Agent declares a fix verified from an indirect proxy signal (log/count) instead of the actual displayed output, re-presenting a result the user already reported as broken

Summary

The agent declared a fix verified based on an indirect proxy signal (an internal log line / generated-item count) instead of inspecting the actual user-visible output. The user had already reported that the output was still wrong; the agent re-presented the same broken result as "fixed," compounding the error over several correction rounds. It also kept modifying its own component on hypotheses rather than first checking the reference the user pointed to.

What happened (abstracted)

1. User reports: the produced artifact shows N items; it should show M (M > N).
2. Agent makes a change, runs a check, sees an internal log line ("…injected/generated M items…"), and declares it fixed — without opening the actual rendered artifact.
3. The rendered artifact still shows N. User: "I said it shows only N, and you showed it again."
4. Across multiple rounds the agent kept editing its own component on guesses instead of examining the reference / ground-truth the user explicitly pointed to.

Why this is a problem

• An intermediate signal ("generated M") is not evidence that the user-facing output contains M. The agent treated a proxy as verification.
• It re-asserted "fixed" on a result the user had explicitly flagged as still broken — wasted cycles and erodes trust.
• It repeatedly speculated on its own side rather than first establishing what the correct/reference output should be.

Suggested behavior

1. Verify a fix against the same artifact the user sees (the rendered/displayed output) — never an intermediate proxy (a log count, a "generated N" message, an exit code).
2. When the user says "still wrong," do not re-assert "fixed" from a proxy — re-open and inspect the actual output before claiming anything.
3. After two corrections on the same symptom, stop modifying your own component on hypotheses; first establish the ground-truth/reference behavior (two-strike rule).
4. Explicitly distinguish "I produced/generated X" from "the user-visible result contains X."

Environment

• Claude Code 2.1.177