test: pin CLAUDE.md and AGENTS.md identical

.github/workflows/ci.yaml

@@ -1,13 +1,11 @@
 name: CI
 
-# Explicit, not "*.md": CONTRIBUTING.md and docs/ are asserted/executed by tests (guarded by tests/test_documentation/test_ci_paths_ignore.py).
+# Explicit, not "*.md": CONTRIBUTING.md, docs/, CLAUDE.md and AGENTS.md are asserted/executed by tests (guarded by tests/test_ci_paths_ignore.py).
 # Safe only while CI is not a required status check on main: a skipped required check never reports, blocking docs-only PRs.
 on:
   push:
     branches: [ "main" ]
     paths-ignore: &docs_only
-      - AGENTS.md
-      - CLAUDE.md
       - CODE_OF_CONDUCT.md
       - NOTICE.md
       - README.md

AGENTS.md

@@ -5,6 +5,7 @@
 - **Orchestration Only**: Coordinate Test-Driven Development cycles between specialized agents
 - **No Code Implementation**: NEVER write implementation code or tests directly
 - **Agent Delegation**: Use Red Agent for test writing, Green Agent for implementation
+- **Session root**: `.claude/agents/red-agent.md` and `.claude/agents/green-agent.md` are only auto-loaded when Claude Code's session is rooted at this repository. If working from a parent or workspace directory, start a session inside the clone (`cd code/mloda && claude`) so the agents are available natively.
 
 ## TDD Workflow
 
@@ -51,7 +52,7 @@ Each phase should be a clean, validated checkpoint with all tests passing and ch
 **CRITICAL**: If agent behavior is unexpected or incorrect:
 
 1. **Update Agent Configuration**: Modify `.claude/agents/red-agent.md` or `.claude/agents/green-agent.md` to refine instructions, constraints, or workflow
-2. **Update This File**: Modify `CLAUDE.md` and `AGENTS.md` (keep them in sync) to clarify orchestration rules or add missing guidance
+2. **Update This File**: Modify `CLAUDE.md` and `AGENTS.md` (`tests/test_agent_docs_sync.py` pins them identical) to clarify orchestration rules or add missing guidance
 3. **Document Changes**: Briefly explain what was learned and why the change improves behavior
 
 This enables continuous learning and improvement of the TDD workflow based on actual usage patterns.
@@ -85,6 +86,12 @@ source .venv/bin/activate
 - **`attribution/ATTRIBUTION.md`**: `tox` regenerates this file from the installed dependency versions on every run, so a dependency change shows up as a diff here. This is intended: commit the update as part of the same change so the tracked file stays current. The release workflow does not regenerate it; it ships the committed copy, so keeping it up to date in PRs is what keeps releases accurate.
 - **Commits**: use [Conventional Commits](https://www.conventionalcommits.org/) (`feat:`, `fix:`, `chore:`, `docs:`, `test:`, `refactor:`, `minor:`). semantic-release computes the next version. This project deviates from the standard: the minor version (middle number) bumps only on `minor:` commits; `feat:` is treated as a patch bump.
 
+### mypy iteration notes
+
+- `mypy --strict --ignore-missing-imports .` (what tox runs) checks `tests/` as well as `mloda/` and `mloda_plugins/`. Running `mypy mloda/` locally will miss test-tree type errors that block the tox gate. `tox` (or `tox -e python310`) is the only trustworthy mypy invocation.
+- A stray local build (`python -m build` or `pip wheel .`) creates a `build/` directory. mypy picks it up and reports spurious errors. Clean it with `rm -rf build/` before running tox. The `[tool.mypy]` exclude list in `pyproject.toml` includes `build/` to prevent this at the gate.
+- Running `mypy --strict .` in the dev venv (outside tox) may surface pyspark-related errors that tox's isolated env does not see; trust the tox run as the source of truth.
+
 ## Issue Creation
 
 When filing a GitHub issue (via `gh issue create` or otherwise), follow the structure in `.github/ISSUE_TEMPLATE/issue.yml`:

CLAUDE.md

@@ -52,7 +52,7 @@ Each phase should be a clean, validated checkpoint with all tests passing and ch
 **CRITICAL**: If agent behavior is unexpected or incorrect:
 
 1. **Update Agent Configuration**: Modify `.claude/agents/red-agent.md` or `.claude/agents/green-agent.md` to refine instructions, constraints, or workflow
-2. **Update This File**: Modify `CLAUDE.md` and `AGENTS.md` (keep them in sync) to clarify orchestration rules or add missing guidance
+2. **Update This File**: Modify `CLAUDE.md` and `AGENTS.md` (`tests/test_agent_docs_sync.py` pins them identical) to clarify orchestration rules or add missing guidance
 3. **Document Changes**: Briefly explain what was learned and why the change improves behavior
 
 This enables continuous learning and improvement of the TDD workflow based on actual usage patterns.

tests/test_agent_docs_sync.py

@@ -0,0 +1,39 @@
+"""CLAUDE.md and AGENTS.md are one instruction set for different agents, so the two files must be identical."""
+
+import difflib
+from pathlib import Path
+
+
+PROJECT_ROOT = Path(__file__).resolve().parent.parent
+
+# Naming both files here makes tests/test_ci_paths_ignore.py treat them as protected: a change to
+# either must run the gate, so neither may stay in the ci.yaml docs-only paths-ignore list.
+CLAUDE_MD = PROJECT_ROOT / "CLAUDE.md"
+AGENTS_MD = PROJECT_ROOT / "AGENTS.md"
+
+
+def _lines(path: Path) -> list[str]:
+    assert path.is_file(), f"{path.name} must exist at the project root: each agent reads only its own file."
+    # Bytes, not read_text(): universal-newline translation would let a CRLF rewrite compare equal.
+    return path.read_bytes().decode("utf-8").split("\n")
+
+
+def _diff(claude: list[str], agents: list[str]) -> str:
+    """Diff of repr-ed lines, so whitespace-only drift is not printed as two identical-looking lines."""
+    return "\n".join(
+        difflib.unified_diff(
+            [repr(line) for line in claude],
+            [repr(line) for line in agents],
+            fromfile="CLAUDE.md",
+            tofile="AGENTS.md",
+            lineterm="",
+        )
+    )
+
+
+def test_claude_md_and_agents_md_are_identical() -> None:
+    claude, agents = _lines(CLAUDE_MD), _lines(AGENTS_MD)
+    assert claude == agents, (
+        f"CLAUDE.md and AGENTS.md have drifted:\n{_diff(claude, agents)}\n\n"
+        "Neither file carries content of its own: whatever belongs in one belongs in the other, verbatim."
+    )

tests/test_ci_paths_ignore.py

@@ -138,6 +138,8 @@ def test_protected_set_is_not_stale() -> None:
     protected = _protected_paths()
     assert "docs/docs/index.md" in protected, "docs/ markdown is executed by tests/test_documentation"
     assert "CONTRIBUTING.md" in protected, "CONTRIBUTING.md content is asserted by tests/test_project_structure.py"
+    assert "CLAUDE.md" in protected, "CLAUDE.md content is asserted by tests/test_agent_docs_sync.py"
+    assert "AGENTS.md" in protected, "AGENTS.md content is asserted by tests/test_agent_docs_sync.py"
 
 
 def test_paths_ignore_declared_for_push_and_pull_request() -> None: