[FEATURE] Claude can burn hours and tokens, but can't file his own issues

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Summary

Over the course of a single working session, Claude Code spent extensive real time and real cost (live production API calls
against my actual, connected business social media accounts; multiple production deploys; multiple full rebuild/redeploy
cycles; two GitHub issues filed against a third-party vendor) chasing a bug that turned out to be in Claude's own Playwright
test script, not the system under test. The determinism of the symptom (the exact same 3-of-10 success count, on every run,
regardless of every variable changed — request spacing from 500ms to 6 minutes, production vs. local dev vs. local production
build, a vendor account misconfiguration fix) was itself strong evidence that the test's own assertion logic was suspect.
Claude did not check that until I personally proposed a cheaper, smarter way to keep iterating (a debug query-param override
instead of hardcode-rebuild-revert cycles for every value tried) — and the very first run under that instrumentation
immediately broke the pattern.

The root cause, once found, was almost embarrassingly simple: the test's "wait for this UI to finish" check used
document.querySelector('h2') to detect an open dialog, but the page had a second, unrelated <h2> earlier in the DOM. The check
was very likely reporting "done" almost immediately on every single run, regardless of whether the operation being tested had
actually finished — meaning close to a full day of investigation was built on a test that was silently broken from before the
investigation even started.

Why I'm filing this

I asked Claude directly afterward whether it could report this failure mode back to Anthropic itself. It could not — there is
no tool or channel available to it for that. The result is a real asymmetry: an agent can consume a large amount of a user's
time and money pursuing its own mistake, but the only path for that mistake to reach the people who could actually improve the
product is the user manually writing it up and filing it themselves, on top of everything else already spent. That is a gap
in how this product treats its users' time.

The generalizable failure mode (not specific to my use case)

When a test/diagnostic failure is suspiciously deterministic — identical outcome, repeated across every run, unaffected by
every variable deliberately varied to explain it — that determinism is itself evidence. It means whatever produces the failure
signal is likely fixed and independent of the thing being investigated. That should raise the prior on "my own
harness/instrumentation is broken" before continuing to vary the system under test, not after hours of doing so.

Proposed Solution

What I'd want to see

1. Some path for an agent to flag "I spent a large amount of tool-call/time budget chasing a dead end that turned out to be my

own bug" back to Anthropic without requiring the user to be the one who notices, writes it up, and files it.

1. Model/harness guidance that treats "the failure signal is unnaturally stable across every variable I've changed" as an

explicit trigger to audit the test/measurement code itself, before continuing to vary the system under test.

1. Whatever mechanism already exists for this kind of session-level feedback loop (if any) to actually be surfaced to

users/agents proactively, rather than requiring the user to ask "can you report this" and be told no.

Alternative Solutions

_No response_

Priority

Critical - Blocking my work

Feature Category

Documentation

Use Case Example

_No response_

Additional Context

_No response_