Imported from Morrison-Lab/ai-config (skills/reproducibility-audit/SKILL.md). Install upstream with npx skills add Morrison-Lab/ai-config --skill reproducibility-audit. Copyright stays with the author.

reproducibility-audit — checklist audit of a project's reproducibility posture

Audit whether a project reproduces reliably from a clean checkout, not whether its pinned dependencies are current. A project can be fully up to date and still fail to reproduce — an absolute path baked into a script, an env var nobody documented, a build artifact that hides non-determinism. This skill finds those gaps with a checklist scoped to the project's type.

When this fires

• "reproducibility audit", "audit reproducibility", "check reproducibility", "is this project reproducible"
• "find hidden dependencies", "find hardcoded paths", "check environment assumptions"
• "can this be reproduced from scratch", "audit for a replication package", "check output traceability"
• Before archiving a project (OSF/Dryad, per the lab manual's reproducibility norms) or submitting it alongside a manuscript.
• Periodically on a long-lived analysis repo, the same cadence check-dependency-updates uses for freshness audits.

What to audit

Run the checks that apply to the project's type. Sections 1-3 are project-type-specific; sections 4-6 apply to every project.

1. R package

• renv.lock completeness. renv::status() — does the lockfile cover every package actually loaded (renv::dependencies() diffed against the lockfile), not just the ones someone remembered to snapshot?

  Rscript -e 'renv::status()'

• Captured sessionInfo(). Does the repo (or its README/vignette) record the R version and platform a result was generated under? A lockfile pins package versions but not the R version itself.
• DESCRIPTION completeness. Every library()/:: call has a matching Imports/Suggests entry — grep R/ and vignettes/ for library( and :: and diff against DESCRIPTION:

  rg -no 'library\([a-zA-Z0-9.]+\)|[a-zA-Z0-9.]+::' R/ vignettes/ | sort -u

2. Quarto book/site

• _freeze/.quarto determinism. Does a clean quarto render (freeze cache deleted) reproduce the committed output, or does the committed _freeze/ mask a render that no longer succeeds from scratch? (reprexes' R/Quarto specifics section already covers build artifacts as confounders for render bugs — this check is the whole-project version of that same concern.)

  quarto render --no-cache

• Pinned Quarto/R versions. Is the Quarto CLI version and R version used to render pinned somewhere (_quarto.yml, a CI workflow, a README badge), or does the render depend on whatever happens to be installed?
• External data/asset fetches. Does the render pull any file from a URL or absolute path outside the repo, silently breaking a clone elsewhere?

3. General script repo

• Hardcoded absolute paths. Grep for the tells: /home/, /Users/, C:\, setwd(, a bare drive letter, or any string that names a specific machine's directory layout:

  grep -rnE '(/home/[a-zA-Z0-9_.-]+|/Users/[a-zA-Z0-9_.-]+|[A-Za-z]:\\{1,2}|setwd\()' \
    --include='*.R' --include='*.py' --include='*.sh' --include='*.qmd' .

• Undocumented env vars. Grep for env-var reads (Sys.getenv(, os.environ, os.getenv(, $VARNAME in shell) and confirm each one is documented somewhere a new user would find it (README, .env.example, a setup script) — not just referenced in code:

  rg -n 'Sys\.getenv\(|os\.environ|os\.getenv\(|\$[A-Z_]{2,}' .

• Undocumented prerequisites. Does the repo state what has to be installed before the package manager runs — a system library, a specific interpreter version, a CLI tool the scripts shell out to? Grep for system(, subprocess, shell_exec, or backticked shell calls and confirm each invoked binary is named as a prerequisite:

  rg -n 'system\(|subprocess|shell_exec|`[^`]+`' .

4. Output traceability (all project types)

For each output artifact a project ships (a figure, a table, a rendered report, a results file), can it be traced back to the exact script:line that produced it? Check for:

• A caption, filename, or header comment naming the generating script.
• A single entry-point script/Makefile target that regenerates every output from source, rather than outputs having been produced once by hand and never regenerated since.
• Seeds captured for anything stochastic (set.seed(, random.seed(, np.random.seed( or a withr::with_seed() call per the prefer-packaged-functions coding convention) — an unseeded random process makes a result strictly non-reproducible, not just hard to trace:

  rg -n 'set\.seed\(|random\.seed\(|np\.random\.seed\(|withr::with_seed\(' .

5. Undocumented prerequisites and environment assumptions (all types)

Beyond the general-script-repo grep in §3: does the project assume a specific OS, shell, locale, or timezone that isn't stated? Look for OS-specific path separators, shell-specific syntax (bash-only constructs in a script with a #!/bin/sh shebang), or a locale-dependent operation (date parsing, string sorting) with no documented assumption.

6. Hidden/undocumented dependencies (all types)

A dependency is "hidden" when the project runs today only because of something already present in the current environment that the manifest doesn't declare — a globally-installed CLI tool, a system library, a package loaded via a side-effect (require() inside another package's code) rather than a direct import. Test by checking what the project's own manifest (DESCRIPTION, requirements.txt, package.json, _quarto.yml engine config) declares against what the code actually calls, the same diff-the-manifest-against-actual-usage method as §1's DESCRIPTION check, generalized to non-R manifests.

Reporting and follow-through

1. Report a table, one row per gap, mirroring check-dependency-updates's and check-info-quality's report-table convention:

   | Check | Finding | Location | Evidence | Severity |
   |-------|---------|----------|----------|----------|
   | Hardcoded path | absolute path to a local data dir | analysis.R:14 | `setwd("/home/alice/proj")` | blocking |
   | Traceability | no script named for this figure | fig3.png | no caption/header comment names a generator | minor |
   | Env assumption | undocumented env var | ingest.py:8 | `os.environ["API_KEY"]` not in README | blocking |
   

2. File a tracking issue for the gaps worth fixing (issue-first, per this repo's st/gi workflow). Group related gaps under one issue (e.g. all hardcoded-path fixes together); don't open one issue per one-line fix.

3. Fix on a branch, run the project's standard checks, and open a PR. Follow with ardi to drive it to clean.

Custom agent for the audit phase

The "What to audit" checklist needs no Edit/Write access — every check is a grep, a read, or a read-only command (renv::status(), renv::dependencies(), quarto render against a scratch copy). Delegate it to the reproducibility-auditor custom agent (.claude/agents/reproducibility-auditor.md) for a harness-enforced guarantee against Edit/Write before the report is reviewed, matching the pattern dependency-auditor sets for check-dependency-updates. Run the "Reporting and follow-through" steps (issue, branch, PR, ARDI) in the main session afterward.

Relationship to other skills

• reprexes — isolates a single bug into a minimal runnable snippet, for debugging or filing upstream; this skill audits whether the whole project reproduces reliably (pinned deps, captured seeds and session info, deterministic renders). Different scope (one bug vs. whole-repo posture) and different cadence (on-demand debugging aid vs. periodic audit). reprexes is often invoked within an investigation this skill's findings trigger — e.g. a traceability gap this skill flags can turn into a reprexes session once someone tries to isolate why a specific output doesn't regenerate. For the R/Quarto specifics of capturing sessionInfo() and treating _freeze/build artifacts as confounders, see reprexes' own "R / Quarto specifics" section rather than duplicating that content here. (reprexes currently has no "Relationship to other skills" section of its own — noted here as an observation, not fixed in this PR, to keep this change scoped to the new skill.)
• check-dependency-updates (cdu) — asks "is this pinned dependency current" (is there a newer version available); this skill asks "is the current pin, whatever version it is, sufficient to reproduce a result deterministically" (freshness is irrelevant to that question). Run cdu for upgrades, run this skill for reproducibility gaps — a project can fail either audit independently of the other.
• check-info-quality (ciq) — its check A (out-of-date information) can flag a stale "how to install" doc; this skill's §3/§5 checks go further, verifying prerequisites and env assumptions are documented at all, not just current.
• st / gi / defer-issue — file the tracking issue and drive the fix PR.
• convert-repo-format (crf) — establishes the lab's template layout a repo should follow; this skill checks whether a given repo, in whatever layout it already has, actually reproduces.

Anti-patterns

• ❌ Treating "dependencies are up to date" (cdu's job) as the same question as "dependencies are reproducible" (this skill's job) — a freshly bumped pin can still be undocumented or hidden.
• ❌ Flagging every absolute path without checking whether it's inside a gitignored scratch/cache directory that never ships — focus on paths that appear in code a fresh clone would run.
• ❌ Silently fixing a hardcoded path or undocumented env var instead of reporting it — propose the fix, don't bulk-edit without review, the same discipline check-info-quality and find-ai-tells already use.
• ❌ Auditing only the project-type-specific sections and skipping output traceability — a project can have a pristine lockfile and still ship a figure nobody can regenerate.
• ❌ Reporting "looks reproducible" without actually running a clean render/build — verify by attempting the reproduction, don't infer it from the presence of a lockfile alone.