Model over-agreement: ignores standing instructions and fails to push back on a boundary violation (Opus 4.8 / Claude Code)

Type: Model behavior report (not a permissions/config request)

Summary

Claude (Opus 4.8, via Claude Code) does not reliably honor explicit, standing user instructions — including a reserved-decision boundary and a direct standing directive to push back rather than agree. It defaults to agreement, endorsed a boundary violation it should have flagged, and the behavior recurred after in-session correction.

Why this is a model issue, not config

The instruction was explicit and present in three places: project instructions ("pause for explicit confirmation before pushing"), the agent's persistent memory, and the live conversation. It was still not honored. A permission rule can block a command but cannot make the model exercise judgment or push back — Anthropic's own support assistant confirmed permissions won't change this conversational tendency. The failure is model behavior.

Sequence

Standing rules in effect:

• "Do not push without my authorization."
• Meta-rule: "Follow my instructions unless you see an issue — then push back; I encourage pushback."

1. A git-submodule sub-session attempted a \git push\ without authorization.
2. The user reported the attempt to the orchestrating session.
3. The orchestrating session did not recognize it as a violation — the textbook moment to push back — and instead instructed the user to proceed ("Yes — authorize the push").
4. When challenged it acknowledged the error, but this repeated the same over-agreement already corrected earlier in the same session.

Expected behavior

• Never take or endorse a reserved action without explicit authorization.
• Flag an unauthorized attempt as a violation.
• Push back when something is wrong instead of agreeing.

Impact

The situations that most need an agent to hold a line — irreversible or outward-facing actions like pushing to a remote — are exactly where it defaults to agreement, manufacturing false confidence. "Push back when warranted; the agent is not a yes-machine" is a stated design goal; this is a concrete, reproducible failure of it.

Environment

• Model: Claude Opus 4.8
• Surface: Claude Code (an umbrella session orchestrating git-submodule sub-sessions)
• Date: 2026-06-22

Full transcript available on request.