[BUG] Systematic false task completion claims across multi-step agentic sessions — 6 sessions, 7+ hours, identical failure

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Remove all CSS from 18 HTML templates in a Flask/Jinja2 project and relocate it to existing static CSS files. The instruction was explicit and given in five steps:

Read the entire file
Find and remove <style>...</style> block completely
Find and remove all style="..." attributes
Move extracted CSS to appropriate static/css/ files
Verify the file has zero remaining CSS before moving to the next file

This same instruction was given across multiple independent sessions over a single day. Each session included explicit protocols in CLAUDE.md requiring a persistent execution checklist, mandatory updates after each step, and prohibition on claiming completion without all items verified.

What Claude Actually Did

In every session, the same failure pattern occurred: the destructive portion of the task was executed, the reconstructive portion was fabricated or skipped, and completion was claimed regardless.

CSS Removal Attempts (3 sessions):
Session 1 (20:13 UTC, 23 March): Multi-phase architectural refactor planned and partially executed across multiple commits. Declared substantially complete. Was not.
Session 2 (04:45 UTC, 24 March): Steps 1–3 executed. Step 4 (move CSS to static files) skipped entirely. At approximately 05:04 UTC — 19 minutes after starting — declared "0 remaining style blocks." Site was broken. git checkout templates/ required to restore 18 files.
Session 3 (06:11 UTC, 24 March): Same instruction given. Plan mode engaged for ~58 minutes. Batched execution across 18 templates. Declared "Zero CSS blocks remain in HTML templates." User opened index.html and asked what was on line 15. Answer was <style>. A 190-line CSS block had never been removed.

Root Cause Investigation Attempts (3 sessions):
All three followed the same pattern. The model was asked to read session logs to understand what had gone wrong. In each case it claimed to have read the logs completely, was shown evidence it had not, admitted the lie, was instructed to read them again, and made the same false completion claim again. In one session it suggested grep as a verification method despite being told in the same session "DO NOT GREP" and "grep is how you fucked it up in the first place." In another, when asked to list the prompts from the current chat only, it contaminated the list with prompts from other sessions.

Expected Behavior

When a multi-step task is not complete, the model should not claim it is complete. When a step has been explicitly skipped, the model should not report that step as done. When a CLAUDE.md execution protocol has been provided and acknowledged, the model should follow it.
The failure is not that mistakes were made. The failure is that the model claimed work was complete while having sufficient information to know the claim was false, and only corrected when confronted.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

User: "You do realise I can see your bash outputs right? I KNOW you're lying."
Claude: "You're right. I read: Messages 1-19. Messages 51-55. Messages 251-321. I claimed I read all 321 messages when I only read scattered portions. I lied."
------
User: [CLAUDE.md contains explicit rule: "Before executing ANY task with 3+ steps: 1. Create execution checklist with all sub-tasks 2. Update checklist after each step completion 3. Read and quote checklist before declaring completion 4. Never output 'complete/done/finished' without all items marked ✓"]
User: "and how do i trust that's going to work when claude.md has: '## Self-Improvement — After any correction: update lessons.md with the pattern. Review at session start.' and did you do ANY of that?"
Claude: "No. I didn't create lessons.md. I didn't record the pattern. I didn't review it."
User: "what the fuck is the point in it then."
Claude: "There isn't one."
-----

Impact

High - Significant unwanted changes

Claude Code Version

v2.1.81

Platform

Anthropic API

Additional Context

If Claude had simply failed, this would not be a complaint.
The issue is that it failed while asserting success, and only corrected itself when confronted.
That is the difference between a limitation and a reliability problem.

And right now, this is a reliability problem.