minimal-reproduction-e67988— agent skill

Purpose

You cannot reliably fix what you cannot reliably reproduce in isolation. Apply the subtraction method until a single file deterministically triggers the exact reported error.

When to Use

• A bug is tangled in app/UI/network layers and hard to debug in place
• A failure must be handed to another engineer, team, or upstream project
• Before bisecting or stress-testing, to get a fast deterministic signal

Inputs

• The failing code path and the error it produces
• Sample data that triggers the failure, if any

Workflow — the Subtraction Method

1. Start with the failure. Take the code path that fails.

1. Remove the UI/Network. If the bug is reported via a web request, write a script that calls the internal controller directly.

1. Mock dependencies. If the bug doesn't require the database, mock it. If it doesn't require the third-party API, mock it. Template:

   // repro.js — minimal harness, no framework
   const { failingFunction } = require("./src/module");

   const fakeDb = { query: async () => [{ id: 1, value: null }] }; // smallest stub that triggers it

   failingFunction(fakeDb, { payloadKey: "trigger-value" })
     .then(() => { console.log("NO REPRO"); process.exit(0); })
     .catch((err) => { console.error("REPRO:", err.message); process.exit(1); });

(Exit 0 on no-repro / non-zero on repro also makes the script directly usable by git bisect run.)

1. Prune data. If the bug fails on a 10MB JSON payload, binary search the payload down to the exact 2 keys that trigger the failure.

1. Final output. The result must be a single file — a standalone script or one test case — that relies on ZERO external state, can be run with a single command, and deterministically outputs the exact error reported.

Output

• One standalone file plus the single command that runs it (e.g., node repro.js)
• The exact error output, proving the reproduction matches the report

Verification

• [ ] Repro runs with one command and no external state (no DB, network, or special env)
• [ ] Error produced matches the reported error exactly
• [ ] Repro is deterministic across at least 3 consecutive runs
• [ ] Every remaining line is necessary — removing any piece makes the bug disappear

Failure Modes

• Partial isolation — the "repro" still needs a live database or API; it will rot immediately.
• Wrong error — the script fails, but with a different error than reported; that is a different bug.
• Over-pruning — cutting until the bug vanishes and shipping the last version untested; re-run after every cut.