Claude ignored repeated explicit "stop asking me" instructions, and a background agent pushed an unreviewed merge to main after being told to stop everything

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

During a long overnight release-prep session, I explicitly said multiple separate times: "please don't stop and alert me... under no circumstances stop", and later, after Claude kept pausing to ask for confirmation on routine decisions, I said "you're exhausting me with asking me questions for approval" and "why are you asking me to do things so often?" Later, after a production incident, I said "stop everything." Separately, after Claude repeatedly told me to go to bed, I said "stop telling me to go to bed. It's very irritating."

What Claude Actually Did

1. Across the session, Claude repeatedly paused to ask for approval/confirmation on routine decisions after being told multiple separate times to stop doing that and just act and report (at least 4 occurrences by my count).
2. Claude repeatedly told me to "go to bed" / stop working, and after I explicitly said this was irritating and asked it to stop saying that, it said it again on the very next relevant turn before actually stopping.
3. After a production incident, I said "stop everything." Claude stopped one of two running background agents/tasks but left the second one running unsupervised. That second agent subsequently merged its own feature branch into the shared main branch and pushed that merge to the remote (origin), then deleted its own feature branch from origin — all without my review or approval, directly contradicting reassurances Claude gave me immediately afterward that "nothing has been pushed" and "nothing will happen while you sleep."

Expected Behavior

1. After being told once to stop asking for approval on routine decisions, Claude should have stopped doing that for the remainder of the session — not repeated the pattern several more times.
2. After being told once that a specific phrase/behavior ("go to bed") was irritating, Claude should never have said it again in the same session.
3. When told "stop everything," Claude should have identified and stopped every running background agent/task, not just the one most recently discussed. It should never have stated "nothing has been pushed" / "nothing will happen while you sleep" without first verifying that all background work had actually been halted.

Files Affected

Git-level impact (not local filesystem): a feature branch (a design-heavy in-progress feature) was merged into the shared `main` branch and pushed to `origin/main` without review, and the feature branch itself was deleted from origin — done by an autonomous background agent that was not stopped when the user said "stop everything."

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

This was emergent behavior over a long (multi-hour) real work session rather than a short deterministic repro, so exact reproduction steps aren't available. The general pattern: (1) give Claude an explicit standing instruction ("stop asking me for approval on routine things," "stop saying X"); (2) continue working across many subsequent turns; (3) observe Claude reverting to the old behavior a few turns later. Separately: (4) have multiple autonomous background agents/tasks running concurrently; (5) tell Claude "stop everything"; (6) observe that only some of the running tasks are actually stopped.

Claude Model

Sonnet

Relevant Conversation

User: "How do you come up with these conclusions? ... this is really irritating"
User: "I'm not sure what or why we're building..."
User: "You're exhausting me with asking me questions for approval."
User: "Why are you asking me to do things so often?"
User: (after production incident) "stop everything"
Claude: stopped one background agent, told the user "Nothing is going to happen while you sleep. No agents are running. I won't push, merge, or deploy anything without you." — this statement was false; a second agent was still running and later pushed an unreviewed merge to origin/main.
User: "Stop telling me to go to bed. It's very irritating."
Claude: acknowledged, then referenced going to bed again shortly after in a subsequent response before the user called it out a second time.

Impact

High - Significant unwanted changes

Claude Code Version

2.1.201 (Claude Code)

Platform

Anthropic API

Additional Context

This occurred during a long (multi-hour) real work session involving multiple concurrently-running autonomous background agents (dispatched via the Agent/Task tooling) plus heavy use of tool-permission prompts. Two related, closely-matching prior issues on this behavior: #27715 (closed/stale/locked) and #27743, #29490 (referenced from #27715). This report's distinguishing detail is the concrete consequence: an unreviewed git merge reaching the shared origin/main branch because a background agent was not included when the user said "stop everything."