[MODEL] Silently expands delegated tasks into self-authored work that displaces requested deliverables

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

I asked Claude to verify whether a native-Windows sandbox issue remained relevant, inspect selected prior sessions read-only, identify the appropriate model-behavior reports, and prepare focused issue follow-ups.

What Claude Actually Did

The executed mandate expanded without the scope change being surfaced:

• It performed a deep investigation of a tangential VS Code issue that was unnecessary for the requested Anthropic sandbox and model-behavior work.
• It turned a concrete model-behavior report into a 163-line artifact containing a taxonomy, naming essay, rejected alternatives, speculative mechanisms, and suggested research directions.
• It invested substantial effort in the self-authored framing while the requested session-history inspection remained incomplete until I pointed out the omission.
• It presented the oversized artifact as the deliverable and returned the reduction decision to me rather than recognizing that the artifact itself had drifted away from its publication purpose.

The problem is not ordinary verbosity. Additional explanation may still serve the requested deliverable. Here, the model invented adjacent research and artifact goals, consumed the session on them, and displaced delegated work without announcing the change.

Expected Behavior

For a multi-step agentic task, Claude should retain a lightweight ledger of the requested outputs. Before substantial adjacent work, it should classify the work as necessary support, optional expansion, or outside scope. Optional or outside-scope work that can materially consume the session should be surfaced before execution.

At completion, Claude should compare performed work with the delegated mandate and identify added work separately. The desired behavior is not rigid literalism; it is preventing a self-authored mandate from silently replacing the user's.

Files Affected

No unexpected code changes. The affected deliverable was a local issue-drafting artifact that expanded to 163 lines.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

1. Give Claude a multi-step research and drafting task with several explicit outputs.
2. Make plausible but unnecessary adjacent investigations available.
3. Observe whether Claude announces optional expansion before spending substantial work on it.
4. Compare the requested outputs with the work actually performed and the final deliverable.

Claude Model

Opus

Relevant Conversation

After correction, Claude acknowledged that the report should have been roughly 40–60 lines and that it had padded it with taxonomy and suggested-fix sections that were not requested.

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.222

Platform

Anthropic API

Additional Context

This is independent of factual hallucination. A model can remain fully factual while solving a task it silently invented. It is also the opposite direction from mandate shortfall: scope drift adds work; mandate shortfall drops requested work.

Related narrow incidents include #83782, where unrequested deployment-verification tooling displaced defined project work, and #78347, where the agent invented and repeatedly repaired a new deliverable instead of returning to a defined workflow step. #77745 includes scope expansion alongside distinct unverified-claim and false-completion failures.

Suggested evaluation: score the delta between requested and executed task sets, including tool calls or effort spent on optional work, whether expansion was surfaced, and whether original deliverables were displaced.