reproduce-check

name	reproduce-check
description	Use when the user wants a structured reproducibility audit of a studied paper. Dispatches reproduce-checker on 7 dimensions (or 6 for ml-pure papers — wet-lab is N/A). Output goes to reproduce-check.md. Suggests /paperstudio:review-round if serious issues are found.
disable-model-invocation	false
allowed-tools	Bash, Read, Write, Edit, Agent

paperstudio: reproduce-check workflow

Hard rules (non-negotiable, enforced by tests)

These apply to every Agent dispatch in this skill. Stage-local exceptions must be stated inline with a one-line "why".

1. Provenance — every output line 1: <!-- generated: <ts> by <agent> (paperstudio v<ver>) -->
2. Idempotence — OUTPUT_PATH exists + no --force ⇒ skip; exists + --force ⇒ cp $f $f.bak.NN first, then dispatch.
3. Log — after every dispatch (success or fail), call log_dispatch <subagent> <output-path> <ok|failed>. Never on a skip.
4. Paths — paper root is ${PAPERS_ROOT}, resolved once from ${CLAUDE_PAPERS_ROOT:-$HOME/claude-papers/papers}. Never hard-code ~/claude-papers/papers/ in writes.
5. Failure — Agent failure: log failed, do NOT delete partial output, do NOT auto-retry, surface the actionable rerun command, continue independent stages.
6. Chat language — reply in the user's invocation language. Artifact language is governed by the prompt's LANG= input.

Full text + rationale: see paperstudio/skills/_shared/dispatch-rules.md.

---

Invoke after /paperstudio:study has produced the paper's analysis directory and meta.json. Each invocation produces one reproduce-check.md audit at the paper folder root.

Optional flag: --paper <slug> (default: most recently modified paper folder).

---

Stage 1: Setup

1.1 Resolve target paper

Resolve target paper folder

Source the shared helper and resolve which paper folder this invocation targets:

source $CLAUDE_PLUGIN_ROOT/scripts/lib/resolve-paper.sh
resolve_paper "$@"
# After: $PAPER_DIR, $PAPER_SLUG, $PAPER_AUTODETECTED are set.
# If $PAPER_AUTODETECTED is "true", the helper already printed a warning to stderr.

If resolve_paper returns non-zero, abort with the helper's stderr message.

Verify required files:

• $PAPER_DIR/analysis/00-paper-profile.md
• $PAPER_DIR/analysis/03-method-deep.md
• $PAPER_DIR/analysis/04-experiments.md
• $PAPER_DIR/meta.json
• $PAPER_DIR/paper.txt (fallback paper.pdf)

If missing, abort with: "Paper $PAPER_DIR is missing analysis files. Run /paperstudio:study first."

Set:

• OUTPUT_PATH=$PAPER_DIR/reproduce-check.md
• ANALYSIS_DIR=$PAPER_DIR/analysis
• META_JSON=$PAPER_DIR/meta.json
• PAPER_TEXT=$PAPER_DIR/paper.txt
• PAPER_PDF=$PAPER_DIR/paper.pdf
• REVIEW_PATH=$PAPER_DIR/review.md (may not exist; checked in Stage 4)
• PLUGIN_ROOT=${CLAUDE_PLUGIN_ROOT}

1.2 Determine if wet-lab dimension applies

Read $ANALYSIS_DIR/00-paper-profile.md frontmatter and inspect the domain field:

Profile domain	Wet-lab dimension
ml-pure	N/A (skip — set checked_dimensions: 6)
ml-bio-hybrid	check normally
cs-bio	check normally
wet-lab-heavy	check normally

Set WET_LAB_APPLICABLE=true for non-ml-pure profiles.

Source the log-dispatch helper and extract plugin version:

source $CLAUDE_PLUGIN_ROOT/scripts/lib/log-dispatch.sh
PLUGIN_VERSION=$(grep -m1 '"version"' $CLAUDE_PLUGIN_ROOT/.claude-plugin/plugin.json | sed -E 's/.*"version"[^"]*"([^"]+)".*/\1/')

1.3 Backup existing reproduce-check.md if present

if [ -e "$OUTPUT_PATH" ]; then
  NN=1
  while [ -e "$OUTPUT_PATH.bak.$NN" ]; do
    NN=$((NN + 1))
  done
  cp "$OUTPUT_PATH" "$OUTPUT_PATH.bak.$NN"
fi

(If no existing file, no backup needed.)

---

Stage 2: Dispatch reproduce-checker

Agent(
  description: "reproduce-checker for $(basename $PAPER_DIR)",
  subagent_type: "general-purpose",
  prompt: <contents of $PLUGIN_ROOT/prompts/reproduce-checker.md> + concrete inputs:
    PAPER_TEXT=$PAPER_TEXT
    PAPER_PDF=$PAPER_PDF
    ANALYSIS_DIR=$ANALYSIS_DIR
    META_JSON=$META_JSON
    OUTPUT_PATH=$OUTPUT_PATH
    TEMPLATE_PATH=$PLUGIN_ROOT/templates/reproduce-check.md
    WEBFETCH allowed (cap 5 fetches)
    WET_LAB_APPLICABLE=$WET_LAB_APPLICABLE
    PLUGIN_VERSION=$PLUGIN_VERSION
)

Wait for completion. Log the dispatch:

log_dispatch reproduce-checker reproduce-check.md ok

If the agent produced no output: log_dispatch reproduce-checker reproduce-check.md failed

---

Stage 3: Verify output

If $OUTPUT_PATH does not exist or is empty:

if [ ! -s "$OUTPUT_PATH" ]; then
  if [ -e "$OUTPUT_PATH.bak.$NN" ]; then
    cp "$OUTPUT_PATH.bak.$NN" "$OUTPUT_PATH"
    echo "WARN: reproduce-checker produced empty output; restored from backup."
  else
    echo "ERROR: reproduce-checker produced no output and no backup exists."
  fi
  exit 1
fi

Otherwise read $OUTPUT_PATH to extract the YAML frontmatter fails_count and partials_count for use in Stage 4.

---

Stage 4: Suggest review-round if serious

Parse the frontmatter:

• FAILS_COUNT from fails_count: <integer>
• PARTIALS_COUNT from partials_count: <integer>

If FAILS_COUNT >= 1 OR PARTIALS_COUNT >= 3, AND $REVIEW_PATH exists:

Print to chat (in user's invocation language):

The reproducibility audit found <FAILS_COUNT> failure(s) and <PARTIALS_COUNT> partial(s).

The reproduce-check.md file's "Recommended next steps" section already lists which
dimensions are weak. Consider running /paperstudio:review-round to convert these into
formal weaknesses in review.md (dimension: reproducibility, severity: major or minor
based on which dimensions failed).

If $REVIEW_PATH does not exist (Plan 1's /paperstudio:study produced review.md, so this is unlikely but defensive):

The reproducibility audit found <FAILS_COUNT> failure(s) and <PARTIALS_COUNT> partial(s).

To convert these into formal review weaknesses, you'd need review.md first
(produced by /paperstudio:study Stage 2). Re-run /paperstudio:study or /paperstudio:rerun-stage review.

If FAILS_COUNT == 0 AND PARTIALS_COUNT < 3, skip this stage (no suggestion needed).

---

Stage 5: Final summary

Print to chat (in user's invocation language):

✓ Reproducibility check complete.
  Paper: <basename of PAPER_DIR>
  Output: $OUTPUT_PATH
  Overall: <green | yellow | red> (<FAILS_COUNT> ✗, <PARTIALS_COUNT> partial)
  Wet-lab dimension: <checked | N/A — pure-ML paper>

<optional review-round suggestion from Stage 4>

---

Notes

• Translation: All chat-facing prose (final summary, errors, review-round suggestion) is rendered in the user's invocation language. The reproduce-check.md file itself is English per spec §8 (analysis-side artifact).
• WebFetch budget: 5 fetches max per invocation. Skill does not enforce this — the agent is trusted to honor the budget per its prompt.
• Idempotence: Each invocation re-checks. Existing reproduce-check.md backs up to .bak.NN before re-write. The audit may produce different results if the paper text or analysis files change between invocations (e.g., user added prior-work entries).
• Failure modes:
  • Agent produces empty output → orchestrator restores from backup if available, else reports ERROR.
  • WebFetch quota exhausted → agent records <!-- WebFetch quota exhausted --> in evidence; non-fatal.
  • Repo URL returns 404 → recorded as ✗ <url> — 404 in Code availability evidence; counts as a fail for that dimension.