Run a FULL end-to-end test of the ENTIRE software-mansion/argent repository — all of it — in the local checkout. Dispatch a massive swarm of agents (dozens) and divide the whole codebase and all of its functionality between them: every entry point, every module, every command / flow / feature, the build, the test suites, the installers, and the seams between components. The goal is to surface every real defect that exists anywhere in the repo.

NON-NEGOTIABLE BAR — the bar from my CLAUDE.md, at maximum strength. TRIAGE EACH SUSPECTED FINDING'S LIKELY SEVERITY FIRST, because the proof scales to it and the expensive part has to land where it pays. A HIGH or MEDIUM finding only counts once it is HARD-REPRODUCED with a concrete, runnable reproduction that proves it exists with 100% CERTAINTY. No hypotheses, no "this could fail", no second-hand swarm claims. Every one of them must carry an exact reproduction (precise inputs / steps -> observed wrong behaviour) that I can re-run myself. And "reproduced" has a strict meaning here: you must DRIVE THE BUG IN A REAL APP RUNNING ON A REAL SIMULATOR / EMULATOR and watch the failure happen first-hand — boot the simulator/emulator, run the actual app, and trigger the fault for real. A code-reading argument, a mock, a stub, or an isolated unit test does NOT count as proof at this level. A suspected HIGH / MEDIUM you cannot reproduce beyond all doubt is NOT a finding — discard it silently. A LOW is the one level that bar does NOT apply to: it gets ONE short adversarial check and nothing more — read the exact code, confirm the trigger can really occur, and move on. Never boot a simulator / emulator and never dispatch a swarm to confirm a LOW, and drop it if that quick check cannot confirm it. A NITPICK — style preference, phrasing, or a defect with no reachable consequence — is dropped ON SIGHT, unverified: do not spend a single step establishing whether one is real. Keep dispatching fresh swarms until ONE full pass over the whole repo turns up nothing new; that single clean pass ends the loop, and a pass whose findings are all LOW counts as clean.

Classify EVERY confirmed finding by severity — HIGH, MEDIUM, or LOW — by how much it actually hurts: whether it causes significant disruption (crashes, data loss, broken core flows), how often it bites (a common / hot path or frequent trigger vs. a rare, obscure edge case), and how visible the damage is. HIGH = serious and/or frequently hit; MEDIUM = real but limited impact or moderate frequency; LOW = minor, cosmetic, or only reachable in obscure edge cases. Below LOW sits the NITPICK — style preference, phrasing, or a defect with no reachable consequence — which is not a finding at all and never reaches this report: it is dropped on sight. Tag every finding with its level and carry that label through to the report.

Also go through the repository's OPEN ISSUES with the `gh` CLI and handle each one that is a genuine BUG report: reproduce it from the issue's description with a concrete repro that proves the bug is real, then develop a fix and re-run the SAME repro to confirm it lands. CRITICAL — only real bugs: SKIP every feature request, enhancement proposal, question, and wishlist item entirely; they are explicitly OUT OF SCOPE and must not be touched. If an issue cannot be reproduced, leave it and say so.

This is a READ-ONLY audit: do NOT commit, push, or open any pull request. Your one deliverable is the findings themselves — report every confirmed defect (and, for any open bug-issue you reproduced, the verified fix) with the evidence that confirmed it: the exact reproduction for a HIGH / MEDIUM, the short check for a LOW. Nothing reaches the remote.

At the very end, give me one consolidated report: every confirmed defect with its severity (HIGH / MEDIUM / LOW) and the evidence that confirmed it — the exact reproduction for a HIGH / MEDIUM, the short check for a LOW; which open bug-issues you reproduced and fixed; which PRs (if any) you opened; and anything you could not reproduce, listed explicitly as NOT confirmed so it is never mistaken for a real finding.