docs(release): prepare v0.2.0 release candidate

Author: reacher-z

.env.example

@@ -14,10 +14,10 @@ GEMINI_LOCATION=global   # or us-central1 / europe-west4
 # Needed for TIGER-Lab/EditReward-Bench. Set this OR run `huggingface-cli login`.
 # HF_TOKEN=hf_xxx
 
-# ---- Optional: override the score-guidelines path ----
-# Defaults to <repo>/score-guidelines/. Only set this if you've moved the
-# templates or are running the package from an unusual install layout.
-# REWARDHARNESS_TEMPLATES_DIR=/absolute/path/to/score-guidelines
+# ---- Optional: override packaged score guidelines ----
+# Defaults to rewardharness/resources/score_guidelines inside the installed
+# package. Set this only for a deliberate custom template bundle.
+# REWARDHARNESS_TEMPLATES_DIR=/absolute/path/to/score_guidelines
 
 # ---- Optional: swap in a different Sub-Agent VLM ----
 # The OpenAI-compatible model id sent to vLLM /v1/chat/completions.

.github/ISSUE_TEMPLATE/bug_report.md

@@ -11,6 +11,7 @@ A one-line description of what went wrong.
 ## Reproduction
 ```bash
 # Exact command(s). Include `make check` output if relevant.
+# Prefer the v0.2 CLI form, for example: rewardharness inspect
 ```
 
 ## Expected vs. actual
@@ -22,6 +23,7 @@ A one-line description of what went wrong.
 - Python: `python --version`
 - Key package versions: `pip show vllm google-genai openai | grep -E '^(Name|Version)'`
 - Output of `make check` (or `python scripts/check_env.py`):
+- Import style used: `rewardharness.*` or deprecated `src.*` compatibility layer
 
 ## Logs / stack trace
 <details>

.github/ISSUE_TEMPLATE/config.yml

@@ -1,5 +1,8 @@
 blank_issues_enabled: false
 contact_links:
+  - name: v0.2 migration guide
+    url: https://github.com/TIGER-AI-Lab/RewardHarness/blob/main/README.md
+    about: Public package, CLI, Library schema, and compatibility-layer guidance.
   - name: Paper
     url: https://arxiv.org/abs/2605.08703
     about: Read the paper before filing a question about results or methodology.

.github/ISSUE_TEMPLATE/feature_request.md

@@ -16,3 +16,7 @@ What would the smallest version of the fix look like? File paths + behavior, not
 
 ## Out of scope
 (optional) Things you are NOT proposing — helps reviewers stay focused.
+
+## Compatibility
+State whether the proposal changes the public `rewardharness.*` API, Library
+schema, checkpoint schema, or CLI output.

.github/dependabot.yml

@@ -1,7 +1,7 @@
 version: 2
 
 updates:
-  # Python dependencies (core + vLLM extra)
+  # Python dependencies (runtime, development/release tooling, and vLLM extra)
   - package-ecosystem: pip
     directory: "/"
     schedule:

.github/workflows/release.yml

@@ -0,0 +1,62 @@
+name: Release
+
+on:
+  push:
+    tags: ["v*"]
+
+permissions:
+  contents: read
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+      - run: python -m pip install -r requirements-dev.txt
+      - name: Verify tag and package version
+        run: |
+          package_version=$(python -c 'import rewardharness; print(rewardharness.__version__)')
+          tag_version=${GITHUB_REF_NAME#v}
+          test "${tag_version//-/}" = "$package_version"
+      - run: make release-check
+      - uses: actions/upload-artifact@v4
+        with:
+          name: distributions
+          path: dist/
+          if-no-files-found: error
+
+  pypi:
+    needs: build
+    runs-on: ubuntu-latest
+    environment: pypi
+    permissions:
+      id-token: write
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: distributions
+          path: dist/
+      - uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          attestations: true
+
+  github:
+    needs: build
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: distributions
+          path: dist/
+      - name: Create GitHub release
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          prerelease=""
+          [[ "$GITHUB_REF_NAME" == *-* ]] && prerelease="--prerelease"
+          gh release create "$GITHUB_REF_NAME" dist/* --generate-notes $prerelease

CHANGELOG.md

@@ -1,8 +1,8 @@
 cff-version: 1.2.0
 message: "If you use this software, please cite the paper."
 title: "RewardHarness: Self-Evolving Agentic Post-Training"
-version: "0.1.2"
-date-released: "2026-05-16"
+version: "0.2.0-rc1"
+date-released: "2026-08-17"
 type: software
 authors:
   - given-names: Yuxuan

CITATION.cff

@@ -1,5 +1,9 @@
 # RewardHarness
 
+The canonical import namespace is `rewardharness`; `src` contains compatibility
+adapters only. New code must not introduce dependencies from `rewardharness`
+back to `src`.
+
 Self-evolving reward model for image editing quality evaluation.
 Qwen2.5-VL-7B (vLLM) as Sub-Agent; Gemini `gemini-3.1-pro-preview` as orchestration layer (Router, ChainAnalyzer, Evolver).
 
@@ -12,17 +16,17 @@ Gemini (orch layer)          Qwen vLLM (sub-agent + tools)
   Evolver._validate_tool()
 ```
 
-- **Router** (`src/router.py`): selects relevant Skills/Tools from Library per editing prompt
-- **ChainAnalyzer** (`src/chain_analyzer.py`): analyzes reasoning chains → improvement signals (skill_updates, tool_updates)
-- **Evolver** (`src/evolver.py`): applies signals to Library; validates tool prompts via vLLM; snapshot/restore for rollback
-- **SubAgent** (`src/sub_agent.py`): multi-turn Qwen reasoning with `<think>/<tool>/<obs>/<answer>` tags
-- **Library** (`src/library/`): Skills (markdown evaluation guidance) + Tools (VLM system_prompts)
-- **Pipeline** (`src/pipeline.py`): evolution loop with Phase A (skills) / Phase B (tools) / Phase C (pruning)
+- **Router** (`rewardharness/evaluation/router.py`): selects relevant Skills/Tools from Library per editing prompt
+- **ChainAnalyzer** (`rewardharness/evolution/analyzer.py`): analyzes reasoning chains → improvement signals (skill_updates, tool_updates)
+- **Evolver** (`rewardharness/evolution/evolver.py`): applies signals to Library; validates tool prompts via vLLM; snapshot/restore for rollback
+- **SubAgent** (`rewardharness/evaluation/engine.py`): multi-turn Qwen reasoning with `<think>/<tool>/<obs>/<answer>` tags
+- **Library** (`rewardharness/resources/library/`): Skills (markdown evaluation guidance) + Tools (VLM system_prompts)
+- **Pipeline** (`rewardharness/evolution/pipeline.py`): evolution loop with Phase A (skills) / Phase B (tools) / Phase C (pruning)
 
 ## Commands
 
 ```bash
-# Tests (107 tests, ~2s)
+# Tests (138 tests, a few seconds)
 python -m pytest tests/ -v
 
 # Evolution (main experiment)

CLAUDE.md

@@ -10,6 +10,7 @@ Thanks for considering a contribution. RewardHarness is a research codebase, so
    make install-dev
    make check    # env / credentials / endpoints
    make test     # mocked tests, ~2 s
+   make quality  # Ruff, mypy, rating-integrity and shell syntax gates
    ```
    PRs that break the test suite won't be merged.
 3. **Keep PRs focused.** One PR = one logical change. If you're refactoring while adding a feature, split into two PRs.
@@ -45,11 +46,16 @@ One-line subject (≤72 chars), then optional body. Reference issues with `#NNN`
 
 ## Tests
 
-- All new code in `src/` should have a corresponding mocked test in `tests/`.
-- Tests must not make real network calls. If you need a Gemini response, patch `src.router.call_gemini`. If you need a Sub-Agent completion, patch `src.sub_agent.OpenAI`. See `tests/README.md` for the patterns.
+- All new code in `rewardharness/` should have a corresponding mocked test in `tests/`.
+- Tests must not make real network calls. Inject or patch the client boundary;
+  see `tests/README.md` for the current patterns.
 
 ## Releases
 
 Versions follow [SemVer](https://semver.org/). Before tagging, maintainers run
 `make release-check`, cut a tag + GitHub Release, and add a `## [x.y.z]` block
 to `CHANGELOG.md`.
+
+Release candidates use PEP 440 versions such as `0.2.0rc1` and matching tags
+such as `v0.2.0-rc1`. The tag workflow publishes the same validated artifacts
+to PyPI and GitHub through trusted publishing.