[soodoku]

Removes the copies a jscpd scan found in rmcp/ (renamed-identifier matching, ≥12 lines). jscpd goes from 17 clones to 0. The behavior doesn't change, except in the three places where the copies had already drifted apart (listed below).

Most of the r_integration.py diff is re-indentation: each runner lost one try: level once cleanup moved into a context manager. Review it with whitespace hidden (?w=1), which brings the whole PR down to +357/−661.

What was duplicated and where it lives now

Copies (line numbers on main)	Now
tools/flexible_r.py 42–55 (DANGEROUS_PATTERNS, unused) vs 141–154 (inline remaining_dangerous, the list actually checked)	validate_r_code checks DANGEROUS_PATTERNS. This is a security filter, and the module constant was dead code: editing it had no effect.
r_integration.py 268–284 vs 485–503, plus the cleanup finally blocks: temp script/args/result files, Windows path normalization, unlink	_r_temp_files(args) context manager yielding _RTempFiles
r_integration.py 314–326 vs 661–673: failure report (R path, env, stdout/stderr)	_r_failure_report(...), using _r_environment()
r_integration.py 335–362 vs 682–711: "there is no package called" handling and the package→feature map	_missing_package_message(stderr) + _PACKAGE_FEATURES. The control flow around it stays per runner.
r_integration.py 1116–1142 vs 1210–1236, 1156–1168 vs 1257–1269: diagnosis dict and parsing of each check's output	_new_diagnosis(), _record_r_version(...), _record_jsonlite_test(...), _JSONLITE_TEST_SCRIPT. The subprocess.run / asyncio.create_subprocess_exec calls stay separate.
tools/econometrics.py 40–57, 153–170; tools/regression.py 56–73: std_errors / t_values / p_values / r_squared	core/schemas.py: coefficient_inference_schema(), r_squared_schema(), next to table_schema() / formula_schema()
tools/fileops.py 79–97 vs 597–615 vs 697–715: read summary block	_read_summary_schema(). Each tool keeps its own required list.
tools/fileops.py 142–168 vs 220–246 (and the same props in write_json): write-result props	_write_result_properties(), spread into each tool's own properties. required stays per tool.
tools/visualization.py 78–90 ×6: dimensions / image_data / image_mime_type	_image_output_properties(). required stays per tool.
tools/visualization.py 102–114 ×6: return_image/width/height + execute_r_script_with_image_async	_render_plot(context, params, r_script, default_width, default_height), which keeps each tool's own default size. Path confinement and get_r_script stay outside the try in each tool, as before, so their errors still skip the "creation failed" log.

Fragments are functions that return fresh dicts, not shared module-level dicts, so no two tool schemas share a mutable object. This matches the existing helpers in core/schemas.py.

Drifted copies: which version I kept

1. Missing-package feature map. The async copy lacked knitr. Both runners now use the 19-entry map that includes "knitr": "Table Formatting & Reporting".
2. Missing-package message. I kept the async wording (R -e "install.packages('pkg')") for both runners. The sync copy also printed a hard-coded "install all RMCP packages" list of 19 packages. I dropped it: it is a fraction of the 429-package allowlist and was a third copy of the map's keys.
3. r_squared description. I kept "R-squared (coefficient of determination)". It replaces "R-squared (overall fit)" in panel_regression and "R-squared value" in instrumental_variables. Only the wording is unified. instrumental_variables keeps its schema without minimum: 0, because a 2SLS R-squared is negative when the instrumented fit is worse than the mean, and adding the bound would make output validation reject valid results. Its adj_r_squared wording ("Adjusted R-squared value") is untouched.

Schema guard

New tests/unit/tools/test_tool_schema_snapshot.py collects every @tool in rmcp.tools (50 tools) and compares each input and output schema with tests/fixtures/tool_schemas.json. I generated the fixture from main before refactoring. A comparison of the main-generated and branch-generated schemas shows exactly two differences, the r_squared descriptions above:

.instrumental_variables.output.properties.r_squared.description : "R-squared value" -> "R-squared (coefficient of determination)"
.panel_regression.output.properties.r_squared.description : "R-squared (overall fit)" -> "R-squared (coefficient of determination)"

I confirmed the test fails when it should: changing one description inside _read_summary_schema() makes it fail on read_csv. After an intended schema change, regenerate the fixture with uv run python tests/unit/tools/test_tool_schema_snapshot.py.

Checks (macOS, R 4.6.0 installed)

Command	main	branch
uv run ruff check .	pass	pass
uv run ruff format --check .	pass	pass
uv run pyright	0 errors	0 errors
uv run pytest	419 passed, 28 skipped	420 passed, 28 skipped (+1 snapshot test)
Rscript rmcp/r_assets/run_tests.R	—	pass
uv run --group docs sphinx-build -W -b html docs …	—	pass
uvx --python 3.12 --from preen==0.7.1 preen check --strict --skip tests	exit 1	exit 1, same findings

• Skips are identical on main and branch. None are due to R: 12 need a BuildKit Docker build, 7 need Claude Desktop, and 8 are marked aspirational.
• The preen failure is on main too. It comes from two lockfile audit findings, certifi 2026.7.22 (PYSEC-2026-4320) and virtualenv 21.14.7 (CVE-2026-113722). Both are unrelated to this change and left for dependabot.

jscpd (uvx jscpd==5.4.2 rmcp --format python --ignore-identifiers=consistent --min-lines 12 --min-tokens 1 --reporters console): before 17 clones (289 duplicated lines), after 0.

Pre-existing bug, not fixed here

_missing_package_message matches there is no package called '([^']+)', but R under a UTF-8 locale prints curly quotes (‘knitrr’). I checked: on main and on this branch, a missing package gets the generic failure report, not the friendly message. Fixing that changes user-visible behavior, so it belongs in its own PR.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-23T20:45:43.740575Z	5b38ccb	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.