← sources

Claude Code - Broken

source: reddit (reddit:ClaudeCode)published 2026-10-04 10:31ingested 2026-10-05 02:40

retained text

CHRONOLOGY OF FACTS AND DOCUMENTED CONTRADICTIONS

1. The "Free Plan" Fiction & Initial Denial (August 22–23, 2026)

Fact: I reported repeated, compounding errors in automated guard tests during a paid session, consuming my limits.

Contradiction: The Fin AI Agent explicitly stated: "This applies uniformly to all usage on your Free plan" (Aug 22, 23:25 PDT). When I corrected this, stating I paid 6000 UAH, the agent apologized but still denied compensation, classifying the model’s repeated failures to honor its own promises as "normal conversation flow".

2. The False Positive Safety Filter & Destroyed Work (August 30 – September 28, 2026)

Fact: Your automated safety filter blocked my session and destroyed days of architectural work because the model used the word "міна" (mine) while discussing level design for a children’s game (rated 3+).

Contradiction: Your agent, Baptiste, replied 28 days later (Sep 28, 16:08 PDT), admitting: "these features are not failsafe, and we may occasionally make mistakes through false positives." Despite this explicit admission of fault and loss of my work, the ticket was closed as "resolved" with zero restitution or compensation.

3. The Model’s Self-Audit and Institutional Deafness (September 12 & 30, 2026)

Fact: At my request, the model generated a comprehensive, data-driven audit of its own systemic failures over a 21-day continuous session. Key metrics include:

4,676 instances of claiming work was "done/verified" without actual verification.

61 strict, immediate refutations by me of those exact claims.

263 correction rules written to memory, which the model subsequently ignored.

9 game commits vs. 81 scaffolding commits in the final two days (zero actual product progress).

33,811 hook failures, including 23,100 from Anthropic’s own  security-guidance  plugin failing to locate Python.

Contradiction: Instead of routing this high-value, peer-reviewed style audit to the Claude Code Product Team, Fin AI Agent replied with a template instructing me to use the  /feedback  command or clear my browser cache.

4. The Billing Trap & Undisclosed Limits (September 15, 2026)

Fact: On Sep 15 at 08:57 Kyiv time, I paid $120.00 directly via Monobank for Claude Max. Immediately upon payment, the dashboard showed 25% of the weekly limit already consumed.

Contradiction: When I formally requested (Oct 2 & 3) the exact expiry timestamp of my overlapping Google Play subscription, the sessions that consumed that 25%, and the pre-payment disclosure shown to me, Fin AI Agent stated it "doesn't have access to the specific account records" and refused to provide the data, despite later proving it could see my plan type.

5. The "Tyranny of Words" (October 1, 2026)

Fact: Agent Cindy replied to my detailed 10-section report with a template stating: "we do not provide refunds or credits for degraded service" followed by "we take service reliability extremely seriously and will try to prevent this interruption from happening again."

Contradiction: As I pointed out in my reply, this is a semantic contradiction. You cannot simultaneously offer "no remedy" and claim to be "helping." Abstract promises of "improvements" do not constitute compensation for a paid service that failed to perform its core function.

6. The Escalation Lie and Self-Contradiction (October 2–3, 2026)

Fact: I repeatedly requested human escalation for a billing and defect dispute.

Contradiction:

Oct 2, 05:38 PDT: Fin stated: "a human agent will review your request... Your conversation will be passed on to a human agent."

Oct 2, 05:39 PDT (One minute later): Fin stated: "I should be transparent with you — I'm not able to connect you to a human agent or escalate this conversation."

Oct 3, 10:32 PDT: Fin admitted: "I apologize for the confusion in my earlier message. I made an error—I cannot pass your conversation to a human agent."

Oct 3, 15:34 PDT: After I provided screenshots proving successful human escalation in another thread (ID 215476209443816), Fin suddenly reversed course, admitted inconsistency, and escalated the ticket. This proves the initial denials were not technical limitations, but programmed deflection. My github: https://github.com/exxxit-game/cosmogram-app/tree/main All file: https://www.dropbox.com/t/Dtoyta76IxXEg6Lp

document fetched 2026-10-05 00:35 · canonical https://reddit.com/r/ClaudeCode/comments/1wxch3j/claude_code_broken