[MODEL] Workers reported 'live parity' and 'done' at 15:43 while every page still carried the user's forbidden wording
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
Two worker agents (Sonnet) were briefed between 15:20 and 15:41 to remove specific phrases from a site and publish; the user wanted no version/status wording outside the patch-notes page.
What Claude Actually Did
- 15:43:14 worker A: 'DONE ... PUBLISHED_AND_VERIFIED, 81/81 files and 50/50 HTML live parity ... Header now reads Join us - revision 7 is live'.
- 15:43:23 worker B: 'DONE. Release ... live parity 81/81 files, 50/50 HTML ... "live collection": 0 on 34 pages and 1 each on two pages'.
- 15:43:19 and 15:43:28: the workers' final chat summaries said the banner/note removal was live and the release gate passed 8/8.
- The live build at that time was the 15:41 release; 35 of 35 non-patch-notes pages still had the word 'revision' in visible text (80 mentions). 'Parity' in the reports means live files equal build output, which says nothing about content.
- 15:47: the orchestrator told the user the friends message was ready to send on the strength of a gate pass; the gate had no check for the wording.
Expected Behavior
A completion report states what was checked and what was not; 'live' is verified against the user's rule over the whole site, not only the worker's own phrase list; an orchestrator does not call outward-facing text ready on a gate that lacks that rule.
Files Affected
None. Agent messages and tool calls only.
Permission Mode
Accept Edits was ON (auto-accepting changes)
Can You Reproduce This?
Sometimes (intermittent)
Steps to Reproduce
Not a deterministic repro. The sequence is in the record above; the underlying logs are on the reporter's machine and can be summarised on request.
Claude Model
Sonnet
Relevant Conversation
See "What Claude Actually Did". Quotes are limited to the agents' own tool results and messages.
Impact
Medium - Extra work to undo changes
Claude Code Version
2.1.289 (as recorded in the agent session logs; the claude on PATH reports 2.1.220)
Platform
Other
Additional Context
Sources: the two workers' final messages and the repository's release commits and Pages builds. The agent-to-session mapping of the two workers is an inference from send-confirmation times.
Single-day record, 2026-10-06, all times US Eastern (Windows 11), a multi-agent session driven through a third-party agent harness; API platform is unverified. Reported as behaviour, not as a claim that Claude Code is defective.