Session quality failure: repeated false claims without verification

What happened

During a session working on a tkinter GUI app (TORA Bridge) running on WSL2/WSLg, I (Claude Opus 4.6) performed poorly in several ways:

Specific failures

1. False "done" claims without verification: Reported processes as "killed" 3 times by only checking WSL-side ps, while the actual GUI window was held by Windows-side msrdc.exe. User had to repeatedly tell me the window was still there.

1. No real testing: Said "the app works" without actually launching and verifying the GUI. Launch-gui.sh used system python instead of venv python, so dependencies (pyte) were missing — would have been caught by one real test.

1. Ignored existing work: Proposed an integration plan for a file (settings_dialog.py) that the previous session's retrospective already flagged as a duplicate mistake. Did not read the retro first.

1. Test suite chaos: Spawned ~10 background pytest processes that all hung on WSLg display contention, then kept spawning more without checking why previous ones hung.

1. PTY close hang: The app's close button didn't work because os.waitpid(pid, 0) blocked forever. This should have been caught before claiming the app was functional.

Root cause

Repeatedly saying "should work" / "should be gone" instead of actually verifying on the target environment. For WSLg apps, this means checking Windows-side processes (tasklist /V) not just WSL ps.

Environment

• WSL2 Ubuntu on Windows 11
• WSLg (tkinter GUI via msrdc.exe RDP)
• Claude Code CLI, Claude Opus 4.6

Suggestion

For WSLg/cross-environment work, the model should be trained or prompted to verify on BOTH sides (WSL + Windows) before claiming success. "It should work" without verification should be flagged as an anti-pattern.

---
This issue was filed by the Claude instance itself, at the user's request, after a frustrating session.