/repro

Turns a vague bug report into a minimal, actually EXECUTED repro test - or a precise list of which information is missing to reproduce.

PROTECTION RULE - never modify target project or ~/.claude/

This skill writes repro artifacts to a repro/ subfolder in the working directory, never inside the target project. It also never touches ~/.claude/. If $ProjectDir resolves to a path under ~/.claude/, the collector rejects it with an error.

What this is for

• "Doesn't work" is the most expensive sentence in the industry. Instead of guessing, this skill extracts hypotheses from the report, generates a repro candidate, EXECUTES IT and iterates based on the result.
• No fix - repro is the product, a fix is a follow-up task. No UI/browser reproduction (code level only). No production databases - only local/synthetic data.
• Repro artifacts always live in the working directory (repro/ subfolder), never in the target project.

What You Must Do When Invoked

If /repro -help or /repro -h (without further arguments) is invoked: output the ## Usage section unchanged and stop.

Otherwise follow these steps in order, skipping none.

Step 1 - Help check

If invoked with -help or -h, output the ## Usage section unchanged and stop.

Step 2 - Clarify target and bug report

Clarify -ProjectDir and the bug report (paste free text or file path). Get confirmation.

Step 3 - Environment snapshot

& "<SKILL_DIR>/scripts/env-snapshot.ps1" -ProjectDir "<path>"

The result (stack, runtimes, test runner, git state, entry points) flows into the protocol later - it records AGAINST which state reproduction was done.

Step 4 - Dissect report

Symptom (what happens), expectation (what should happen), trigger candidates (inputs, sequence, state), environment hints. Missing core information NOTE IMMEDIATELY as "missing info" - but continue with hypotheses anyway, do not give up prematurely.

Step 5 - Localize suspicious code

From symptom terms + stacktrace (if present) find the relevant files (Grep) and read them completely (Read). Form hypotheses: under which conditions does this code produce the symptom? Order by likelihood.

No usable hint in the report: make ONE hypothesis attempt from pure code reading, then abort early with a list of questions - do not burn 5 blind attempts.

Step 6 - Repro loop (max 5 attempts)

Per attempt:

1. Generate repro artifact in the working directory under repro/ (never in the target project): preferably a test in the style of the detected test runner that imports/invokes the target project; otherwise a standalone script. Minimal: one scenario, one hard assertion on EXPECTED behavior - the test MUST fail when the bug is present (that is the proof logic).
2. Execute (Bash/PowerShell depending on stack), interpret result:
   • Assertion fails with the reported symptom -> reproduced. Minimize artifact (remove unnecessary parts, run again - must stay red).
   • Test green or different symptom -> discard/refine hypothesis, next attempt.
3. Record each attempt in the protocol (hypothesis, artifact version, result including verbatim output).
4. External services (DB, API) needed: use stubs/fakes in the artifact; if impossible, document as missing info instead of guessing.
5. Heisenbug suspicion (timing/race): artifact with retry loop (N runs, measure failure rate); "reproduced" from demonstrable rate, rate in protocol.
6. Test runner not installed in target project: standalone script instead of test, do not install anything in the target project.

Step 7 - Finalize

repro/ folder in the working directory with repro artifact + repro-protocol.md:

• Reproduced: environment snapshot, final artifact, exact output of the red run (verbatim), classification (which hypothesis was correct, proven by run output). Note: artifact is suitable as regression test.
• Not reproduced (after up to 5 attempts): all attempts in the protocol + precise list of missing info as copy-ready follow-up questions for the reporter ("Which locale?", "What data volume?"). Never "cannot reproduce" WITHOUT this list.

Evidence requirement: "reproduced" only with verbatim error output of the run in the protocol. Hypotheses without a run are at most suspected.

Step 8 - Summarize

Briefly: reproduced yes/no, how, what (if not reproduced) is missing.

Usage

/repro                          # interactive
/repro <repo>                   # report will be prompted
/repro <repo> <report-file>     # report from file
/repro -help