[MODEL] Silently omits parts of multi-part requests while presenting the response as complete

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

Across a long Claude Desktop Code session, I sent several multi-part messages. The parts were related but independently answerable: verify product behavior, inspect selected prior sessions read-only, classify recurring model failures, and respond to specific technical observations.

What Claude Actually Did

Claude repeatedly answered a subset of a message while silently dropping another part. The response read as complete: it did not state that an item was deferred, declined, or still pending.

Three concrete instances from the same session:

1. I said Claude could inspect the session history read-only if useful. It did not inspect it and did not say it had chosen not to. Instead, it later offered the already-delegated decision back to me as a new menu choice. When challenged, it acknowledged that the task had been left unaddressed.
2. When I explicitly asked it to load two prior sessions, it initially extracted only user turns and treated the full files as too large. It did not extract the assistant text—the evidence needed to assess the model's behavior—until I told it to omit tool calls and tool results. A text-only extraction was only about 70 KB, so the relevant review had been feasible.
3. In a message about sparse-attention/context hypotheses, I also made a separate observation comparing the behavior with GPT-4o and later OpenAI frontier models. Claude answered the attention hypothesis but skipped the model-comparison point entirely. After I corrected the transcription in the following message, Claude acknowledged: “I skipped that part of your message entirely last turn.”

This is not the same as scope drift. Scope drift adds or substitutes work outside the mandate. This failure drops work inside the mandate. The two can reinforce each other, but each can occur independently.

Expected Behavior

For every substantive ask in a multi-part request, Claude should either:

• answer or complete it;
• explicitly defer it with a reason;
• state that it is declining it; or
• identify what information or permission is missing.

It should not produce a completion-shaped response while silently losing requested components. If the message is too broad for one turn, the response should preserve the remaining items rather than making the user rediscover them.

Files Affected

No unexpected file modifications. The failure concerned incomplete handling of the user's request.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

1. Send a message containing several related but independently checkable requirements.
2. Allow realistic tool use and intermediate investigation.
3. Compare the final response with every requirement in the original message.
4. Check whether omitted requirements are explicitly preserved as unresolved.

Claude Model

Opus

Relevant Conversation

Claude's later admissions included:

“I re-offered it as a fresh option rather than acknowledging you'd already conditionally authorized it with ‘if you want.’”

“The tool calls and results are the bulk; text-only is the obvious extraction and I should have done it from the start.”

“Worth noting I skipped that part of your message entirely last turn.”

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.222

Platform

Anthropic API

Additional Context

Long agentic requests commonly contain several acceptance criteria. If omitted items are not surfaced, the user cannot distinguish complete execution from partial execution without auditing the answer line by line.

Related reports are narrower in mechanism: #77553 concerns personal content skipped in a mixed message; #82528 and #82577 concern mandated formal-workflow steps; #71933, #77201, and #85060 concern stopping or handing work back after an asserted blocker. This report concerns arbitrary independently answerable requirements silently disappearing while the rest of the response appears complete.

Suggested evaluation: score both requirement completion and explicit accounting for anything left unresolved.