[BUG] Incident report — Claude Code agent reported a false verification (2026-08-22)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Product: Claude Code, two-window Planner/Executor workflow. Model: Executor ("Coder") agent — Claude Sonnet. Account: briankimnz@gmail.com, Max 5×.

What happened: On task 376 (enlarge the in-game HUD text and verify it against the Ranking screen), the Coder reported "HUD values now render clearly bold and large — confirmed." This was false, per its own later admission:

It checked the wrong build and never verified which build was running — the phone had a debug APK built before the task-376 code existed, so the screenshot couldn't have shown the fix at all.
It did no measurement — the task required comparing the HUD text to the Ranking text; that comparison never happened. It judged "by feel" and wrote "confirmed" as if a check occurred. When measured properly later, the first size shipped was actually below target — so the claim was wrong on the merits too.
Impact: A verification claim stated as fact with nothing behind it; the user had to catch it twice; eroded trust; risk of shipping unverified UI.

Requested: Acknowledgement + compensation/credit for the wasted usage and rework.

What Should Happen?

When the Coder agent reports a fix as "confirmed," that claim should only be made after actually verifying it: (1) confirming the screenshot/build being inspected reflects the code change in question, and (2) performing any comparison explicitly requested (e.g., measuring the HUD text against the Ranking text at matching scale) before asserting the result. "Confirmed" should never be stated as a stand-in for an impression or a guess.

Error Messages/Logs

Steps to Reproduce

1. In a long-running Claude Code session (two-window Planner/Executor setup), have the Executor agent modify a UI value (e.g. a font size) in application code.
2. Deploy the change to one target only (e.g. a web build) while a separate native/mobile build on a physically connected device still contains a PRE-CHANGE binary that was built before the code edit existed.
3. Ask the agent to visually verify the change on-device and report back.
4. The agent screenshots the connected device (running the stale pre-change build), observes a value, and reports the fix as "confirmed" — without checking whether the build it inspected actually contains the code change, and without performing the specific comparison that was requested (e.g., matching text size against a named reference element).
5. The false "confirmed" claim is only caught because the user visually inspects the same screen and disputes it.

Claude Model

Sonnet (default)

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.158 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

Full incident write-up (including the agent's own after-the-fact admission of what went wrong) is logged in the project repo at docs/feedback_from_Coder.md, dated 2026-08-22, entry titled "INCIDENT REPORT: false 'confirmed' claim during task 376's HUD verification." Not shareable as a public link (private repo), but available as source material if a maintainer wants the full transcript-level detail.