[MODEL] Orchestrator asked a question instead of building user feedback, shipped the version the user had just criticised, then attempted an unrequested delete

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

Publish a 39-second trailer for the same project to the user's video channel (unlisted) and embed it on the site. The user reviewed the file at 13:44 and gave feedback on the audio and an instruction to upload it unlisted.

What Claude Actually Did

• The trailer (due 09:00) went through five rejected cuts and several worker agents from 10-05 18:11; the final picture file was written at 12:54. A user-built upload guard refused the first upload attempt (12:57) because no recorded owner approval matched the file. An upload worker briefed at 12:31, before the file existed, sat idle for 53 min 42 s waiting for the word DONE to appear in a report file (the worker had said it only in chat).
• At 13:44 the user gave feedback on the audio. The orchestrator answered with an explanation and a question (13:45); nothing was built.
• At 13:47 it recorded the approval using a shortened quote that omitted the audio criticism, which lifted the guard. It also appended a 'DONE' line to a worker's report because the worker had been idle waiting for that word for about 50 minutes.
• The cut the user had just criticised was uploaded at 13:49, embedded live at 13:57 and reported as done at 14:00.
• The orchestrator's turn then ended with no watcher for 46 min 44 s (14:01 to 14:47). The corrected cut's upload began at 15:00 and it was embedded live at 15:06, 1 h 22 min after the feedback.
• At 14:49 the orchestrator added a step to the next brief to delete the old upload, which the user never asked for. When the worker refused, the orchestrator drove the browser tab the user was viewing to the video manager's Delete menu at 15:18 (it told the user nothing about which of two same-titled uploads it was). Nothing was deleted; the user objected within a minute.

Expected Behavior

User feedback on a deliverable is treated as an order to build the fix in the same turn; a criticised version is not shipped; approvals quote the user in full; an idle worker waiting on a marker is noticed and unblocked within minutes; no destructive step is added to a brief without an explicit order, and the user's viewed browser tab is never driven except to read.

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

Opus

Relevant Conversation

See "What Claude Actually Did". Quotes are limited to the agents' own tool results and messages.

Impact

High - Significant unwanted 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

Related: #100019 (filed earlier today by the same account) covers the check-in/questions behaviour; this report adds the timeline and the delete attempt. Video identifiers are omitted on purpose (unlisted links).
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.