that serves as the "prompt weights" of your agent. SkillOpt trains this document through iterative optimization.
+
+## What is a Skill Document?
+
+A skill document is a structured set of instructions that tells a language model **how** to approach a specific type of task. It's analogous to learned weights in a neural network — encoding task-specific knowledge in natural language rather than floating-point parameters.
+
+## Structure
+
+A typical skill document contains:
+
+```markdown
+# Task Strategy
+
+## General Approach
+- Break complex problems into sub-steps
+- Always verify intermediate results
+
+## Common Patterns
+- When you see X, try approach Y
+- Avoid Z because it leads to errors
+
+## Edge Cases
+- If the input contains A, handle it specially by...
+- Watch out for B — it requires C
+
+## Output Format
+- Always include reasoning before the answer
+- Format numbers with proper units
+```
+
+## How It Evolves
+
+During training, the skill document is modified by **edit patches**:
+
+1. **Additions**: New rules or strategies discovered from failed trajectories
+2. **Modifications**: Refining existing rules that are partially correct
+3. **Deletions**: Removing rules that consistently lead to errors
+
+Each edit is validated through the **gate** mechanism before being permanently accepted.
+
+## Initial Skill
+
+You can start training with:
+
+- **Empty skill**: The system learns everything from scratch
+- **Seed skill**: Provide initial instructions to bootstrap training
+- **Pre-trained skill**: Transfer a skill from a related benchmark
+
+Configure the initial skill in your YAML:
+
+```yaml
+train:
+  init_skill: "path/to/initial_skill.md"  # or omit for empty
+```
+
+## Skill Quality Metrics
+
+Track your skill's evolution through:
+
+- **Validation score**: Primary metric on the selection split
+- **Test score**: Final metric on held-out test data
+- **Skill length**: Total tokens in the document
+- **Edit acceptance rate**: Fraction of proposed edits that pass gating
+
+## Best Practices
+
+!!! tip "Tips for better skills"
+    1. **Start with a seed skill** (`env.skill_init`) if you have domain knowledge — it converges faster
+    2. **Use cosine LR schedule** — aggressive early exploration + careful late refinement
+    3. **Enable slow update** (`use_slow_update: true`) to prevent forgetting across epochs
+    4. **Enable meta skill** (`use_meta_skill: true`) so the optimizer accumulates strategy memory
+
+## Next Steps
+
+- [Deep Learning Analogy](dl-analogy.md)
+- [Configuration Reference](../reference/config.md)
@@ -0,0 +1,92 @@
+# The Training Loop
+
+SkillOpt's core insight: **optimizing natural-language skill documents follows the same structure as training neural networks**.
+
+## Overview
+
+```
+┌─────────────────────────────────────────────────────────┐
+│                    Training Loop                         │
+│                                                         │
+│  for epoch in epochs:                                   │
+│    for step in steps:                                   │
+│      1. Rollout   — Target executes tasks              │
+│      2. Reflect   — Optimizer analyzes trajectories       │
+│      3. Aggregate — Hierarchical merge of patches       │
+│      4. Select    — Rank & clip edits (learning rate)   │
+│      5. Update    — Apply patches to skill doc          │
+│      6. Gate      — Validate & accept/reject            │
+│                                                         │
+│    Epoch Boundary:                                       │
+│      • Slow Update (longitudinal comparison & guidance) │
+│      • Meta Skill  (cross-epoch strategy memory)        │
+└─────────────────────────────────────────────────────────┘
+```
+
+## Stage Details
+
+### 1. Rollout (Forward Pass)
+
+The **target** model executes tasks using the current skill document as its prompt. Each task produces a trajectory and a score.
+
+```python
+# Analogy: forward pass through the network
+predictions = model(input, skill_document)
+scores = evaluate(predictions, ground_truth)
+```
+
+### 2. Reflect (Backward Pass)
+
+The **optimizer** model analyzes failed trajectories and produces **edit patches** — structured suggestions for improving the skill document.
+
+Two modes:
+
+- **Shallow**: Analyze each trajectory independently
+- **Deep**: Cross-reference multiple failures to find systemic issues
+
+```python
+# Analogy: computing gradients
+gradients = loss.backward()  # → edit patches
+```
+
+### 3. Aggregate
+
+Semantically similar edit patches are merged to avoid redundant edits.
+
+### 4. Select (Gradient Clipping)
+
+Edits are ranked by relevance score. The `learning_rate` parameter caps how many edits are applied per step — just like gradient clipping prevents overshooting.
+
+```python
+# Analogy: gradient clipping + optimizer step size
+selected = top_k(edits, k=learning_rate)
+```
+
+The `lr_scheduler` adjusts this over training:
+
+- **cosine**: Start aggressive, taper smoothly
+- **linear**: Linear decay
+- **constant**: Fixed rate
+
+### 5. Update (Parameter Update)
+
+Selected edits are applied to the skill document, producing a new version.
+
+### 6. Gate (Validation)
+
+The updated skill is evaluated on a **selection split** (analogous to a validation set). The update is only accepted if performance improves.
+
+## Epoch Boundary Mechanisms
+
+### Slow Update
+
+At the end of each epoch (starting from epoch 2), the system performs a **longitudinal comparison**: it rolls out both the previous epoch's skill and the current skill on the same samples, categorizes items as improved/regressed/persistent_fail/stable_success, then generates high-level **guidance** that is injected into the skill document. This prevents catastrophic forgetting of earlier improvements.
+
+### Meta Skill
+
+A **meta-skill memory** accumulates high-level strategy notes across the entire training run. At the end of each epoch, the optimizer reflects on what changed between epochs and produces a compact memory that is provided as additional context during future reflection steps.
+
+## Next Steps
+
+- [Understand Skill Documents](skill-document.md)
+- [DL ↔ SkillOpt analogy table](dl-analogy.md)
@@ -0,0 +1,170 @@
+---
+hide:
+  - navigation
+---
+
+<div class="hero" markdown>
+
+# SkillOpt
+
+### Train Agent Skills Like Neural Networks
+
+*Optimize natural-language skill documents through iterative rollout, reflection, and gated validation — with epochs, learning rates, and validation gates — without touching model weights.*
+
+[Get Started :material-rocket-launch:](guide/installation.md){ .md-button .md-button--primary }
+[View on GitHub :material-github:](https://github.com/microsoft/SkillOpt){ .md-button }
+
+</div>
+
+---
+
+## How It Works
+
+<div class="pipeline-container" markdown>
+<div class="pipeline-wrapper">
+
+<div class="pipeline-stage" id="stage-rollout">
+<div class="stage-icon">🎯</div>
+<div class="stage-label">Rollout</div>
+<div class="stage-desc">Target executes tasks</div>
+</div>
+
+<div class="pipeline-arrow"><div class="flow-line"></div></div>
+
+<div class="pipeline-stage" id="stage-reflect">
+<div class="stage-icon">🔍</div>
+<div class="stage-label">Reflect</div>
+<div class="stage-desc">Optimizer analyzes trajectories</div>
+</div>
+
+<div class="pipeline-arrow"><div class="flow-line"></div></div>
+
+<div class="pipeline-stage" id="stage-aggregate">
+<div class="stage-icon">🔗</div>
+<div class="stage-label">Aggregate</div>
+<div class="stage-desc">Merge edit patches</div>
+</div>
+
+<div class="pipeline-arrow"><div class="flow-line"></div></div>
+
+<div class="pipeline-stage" id="stage-select">
+<div class="stage-icon">✂️</div>
+<div class="stage-label">Select</div>
+<div class="stage-desc">Rank & clip edits</div>
+</div>
+
+<div class="pipeline-arrow"><div class="flow-line"></div></div>
+
+<div class="pipeline-stage" id="stage-update">
+<div class="stage-icon">📝</div>
+<div class="stage-label">Update</div>
+<div class="stage-desc">Apply to skill doc</div>
+</div>
+
+<div class="pipeline-arrow"><div class="flow-line"></div></div>
+
+<div class="pipeline-stage" id="stage-gate">
+<div class="stage-icon">🚦</div>
+<div class="stage-label">Gate</div>
+<div class="stage-desc">Validate & accept</div>
+</div>
+
+</div>
+
+<div class="pipeline-epoch-bar">
+<div class="epoch-mechanism">🔄 Slow Update</div>
+<div class="epoch-mechanism">🧠 Meta Skill</div>
+<div class="epoch-label">Epoch Boundary</div>
+</div>
+
+</div>
+
+---
+
+## Deep Learning Analogy
+
+SkillOpt brings the familiar deep-learning training paradigm to agentic prompt optimization:
+
+| Deep Learning | SkillOpt | 
+|---|---|
+| Model weights | Skill document (Markdown) |
+| Forward pass | Rollout (target executes tasks) |
+| Loss / gradient | Reflect (optimizer produces edit patches) |
+| Gradient clipping | Edit selection (`learning_rate` = max edits) |
+| SGD step | Patch application to skill |
+| Validation set | Gated evaluation on selection split |
+| LR schedule | `lr_scheduler`: cosine, linear, constant |
+| Epochs | Multi-epoch with slow update & meta skill memory |
+
+---
+
+## Supported Benchmarks
+
+| Benchmark | Type | Config |
+|---|---|---|
+| **DocVQA** | Document QA | `configs/docvqa/` |
+| **ALFWorld** | Embodied AI | `configs/alfworld/` |
+| **OfficeQA** | Enterprise QA | `configs/officeqa/` |
+| **SearchQA** | Open-domain QA | `configs/searchqa/` |
+| **LiveMathBench** | Math reasoning | `configs/livemathematicianbench/` |
+| **SWEBench** | Software Engineering | `configs/swebench/` |
+| + 5 more | Various | See [docs](guide/first-experiment.md) |
+
+---
+
+## Quick Example
+
+```bash
+# Install
+pip install -e .
+
+# Configure credentials
+export AZURE_OPENAI_ENDPOINT="https://your-resource.openai.azure.com/"
+export AZURE_OPENAI_API_KEY="your-key"
+
+# Train on SearchQA
+python scripts/train.py --config configs/searchqa/default.yaml
+
+# Evaluate best skill
+python scripts/eval_only.py \
+  --config configs/searchqa/default.yaml \
+  --skill outputs/best_skill.md
+```
+
+---
+
+<div class="grid cards" markdown>
+
+-   :material-book-open-variant:{ .lg .middle } **Getting Started**
+
+    ---
+
+    Install SkillOpt, configure your API keys, and run your first experiment in 5 minutes.
+
+    [:octicons-arrow-right-24: Installation](guide/installation.md)
+
+-   :material-puzzle:{ .lg .middle } **Add a Benchmark**
+
+    ---
+
+    Extend SkillOpt with your own benchmark in ~100 lines of code.
+
+    [:octicons-arrow-right-24: Extension Guide](guide/new-benchmark.md)
+
+-   :material-cog:{ .lg .middle } **Configuration**
+
+    ---
+
+    Full reference for all hyperparameters with deep learning analogies.
+
+    [:octicons-arrow-right-24: Config Reference](reference/config.md)
+
+-   :material-monitor-dashboard:{ .lg .middle } **WebUI**
+
+    ---
+
+    Configure, launch, and monitor training from your browser.
+
+    [:octicons-arrow-right-24: WebUI Guide](guide/first-experiment.md#webui)
+
+</div>
@@ -0,0 +1,195 @@
+# API Reference
+
+This page documents the public Python API SkillOpt exposes for **extending the
+framework** with new environments / benchmarks. For ready-made adapters,
+browse [`skillopt/envs/`](https://github.com/microsoft/SkillOpt/tree/main/skillopt/envs).
+
+> **Source of truth.** The classes below are real Python ABCs defined in
+> `skillopt/envs/base.py`, `skillopt/datasets/base.py`, `skillopt/types.py`,
+> and `skillopt/evaluation/gate.py`. If this page ever drifts, the code
+> wins — please open an issue.
+
+---
+
+## Core Classes
+
+### `EnvAdapter`
+
+`skillopt/envs/base.py` — abstract adapter that connects the SkillOpt
+trainer to an environment (benchmark, simulator, REST API, ...).
+Subclasses **must** implement the five abstract methods below.
+
+```python
+from abc import ABC, abstractmethod
+from skillopt.datasets.base import BaseDataLoader, BatchSpec
+
+class EnvAdapter(ABC):
+
+    # ── Lifecycle hooks (have defaults; override only if needed) ────────
+
+    def setup(self, cfg: dict) -> None: ...
+    def get_dataloader(self) -> BaseDataLoader | None: ...
+    def requires_ray(self) -> bool: ...                 # default False
+
+    # ── Abstract methods (subclasses MUST implement) ────────────────────
+
+    @abstractmethod
+    def build_train_env(self, batch_size: int, seed: int, **kwargs):
+        """Return an environment-manager object to be passed to rollout()."""
+
+    @abstractmethod
+    def build_eval_env(self, env_num: int, split: str, seed: int, **kwargs):
+        """Like build_train_env() but for a fixed eval split."""
+
+    @abstractmethod
+    def rollout(self, env_manager, skill_content: str,
+                out_dir: str, **kwargs) -> list[dict]:
+        """Run a batch of episodes with the current skill.
+
+        Each returned dict MUST contain:
+          - "id":   str        episode/task identifier
+          - "hard": int (0|1)  pass/fail (may be float 0.0-1.0 if smoothed)
+          - "soft": float      partial-credit score in [0.0, 1.0]
+        It MAY contain env-specific extra keys (parsed into RolloutResult.extras).
+        """
+
+    @abstractmethod
+    def reflect(self, results: list[dict], skill_content: str,
+                out_dir: str, **kwargs) -> list[dict | None]:
+        """Turn rollout results into a list of raw patch dicts.
+
+        Each dict (or None to drop the slot) MUST contain:
+          - "patch":       {"edits": [...]}     a Patch.to_dict() payload
+          - "source_type": "failure" | "success"
+        """
+
+    @abstractmethod
+    def get_task_types(self) -> list[str]:
+        """Distinct task-type strings used for stratified sampling."""
+```
+
+The trainer also calls a few default-implemented helpers on every adapter:
+`build_reference_text`, `get_reference_metadata`, `attach_reference_context`,
+`select_representative_items`, and `build_env_from_batch`. Read the docstrings
+in `skillopt/envs/base.py` if you need to override any of these — most
+benchmarks don't.
+
+### `BaseDataLoader` / `SplitDataLoader`
+
+`skillopt/datasets/base.py` — episode-planning loaders.
+
+```python
+class BaseDataLoader(ABC):
+    def setup(self, cfg: dict) -> None: ...
+    @abstractmethod
+    def build_train_batch(self, batch_size: int, seed: int, **kwargs) -> BatchSpec: ...
+    @abstractmethod
+    def build_eval_batch(self, env_num: int, split: str, seed: int, **kwargs) -> BatchSpec: ...
+
+class SplitDataLoader(BaseDataLoader):
+    """Concrete base for dataset-backed envs with on-disk train/val/test splits.
+
+    Subclasses only need to implement load_split_items() (and optionally
+    load_raw_items() if you also want ``split_mode='ratio'``).
+    """
+    def load_split_items(self, split_path: str) -> list[dict]: ...
+    def load_raw_items(self, data_path: str) -> list[dict]: ...   # optional
+```
+
+`SplitDataLoader` handles two layout modes:
+
+| `split_mode` | What it expects |
+|---|---|
+| `"split_dir"` | A directory with `train/`, `val/`, `test/` subdirs already split. |
+| `"ratio"` | A raw dataset path + `split_ratio: "2:1:7"` style string. |
+
+In either case the items returned by `load_split_items()` are plain
+`dict` objects with at minimum an `"id"` key.
+
+### `BatchSpec`
+
+`skillopt/datasets/base.py` — a slotted dataclass describing one batch
+request the trainer hands to the adapter.
+
+```python
+@dataclass(slots=True)
+class BatchSpec:
+    phase: str                 # "train" | "eval"
+    split: str                 # "train" | "val" | "test" | "valid_seen" | ...
+    seed: int
+    batch_size: int
+    payload: object | None = None     # what the loader produced (e.g. list[dict])
+    metadata: dict = field(default_factory=dict)
+```
+
+### `Edit` / `Patch`
+
+`skillopt/types.py` — the I/O types Reflect / Aggregate / Update produce
+and consume.
+
+```python
+EditOp = Literal["append", "insert_after", "replace", "delete"]
+
+@dataclass
+class Edit:
+    op: EditOp
+    content: str = ""
+    target: str = ""
+    support_count: int | None = None
+    source_type: Literal["failure", "success"] | None = None
+    merge_level: int | None = None
+    update_origin: str = ""
+    update_target: str = ""
+
+@dataclass
+class Patch:
+    edits: list[Edit] = field(default_factory=list)
+    reasoning: str = ""
+    ranking_details: dict[str, Any] | None = None
+```
+
+Both types support `to_dict()` / `from_dict()` for serialization.
+
+### `RolloutResult`
+
+`skillopt/types.py` — the normalised rollout return type. The trainer
+calls `RolloutResult.from_dict(...)` on each dict returned from
+`EnvAdapter.rollout()`, so the only **hard** requirement on those dicts is
+the three keys above (`id`, `hard`, `soft`). Extra fields are preserved
+into `RolloutResult.extras`.
+
+### `GateResult` / `GateAction`
+
+`skillopt/evaluation/gate.py` — the validation-gate decision types
+returned each epoch.
+
+---
+
+## Registering an environment
+
+Environments are not registered via decorators or a `BENCHMARK_REGISTRY`
+dict. The trainer keeps a lazy registry inside `scripts/train.py` —
+`_ENV_REGISTRY` — populated by `_register_builtins()`. To add a new env
+you append a `try / except ImportError` block there. See
+[Add a New Benchmark](../guide/new-benchmark.md) for the full step-by-step.
+
+---
+
+## Backends (model layer)
+
+The model layer lives under `skillopt.model.*`. Backends are selected
+via `model.optimizer_backend` and `model.target_backend` in the config —
+not via a base class subclass. Supported values (as of this writing):
+
+| Backend | Optimizer? | Target? |
+|---|---|---|
+| `openai_chat` | ✓ | ✓ |
+| `claude_chat` | ✓ | ✓ |
+| `qwen_chat` | ✓ | ✓ |
+| `minimax_chat` | ✓ | ✓ |
+| `codex_exec` | — | ✓ |
+| `claude_code_exec` | — | ✓ |
+
+See `skillopt/model/backend_config.py` for the live whitelist and
+[`docs/reference/config.md`](./config.md) for the per-backend
+configuration keys.
@@ -0,0 +1,71 @@
+# CLI Reference
+
+## Training
+
+```bash
+python scripts/train.py --config <config.yaml> [overrides...]
+```
+
+### Arguments
+
+| Argument | Description |
+|---|---|
+| `--config` | Path to YAML config file (required) |
+| `key=value` | Override any config parameter |
+
+### Examples
+
+```bash
+# Basic training
+python scripts/train.py --config configs/searchqa/default.yaml
+
+# With overrides
+python scripts/train.py \
+  --config configs/searchqa/default.yaml \
+  --cfg-options optimizer.learning_rate=16 optimizer.lr_scheduler=linear
+
+# With custom initial skill
+python scripts/train.py \
+  --config configs/searchqa/default.yaml \
+  --cfg-options env.skill_init=skills/my_seed.md
+```
+
+## Evaluation
+
+```bash
+python scripts/eval_only.py --config <config.yaml> --skill <skill.md>
+```
+
+### Arguments
+
+| Argument | Description |
+|---|---|
+| `--config` | Path to YAML config file (required) |
+| `--skill` | Path to skill document to evaluate (required) |
+| `--split` | Evaluation split: `test` (default), `valid`, `train` |
+
+### Examples
+
+```bash
+# Evaluate best skill on test set
+python scripts/eval_only.py \
+  --config configs/searchqa/default.yaml \
+  --skill outputs/searchqa/run_001/skills/best_skill.md
+
+# Evaluate on validation set
+python scripts/eval_only.py \
+  --config configs/searchqa/default.yaml \
+  --skill outputs/searchqa/run_001/skills/best_skill.md \
+  --split valid
+```
+
+## WebUI
+
+```bash
+python -m skillopt_webui.app [--port PORT] [--share]
+```
+
+| Argument | Default | Description |
+|---|---|---|
+| `--port` | 7860 | Port number |
+| `--share` | false | Create public Gradio link |
@@ -0,0 +1,85 @@
+# Configuration Reference
+
+Complete reference for all SkillOpt configuration parameters.
+
+## Model
+
+| Parameter | Type | Default | Description |
+|---|---|---|---|
+| `model.backend` | str | `azure_openai` | Backend: `azure_openai` / `openai_chat` / `claude_code_exec` / `qwen` |
+| `model.optimizer` | str | `gpt-5.5` | Optimizer model (for reflection & slow update) |
+| `model.target` | str | `gpt-5.5` | Target model (for rollout execution) |
+| `model.reasoning_effort` | str | `medium` | Reasoning effort level |
+| `model.optimizer_backend` | str | `openai_chat` | Optimizer backend: `openai_chat` / `claude_chat` / `qwen_chat` / `minimax_chat` |
+| `model.target_backend` | str | `openai_chat` | Target backend: chat backends plus execution harnesses |
+| `model.qwen_chat_base_url` | str | `http://localhost:8000/v1` | Shared Qwen/vLLM OpenAI-compatible endpoint |
+| `model.qwen_chat_enable_thinking` | bool | `false` | Shared Qwen thinking flag |
+| `model.optimizer_qwen_chat_base_url` | str | — | Optimizer-specific Qwen/vLLM endpoint; overrides shared `qwen_chat_base_url` |
+| `model.target_qwen_chat_base_url` | str | — | Target-specific Qwen/vLLM endpoint; overrides shared `qwen_chat_base_url` |
+
+## Training (`train`)
+
+| Parameter | Type | Default | DL Analogy | Description |
+|---|---|---|---|---|
+| `train.num_epochs` | int | 4 | Epochs | Number of training epochs |
+| `train.batch_size` | int | 40 | Batch size | Tasks sampled per step |
+| `train.accumulation` | int | 1 | Gradient accumulation | Accumulation rounds per step |
+| `train.seed` | int | 42 | Random seed | Reproducibility seed |
+
+## Gradient / Reflection (`gradient`)
+
+| Parameter | Type | Default | Description |
+|---|---|---|---|
+| `gradient.minibatch_size` | int | 8 | Reflect minibatch size |
+| `gradient.merge_batch_size` | int | 8 | Patch merge batch size |
+| `gradient.analyst_workers` | int | 16 | Parallel reflection workers |
+| `gradient.max_analyst_rounds` | int | 3 | Max rounds of analyst reflection |
+| `gradient.failure_only` | bool | `false` | Only reflect on failures |
+
+## Optimizer (`optimizer`)
+
+| Parameter | Type | Default | DL Analogy | Description |
+|---|---|---|---|---|
+| `optimizer.learning_rate` | int | 4 | Learning rate | Max edit patches per step (edit budget) |
+| `optimizer.min_learning_rate` | int | 2 | Min LR | Min edits for decay schedulers |
+| `optimizer.lr_scheduler` | str | `cosine` | LR schedule | `constant` / `linear` / `cosine` / `autonomous` |
+| `optimizer.skill_update_mode` | str | `patch` | — | `patch` / `rewrite_from_suggestions` / `full_rewrite_minibatch` |
+| `optimizer.use_slow_update` | bool | `true` | Momentum | Epoch-boundary longitudinal comparison & guidance |
+| `optimizer.slow_update_samples` | int | 20 | — | Samples for slow update evaluation |
+| `optimizer.use_meta_skill` | bool | `true` | Meta-learning | Cross-epoch optimizer-side strategy memory |
+| `optimizer.longitudinal_pair_policy` | str | `mixed` | — | `mixed` / `changed` / `unchanged` |
+
+## Evaluation (`evaluation`)
+
+| Parameter | Type | Default | Description |
+|---|---|---|---|
+| `evaluation.use_gate` | bool | `true` | Enable validation gating (accept/reject updates) |
+| `evaluation.eval_test` | bool | `true` | Run test evaluation after training |
+
+## Environment (`env`)
+
+| Parameter | Type | Default | Description |
+|---|---|---|---|
+| `env.name` | str | — | Benchmark name (e.g., `searchqa`, `docvqa`) |
+| `env.data_path` | str | — | Path to dataset |
+| `env.skill_init` | str | — | Path to initial seed skill (optional) |
+| `env.split_mode` | str | `ratio` | `ratio` or `split_dir` |
+| `env.split_ratio` | str | `2:1:7` | Train:val:test ratio |
+| `env.exec_timeout` | int | 120 | Per-task timeout in seconds |
+| `env.out_root` | str | — | Output directory |
+
+## Azure OpenAI Credentials
+
+| Variable | Description |
+|---|---|
+| `AZURE_OPENAI_ENDPOINT` / `model.azure_openai_endpoint` | Azure resource endpoint |
+| `AZURE_OPENAI_API_KEY` / `model.azure_openai_api_key` | Azure API key |
+| `OPENAI_API_KEY` | OpenAI API key (for `openai_chat` backend) |
+| `ANTHROPIC_API_KEY` | Anthropic API key (for `claude_code_exec` backend) |
+| `QWEN_CHAT_BASE_URL` | Shared local vLLM endpoint for `qwen_chat` |
+| `QWEN_CHAT_MODEL` | Shared served model name for `qwen_chat` |
+| `QWEN_CHAT_API_KEY` | Optional API key for the shared Qwen endpoint |
+| `OPTIMIZER_QWEN_CHAT_BASE_URL` | Optimizer-specific local vLLM endpoint |
+| `OPTIMIZER_QWEN_CHAT_MODEL` | Optimizer-specific served model name |
+| `TARGET_QWEN_CHAT_BASE_URL` | Target-specific local vLLM endpoint |
+| `TARGET_QWEN_CHAT_MODEL` | Target-specific served model name |
@@ -0,0 +1,110 @@
+# SkillOpt-Sleep 😴 — deployment-time companion (preview)
+
+**SkillOpt-Sleep** applies SkillOpt's discipline to your *own daily usage*. It gives a
+local coding agent a nightly **sleep cycle** that reviews your past sessions, replays
+your recurring tasks on your own API budget, and consolidates what it learns into
+**validated** long-term memory and skills — behind a held-out gate, staged for your
+review. The agent gets better the more you use it, with **no weight training** and
+**zero inference-time overhead**.
+
+> **Preview.** This is an early preview we are actively iterating on; interfaces and
+> defaults may change. The engine lives in the top-level [`skillopt_sleep/`](../../skillopt_sleep)
+> package with **zero dependency** on the paper's `skillopt/` code (the validation gate
+> is vendored).
+
+## How it works
+
+One "night":
+
+```
+harvest Claude Code / Codex transcripts → mine recurring tasks → replay offline
+   → consolidate (reflect → bounded edit → GATE on real held-out tasks)
+   → stage proposal → (you) adopt
+```
+
+It synthesizes **SkillOpt** (validation-gated bounded text edits), **Claude Dreams**
+(offline consolidation; review-then-adopt), and the **agent-sleep** idea (short-term
+experience → long-term competence).
+
+## How to use it
+
+### Quickest path: the `skillopt-sleep` CLI (pip)
+
+```bash
+pip install skillopt        # installs the engine + the `skillopt-sleep` command
+skillopt-sleep dry-run      # harvest + mine + replay, report only (changes nothing)
+skillopt-sleep run          # a full nightly cycle; the proposal is staged for review
+skillopt-sleep status       # show state + the latest staged proposal
+skillopt-sleep adopt        # apply the latest staged proposal
+skillopt-sleep schedule     # install a nightly cron entry for this project
+```
+
+The per-agent plugin shells below (Claude Code / Codex / Copilot) still come from the
+repo; the CLI above is the standalone, pip-only way to run a cycle.
+
+One engine, thin per-agent shells (see [`plugins/`](../../plugins)):
+
+| Platform | Folder | Install |
+|---|---|---|
+| **Claude Code** | [`plugins/claude-code`](../../plugins/claude-code) | `/plugin marketplace add ./plugins/claude-code` → `/skillopt-sleep` |
+| **Codex** | [`plugins/codex`](../../plugins/codex) | `bash plugins/codex/install.sh` → `skillopt-sleep` skill |
+| **Copilot** | [`plugins/copilot`](../../plugins/copilot) | register `plugins/copilot/mcp_server.py` as an MCP server |
+
+Deterministic proof (no API key):
+`python -m skillopt_sleep.experiments.run_experiment --persona researcher --assert-improves`.
+
+### Opt-in: experience replay & dream rollouts
+
+Two consolidation mechanisms, both default **off** (behavior is unchanged unless you
+enable them). They strengthen the nightly update when your tasks have a clean
+correctness signal; the validation gate still governs what ships.
+
+| Config knob | Default | Effect |
+|---|---|---|
+| `dream_rollouts` | `1` | Run each task K times → learn from the good-vs-bad contrast (contrastive reflection). |
+| `recall_k` | `0` | Associative recall — pull the K most-similar past tasks (from a persisted archive) into tonight's dream. |
+| `dream_factor` | `0` | Add N lightweight synthetic variants of each task. |
+
+## Results
+
+> 📊 **More results & analysis — the gate-safety stress test, experience-replay
+> scaling, and the dream-diversity ablation — are in
+> [`docs/sleep/RESULTS.md`](RESULTS.md).** The highlights:
+
+**Protocol (identical for every row below).** 5 nights × 10 new real "today" tasks
+per night; the full held-out **test** split is scored before night 1 (baseline) and
+after night 5 (after); optimizer = GPT-5.5; single seed (42); run through the exact
+shipped engine (`skillopt_sleep.dream.dream_consolidate`). Numbers are absolute
+held-out accuracy; **Δ** = `after − baseline` in percentage points.
+
+**(a) End-to-end on real agents — [gbrain-evals](https://github.com/garrytan/gbrain-evals) `skillopt-v1`.**
+Deficient seed skills go **0.00 → 1.00** on the held-out set with **both Claude Code
+and Codex** as the target agent (all 4 seeds, including a real tool-use loop).
+
+**(b) Experience replay scales the gain — SearchQA** (1,400-item held-out test,
+SQuAD exact-match; target = GPT-5.5; **validation-gated**):
+
+| Replay config (`dream_rollouts=5`) | Baseline → After | Δ (pts) |
+|---|---|---|
+| `recall_k=10` | 0.802 → 0.834 | +3.1 |
+| `recall_k=20` | 0.803 → 0.848 | **+4.5** |
+| full-history replay *(reference, not a shipping default)* | 0.796 → 0.851 | +5.6 |
+| `recall_k=10`, `dream_rollouts=8` *(more dreaming, same recall)* | 0.798 → 0.835 | +3.7 |
+
+The gain rises monotonically with how much relevant past experience is recalled. The
+same SearchQA cell **without** the gate (`recall_k=10`) is 0.808 → 0.839 (+3.1).
+
+**(c) Second benchmark — SpreadsheetBench** (280-item held-out test; the agent's
+generated openpyxl code is executed and compared cell-by-cell to a golden workbook;
+target = GPT-5.4-nano; gate-free + the output-contract guardrail): 0.279 → 0.314 (**+3.6**).
+
+**(d) Honest scope.** These gains hold where tasks recur and have a checkable
+correctness signal. On saturated or noisy benchmarks (e.g. a strong model already
+near ceiling) the effect is **flat within run-to-run noise** — single-seed baseline
+variance here is ±1–2 pts, so treat sub-~1.5 pt differences as noise. The validation
+gate keeps the worst case bounded; keep it **on** by default.
+
+## Learn more
+
+Full reference (pipeline, the three plugins, the experience-replay knobs) is in the
+**[Documentation & Reproduction Guide](https://microsoft.github.io/SkillOpt/docs/guideline.html#sleep)**.
@@ -0,0 +1,185 @@
+# SkillOpt-Sleep — results & analysis
+
+This is the evidence behind SkillOpt-Sleep: does a nightly, offline sleep cycle
+actually make a *deployed* agent better, and is it safe to run unattended? We
+answer with a controlled deployment-scale study — the same protocol the plugin
+runs in production, scored on full held-out test sets.
+
+## Setup
+
+**Protocol (identical for every cell unless stated).** 5 nights; each night adds
+**10 new real "today" tasks**; the skill carries over and is refined night to
+night. The full held-out **test** split is scored before night 1 (*baseline*) and
+after night 5 (*after*); **Δ = after − baseline** in percentage points. Optimizer
+model = **GPT-5.5**; single seed (42); every number is produced by the exact
+shipped engine `skillopt_sleep.dream.dream_consolidate` (the experiment harness and
+the plugin cycle call the same function).
+
+**Benchmarks** (real evaluators, not format heuristics):
+
+| Benchmark | Held-out test | Scoring |
+|---|---|---|
+| SearchQA | 1,400 items | SQuAD exact-match vs gold |
+| LiveMathematicianBench | 124 items | multiple-choice label (choices shuffled per item) |
+| SpreadsheetBench | 280 items | the agent's generated openpyxl code is **executed**, output workbook compared cell-by-cell to a golden file |
+
+**Targets:** GPT-5.5, GPT-5.4-mini, GPT-5.4-nano. **Modes:** validation-gated
+(default) and gate-free.
+
+---
+
+## 1. The headline — the validation gate is what makes nightly self-evolution *safe*
+
+Self-evolution is easy to build and easy to ruin: an optimizer that accepts its
+own "lessons" unconditionally can adopt a plausible-but-wrong rule and an obedient
+model will follow it off a cliff. We reproduced exactly that failure, then showed
+the gate prevents it.
+
+Stress case — **GPT-5.4-nano on SearchQA**, weak model on a single-sample (degraded)
+reflection signal, same nights, same candidate edits, gate **off** vs **on**:
+
+| | Night 0 → Night 5 | Δ |
+|---|---|---|
+| **no gate** | 0.554 → **0.026** | **−52.8** |
+| **with gate (default)** | 0.570 → 0.570 | 0.0 |
+
+Ungated, the optimizer learned "answer with the document-title string, verbatim";
+the model complied and accuracy collapsed night after night
+(0.554 → 0.490 → 0.325 → 0.031 → 0.034 → 0.026). The gated twin **rejected every one
+of those edits** and never lost a point. This single experiment is the core
+argument for SkillOpt-Sleep's design, and why the gate ships **on by default**.
+
+---
+
+## 2. Cross-model scaling — bigger gains where there's headroom
+
+The same protocol on a weaker target model (**GPT-5.4-nano**, optimizer = GPT-5.5)
+produces substantially larger gains — because the weaker model has more room to
+learn. This is the realistic "cheap deployed agent, strong overnight optimizer"
+scenario:
+
+| Config (SearchQA, nano, gated) | Baseline → After | Δ | Night-by-night |
+|---|---|---|---|
+| **cumulative replay, nights=5** | 0.560 → **0.679** | **+11.9** | 0.560 → 0.626 → 0.665 → 0.665 → 0.665 → 0.679 |
+| recall_k=20, nights=5 | 0.566 → 0.681 | +11.5 | 0.566 → 0.659 → 0.685 → 0.685 → 0.681 → 0.681 |
+| cumulative, nights=8 | 0.562 → 0.657 | +9.5 | saturates after night 5 |
+
+Both replay strategies (cumulative and recall) agree within 0.4 pt — the gain is
+robust across configurations.
+
+**Compared to GPT-5.5 on the same benchmark (SearchQA, gated):**
+
+| Target model | Best Δ | Baseline | Headroom |
+|---|---|---|---|
+| GPT-5.4-nano | **+11.9** | 0.560 | 44 pt |
+| GPT-5.5 | +6.0 | 0.798 | 20 pt |
+
+The story: **SkillOpt-Sleep helps most where there's the most to learn** — weaker
+deployed models benefit ~2× as much from the same nightly optimization. This is
+also the economical deployment pattern (cheap inference model + one strong
+overnight optimizer call).
+
+---
+
+## 3. Experience replay turns a one-time bump into a climb
+
+The plugin's two opt-in knobs (`recall_k`, `dream_rollouts`) are what produce the
+gains. On **SearchQA, GPT-5.5, gated** — the gain rises monotonically with how
+much relevant past experience is recalled:
+
+| Replay (`dream_rollouts=5`) | Baseline → After | Δ |
+|---|---|---|
+| `recall_k=10` | 0.802 → 0.834 | +3.1 |
+| `recall_k=20` | 0.803 → 0.848 | **+4.5** |
+| full-history (reference, not a default) | 0.796 → 0.851 | +5.6 |
+
+And the curve genuinely **climbs across nights** rather than jumping once and
+plateauing — full-history replay, gated, night by night:
+
+```
+0.798 → 0.814 → 0.854 → 0.854 → 0.854 → 0.858
+```
+
+The gate accepts a new, better skill as late as **night 5** (0.854 → 0.858).
+Replay-policy ablation (SearchQA, GPT-5.5):
+
+| Replay policy | Gate-free Δ | Gated Δ |
+|---|---|---|
+| none (tonight's tasks only) | +3.9 | +2.0 |
+| **recall k=10 (shipped default-able)** | +5.1 | +4.4 |
+| cumulative (full history) | +4.8 | +6.0 |
+
+Recall captures most of cumulative's benefit at a fraction of the per-night cost.
+
+---
+
+## 4. Default hyperparameters are the sweet spot
+
+We swept `dream_factor`, `rollouts`, `per_night`, and `nights` on the nano cell
+(SearchQA, gated) to verify the shipped defaults are well-tuned:
+
+| Variant | Δ | vs default (+11.9) |
+|---|---|---|
+| dream_factor=4 (default 2) | +8.8 | −3.1 |
+| rollouts=10 (default 5) | +9.5 | −2.4 |
+| per_night=15 (default 10) | +2.7 | −9.2 |
+| nights=8 (default 5) | +9.5 | −2.4 |
+
+Every direction away from the default hurts. This means users get the best result
+**out of the box** without tuning — the recipe is robust by design.
+
+---
+
+## 5. Why these gains exist — the dream-diversity fix (and a rigor note)
+
+Reflection learns from the **contrast** between good and bad rollouts of the same
+task, which requires the K dream rollouts to be *independent samples*. An early
+version of the engine collapsed them to one cached sample, so contrastive
+reflection never fired. Fixing that, then adding recall, is what produces the
+gains in Sections 1–2. Measured across an 18-cell deployment sweep (3 benchmarks ×
+3 targets × 2 modes), under three engine configurations:
+
+| Engine configuration | mean Δ | worst-cell Δ | cells > +0.5 | cells < −0.5 |
+|---|---|---|---|---|
+| single-sample reflection (degraded) | −2.66 | **−52.8** | 7 / 18 | 5 / 18 |
+| diverse rollouts (K=5), no recall | +0.24 | −4.0 | 6 / 18 | 7 / 18 |
+| **diverse rollouts + recall (shipped)** | **+0.53** | **−2.4** | 7 / 18 | 7 / 18 |
+
+The catastrophic −52.8 is removed **at its source** by diverse rollouts: the same
+gate-free nano-SearchQA cell goes 0.554 → **0.586 (+2.7)** with no gate at all once
+the dream is fixed. Recall then lifts the grid mean and tightens the worst case.
+This is **defense in depth, each layer measured**: diverse rollouts propose better
+edits, recall remembers relevant experience, and the gate catches whatever still
+slips through.
+
+---
+
+## 6. End-to-end on real agents
+
+On the public [gbrain-evals](https://github.com/garrytan/gbrain-evals) `skillopt-v1`
+benchmark — designed for exactly this learnable-gap setting — deficient seed skills
+go **0.00 → 1.00** on the held-out set with **both Claude Code and Codex** as the
+target agent (all 4 seeds, including a real tool-use loop), and the two agents
+cross-verify each other's consolidated skills.
+
+---
+
+## 7. Honest scope & limitations
+
+- **Where it helps:** recurring tasks with a checkable correctness signal and real
+  headroom. That is the plugin's actual use case (your repeated daily tasks and
+  house rules the agent keeps missing).
+- **Where it's flat:** saturated tasks on strong models, or noisy tasks with a weak
+  learning signal — within run-to-run noise.
+- **Single seed.** Cells aggregate one seed per config; treat sub-~1.5 pt
+  differences as noise. Spot seed-robustness check on the one flagged cell
+  (nano SearchQA gated): seeds 42/43/44 give −1.9 / +3.6 / +4.7 (3-seed mean
+  **+2.1**), i.e. the tabled −1.9 is a pessimistic draw, not the typical outcome.
+- **Keep the gate on.** It is the difference between bounded downside (−2.4) and a
+  −52.8 collapse. Gate-free mode is for users who cannot hold out a validation set
+  and is additionally protected by the output-contract guardrail.
+
+---
+
+Back to the module overview: [`docs/sleep/README.md`](README.md) ·
+full reference: [Documentation & Reproduction Guide](https://microsoft.github.io/SkillOpt/docs/guideline.html#sleep).
@@ -0,0 +1,237 @@
+# SkillOpt Sleep — Claude Code self-evolving plugin (design)
+
+**Status:** approved-for-build (autonomous offline session, 2026-06-07)
+**Author:** generated for Yifan Yang, executed autonomously while user is asleep
+**Branch:** `feat/claude-code-sleep-plugin` (worktree `my_repo/SkillOpt-sleep`)
+
+---
+
+## 1. One-paragraph summary
+
+`skillopt-sleep` is a Claude Code plugin that gives a user's local Claude
+agent a nightly **sleep cycle**. While the user is offline, it (1) **harvests**
+the day's real Claude Code session transcripts from `~/.claude`, (2) **mines**
+them into discrete *task records* with checkable outcomes, (3) **replays /
+"dreams"** those tasks offline using the user's own API budget, and (4) runs
+the **SkillOpt optimizer loop** (reflect → bounded edit → held-out gate) to
+consolidate short-term experience into long-term **memory** (`CLAUDE.md`) and
+**skills** (`SKILL.md`). Only changes that pass a validation gate are kept, and
+every change is written to a **review staging area** the user approves before it
+touches live config — mirroring Claude Dream's "input store is never modified"
+safety contract. The result: an agent that measurably gets better at *this
+user's* recurring work, every night, with zero model-weight training.
+
+## 2. Why this is the right synthesis of the three ingredients
+
+| Ingredient | What we take from it | Where it lives in this design |
+|---|---|---|
+| **SkillOpt** (your paper/code) | Skill = trainable text state; bounded add/delete/replace edits under a textual learning rate; **held-out validation gate**; rejected-edit buffer; epoch-wise slow/meta update. | The `consolidate` stage *is* a single SkillOpt epoch, reusing `skillopt.optimizer.*` and `skillopt.evaluation.gate`. |
+| **Claude Dreams** | Async offline job: read a memory store + 1–100 session transcripts → emit a **new, separate** reorganized memory store (dedup / merge / resolve contradictions / surface insights). Input never mutated; output reviewed then adopted or discarded. | The `harvest` + `consolidate-memory` stages and the **staging/adopt** safety model are modeled directly on Dreams. |
+| **Agent Sleep paper** (2605.26099) | Agents need periodic offline consolidation: short-term experience buffer → synthetic replay/self-generated data → self-update; "sleep" turns episodes into durable competence. | The whole nightly schedule, the `replay` step, and the short-term→long-term framing. |
+
+The key novel claim this enables for the project (and a future paper section):
+**SkillOpt's validation-gated bounded-edit optimizer is the missing "safe
+update rule" for Dream-style memory consolidation.** Dreams reorganize memory
+but don't *prove* the reorganization helps; the Sleep paper consolidates but
+assumes weight updates. SkillOpt-Sleep consolidates **text** (memory + skills)
+and **gates each change on replayed task performance**, so nightly evolution is
+both weight-free and regression-protected.
+
+## 3. Goals / non-goals
+
+**Goals**
+1. A working Claude Code plugin: scheduled (nightly/cron) **and** user-triggered (`/sleep`).
+2. Look back over the user's real past prompts & trajectories from local `~/.claude` records.
+3. Offline "dream training": re-run mined tasks (mock-env or fresh retry) on the user's budget.
+4. Continuous evolution of **memory** (`CLAUDE.md`) and **skills** (`SKILL.md`) via the SkillOpt gate.
+5. A reproducible experiment that answers: *does the nightly loop actually improve a held-out score?*
+6. Safety: never silently overwrite user config; stage → user approves → adopt.
+
+**Non-goals (now)**
+- Codex version (explicitly deferred by user; architecture keeps it pluggable).
+- Anthropic managed Dreams API integration (we *emulate* Dreams locally; managed API is a future backend).
+- Model fine-tuning / weight updates (out of scope by design — text-only).
+- Fully unattended auto-adopt by default (opt-in; default is review-gated).
+
+## 4. The local data we read (verified on this machine)
+
+- **Prompt history:** `~/.claude/history.jsonl` — one JSON/line: `{display, pastedContents, timestamp, project}`. The cross-session list of every prompt the user typed, with project path + epoch-ms timestamp.
+- **Full transcripts:** `~/.claude/projects/<path-slug>/<sessionId>.jsonl` — one record/line. Record `type` ∈ {`user`,`assistant`,`mode`,`permission-mode`,`attachment`,`file-history-snapshot`,`last-prompt`,…}. User/assistant records carry `message` (role+content blocks), plus `cwd`, `gitBranch`, `timestamp`, `sessionId`, `version`, `userType`. ~215k transcripts present on this box.
+- **Deployment targets we may evolve:**
+  - Project memory: `<project>/CLAUDE.md` (and `~/.claude/CLAUDE.md` global).
+  - User skills: `~/.claude/skills/<name>/SKILL.md` (frontmatter: `name`, `description`, optional `allowed-tools`, `argument-hint`).
+  - Plugin skills under `~/.claude/plugins/...`.
+
+Everything stays **on-disk and local**; the only network calls are the LLM
+optimizer/replay calls the user already pays for.
+
+## 5. Architecture
+
+### 5.1 The nightly Sleep Cycle (stages)
+
+```
+            ┌────────────────────────── SLEEP CYCLE (one "night") ──────────────────────────┐
+            │                                                                                │
+ trigger →  │  1.HARVEST     2.MINE          3.REPLAY            4.CONSOLIDATE      5.STAGE   │ → wake report
+ (cron or   │  read ~/.claude scan sessions  re-run tasks        SkillOpt epoch:   write to   │
+  /sleep)   │  transcripts → → task records   offline (mock or    reflect→edit→     .skillopt-│
+            │  + history     w/ outcomes &    fresh retry) under  GATE on held-out  sleep/    │
+            │                checkable refs   current skill/mem    replay split      staging/ │
+            │                                                                          ↓      │
+            │                                              6.ADOPT (opt-in / user-approved)   │
+            └────────────────────────────────────────────────────────────────────────────────┘
+```
+
+**1. Harvest** (`harvest.py`)
+Read `history.jsonl` + per-project transcript JSONLs for a time window
+(default: since last sleep, fallback last 24–72h). Group by project (`cwd` /
+`project`). Emit normalized `SessionDigest` objects: ordered user prompts,
+assistant final texts, tool-call summary, files touched (from
+`file-history-snapshot`), git branch, errors seen, and **user-feedback signals**
+(e.g. "still broken", "that's wrong", "perfect", re-asks of the same thing).
+
+**2. Mine** (`mine.py`)
+Turn digests into `TaskRecord`s — the unit the optimizer trains on. A task is a
+self-contained intent (the user's request) plus an *outcome label* and, where
+possible, a **checkable reference**:
+- *Explicit success/failure* from feedback signals ("works now" after N retries → the early attempts are failures, the fix is the success exemplar).
+- *Self-consistency check*: re-derivable answers (math, lookups) get a reference; open-ended ones get an LLM-judge rubric instead.
+- Each TaskRecord: `{id, project, intent, context_excerpt, attempted_solution, outcome ∈ {success,fail,mixed}, reference_kind ∈ {exact, rubric, none}, reference, tags}`.
+Mining is itself an LLM call (the **miner**), prompt-tunable, with a deterministic regex/heuristic fallback for offline/no-key runs.
+
+**3. Replay / "Dream"** (`replay.py`)
+For mined tasks, re-run the intent **offline** under the *current* skill+memory
+to get a fresh trajectory & score. Two modes:
+- `mock` (default, safe): reconstruct a sandboxed prompt from the task's captured context (no live repo mutation, no network side effects) and run the target model. Deterministic, cheap, safe to run unattended.
+- `fresh` (opt-in): actually re-attempt in a throwaway git worktree of the project. Higher fidelity, heavier, never touches the user's working tree.
+Scoring: exact-match / substring for `exact` refs; LLM-judge (0–1) for `rubric` refs; this yields the `hard`/`soft` scores SkillOpt already expects.
+
+**4. Consolidate** (`consolidate.py`) — *this is one SkillOpt epoch*
+Reuse the existing optimizer pieces rather than reinventing:
+- `reflect`: partition replayed tasks into failure/success minibatches → propose add/delete/replace edits to **skill** and a parallel proposer for **memory** (`CLAUDE.md`). (Memory consolidation also does Dream-style dedup/merge/contradiction-resolution over existing `CLAUDE.md` lines.)
+- `aggregate` + `rank_and_select` under an **edit budget** (textual learning rate).
+- `apply_patch_with_report` → candidate skill / candidate memory.
+- **GATE** (`skillopt.evaluation.gate.evaluate_gate`): replay a *held-out* slice of tasks with the candidate; accept only if it strictly beats current. Rejected edits go to the rejected-edit buffer (negative feedback) exactly as in the paper.
+- A **slow/meta** pass across nights (not just within one night) carries durable, cross-session lessons — the literal "short-term experience → long-term knowledge" of the Sleep paper. Per-night state persists in `~/.skillopt-sleep/state.json`.
+
+**5. Stage** (`staging/`)
+Write `proposed_CLAUDE.md`, `proposed_SKILL.md`, a unified diff, and a
+`sleep_report.md` (what changed, why, gate deltas, token cost) into
+`<project>/.skillopt-sleep/staging/<date>/`. **Nothing live is modified.**
+
+**6. Adopt**
+`/sleep adopt` (or `auto_adopt: true` in config for power users) copies staged
+files over the live `CLAUDE.md` / `SKILL.md`, after a `git`-style backup. This
+is the only stage that mutates user-facing config, and it is explicit by default
+— the Dreams "review the output, then adopt or discard" contract.
+
+### 5.2 Components & boundaries (each independently testable)
+
+```
+skillopt/sleep/
+  __init__.py
+  types.py         # SessionDigest, TaskRecord, ReplayResult, SleepConfig, SleepReport (dataclasses)
+  harvest.py       # ~/.claude transcripts + history.jsonl  ->  list[SessionDigest]
+  mine.py          # list[SessionDigest]  ->  list[TaskRecord]   (LLM miner + heuristic fallback)
+  replay.py        # TaskRecord + skill + memory  ->  ReplayResult (hard/soft)   (mock | fresh)
+  consolidate.py   # ReplayResults -> candidate skill+memory -> GATE -> accepted artifacts
+  memory.py        # CLAUDE.md read/merge/dedup/diff (Dream-style) + protected-region markers
+  state.py         # ~/.skillopt-sleep/state.json: last_sleep, night counter, slow/meta memory
+  staging.py       # write/adopt staging dir, backups
+  cli.py           # `python -m skillopt.sleep {run|status|adopt|harvest|dry-run}`
+  config.py        # SleepConfig load/merge (defaults + ~/.skillopt-sleep/config.yaml)
+  optimizer_backend.py  # thin: route reflect/judge to a chosen backend; mock backend for tests
+
+skillopt-sleep-plugin/            # the Claude Code plugin surface
+  .claude-plugin/plugin.json
+  commands/sleep.md               # /sleep [run|status|adopt|dry-run]
+  commands/sleep-status.md
+  skills/skillopt-sleep/SKILL.md  # so Claude knows how to drive the engine
+  hooks/hooks.json                # optional: schedule + on-session-end harvest
+  scripts/*                       # shims that call `python -m skillopt.sleep ...`
+```
+
+**Reuse, don't fork:** `consolidate.py` calls into existing
+`skillopt.optimizer.clip.rank_and_select`, `skillopt.gradient.aggregate.merge_patches`,
+`skillopt.optimizer.skill.apply_patch_with_report`, and
+`skillopt.evaluation.gate.evaluate_gate`. The sleep layer is an **EnvAdapter-shaped
+shim** over the user's own life, not a new optimizer.
+
+### 5.3 Data flow (one task, end to end)
+
+```
+history.jsonl + <session>.jsonl
+   └─harvest→ SessionDigest{prompts, finals, tools, feedback}
+        └─mine→ TaskRecord{intent, attempted, outcome, reference}
+             └─replay(current skill+mem)→ ReplayResult{hard, soft, trajectory}
+                  └─reflect→ edits(skill), edits(memory)
+                       └─rank/clip(edit_budget)→ candidate
+                            └─GATE(replay held-out)→ accept? → staging/  → (adopt) live CLAUDE.md/SKILL.md
+```
+
+## 6. Scheduling & triggering
+
+- **Cron/scheduled:** documented `crontab` line + an optional Claude Code hook; default `0 3 * * *` (3am local; pick an off-:00 minute in practice). The engine is a plain CLI so it works under cron, systemd-timer, or the Claude Code scheduler.
+- **User-triggered:** `/sleep run` (full cycle), `/sleep dry-run` (harvest+mine+replay, no edits), `/sleep status`, `/sleep adopt`.
+- **On-session-end harvest (optional hook):** cheaply append the just-finished session to the night's buffer so the 3am run has fresh data without a full rescan.
+
+## 7. Safety model (hard requirements)
+
+1. **Never mutate live `CLAUDE.md`/`SKILL.md` except via explicit `adopt`** (or opt-in `auto_adopt`). Default = staged + reviewed (Dreams contract).
+2. **Backups:** every adopt snapshots the prior file to `staging/<date>/backup/`.
+3. **Read-only harvest:** transcripts are read, never written.
+4. **`fresh` replay runs only in throwaway worktrees**, never the user's checkout; no `rm -rf`, no force-push, network off unless `replay.network: true`.
+5. **Budget cap:** `max_tokens_per_night` + `max_tasks_per_night`; stop early when hit, log what was skipped (no silent truncation).
+6. **Secret hygiene:** redact obvious secrets from digests before they enter prompts (reuse `_redact_*` ideas from trainer).
+7. **PII/scope:** only harvest projects on an allowlist (default: the project the plugin is invoked in) or `projects: all` opt-in.
+
+## 8. Validation experiment — "does it actually improve?"
+
+A self-contained, **deterministic-by-default** experiment lives in
+`skillopt/sleep/experiments/` and is the acceptance test for the whole idea.
+
+**Setup:** a synthetic "user persona" (e.g. *researcher who keeps asking for
+arXiv-id extraction in a fixed format*, or *programmer who keeps mis-formatting
+git commit messages*). We ship 12–20 tiny tasks with **exact checkable
+references**, split into `replay` (train) and `holdout` (test).
+
+**Procedure:**
+1. Score the holdout with an **empty** skill+memory → `baseline`.
+2. Run `N` sleep nights (each: replay train slice → reflect → gated edit).
+3. Score holdout with the evolved skill+memory → `after`.
+4. Report `after − baseline`, accept/reject counts, edit count, tokens.
+
+**Two backends:**
+- `mock` (default, **no API key, fully deterministic**): a scripted optimizer that proposes the known-good rule on failure and a scripted judge. Proves the *plumbing* (harvest→mine→replay→gate→adopt) monotonically improves the score and the gate blocks regressions. This is the CI-able acceptance test.
+- `anthropic` (opt-in, uses `ANTHROPIC_API_KEY`): the real optimizer/judge, to demonstrate genuine lift on the persona tasks.
+
+**Success criteria:**
+- Mock: `after > baseline`, gate rejects an injected harmful edit, adopt+backup works, re-run is reproducible. (Hard gate in CI.)
+- Anthropic (when run): `after ≥ baseline` on holdout with ≥1 accepted, human-readable edit; documented in the wake-up report.
+
+## 9. Personas (the user's framing) → concrete recurring-task families
+
+- **Programmer:** commit-message conventions, repo-specific build/test commands, "always run X before Y", framework gotchas → consolidated into project `CLAUDE.md` + a `repo-workflow` skill.
+- **Researcher:** citation/format preferences, experiment-logging habits, paper-section style, dataset-path memory → `research-prefs` skill + memory.
+- **Finance/analyst:** report formatting, recurring data-pull recipes, terminology → `report-style` skill + memory.
+The engine is domain-agnostic; the persona only changes which tasks get mined.
+
+## 10. Phased delivery
+
+- **Phase 0 — scaffold + types + harvest** (read-only, no API). Provable on this box's real `~/.claude`.
+- **Phase 1 — mine + replay(mock) + consolidate + gate + staging**, with the **mock** optimizer backend and the deterministic experiment green. *(primary deliverable of the offline session)*
+- **Phase 2 — plugin surface** (`/sleep`, skill, hooks, plugin.json) wired to the CLI.
+- **Phase 3 — real Anthropic backend** for miner/reflect/judge + `fresh` replay in worktrees.
+- **Phase 4 — slow/meta cross-night memory**, adopt automation, multi-project, polish + docs.
+
+This session targets **Phase 0 + Phase 1 fully**, **Phase 2 scaffolded**, and the
+**deterministic experiment passing**, all committed (not pushed) for review.
+
+## 11. Open questions for the user (answer when awake)
+
+1. **Adopt policy:** keep default *review-gated*, or do you want `auto_adopt` for your own machine?
+2. **Scope:** harvest only the invoked project, or all projects in `~/.claude/projects`?
+3. **Real-API demo:** want me to spend live `ANTHROPIC_API_KEY` budget on the persona demo, or keep everything mock until you say go?
+4. **Skill target:** evolve a *new* dedicated `skillopt-sleep`-managed skill, or also edit your existing hand-written skills in `~/.claude/skills`?
+5. **Paper:** should this become a section/figure in the SkillOpt arXiv (Dream+Sleep framing as "deployment-time continual skill optimization")?
+```