Model loses context and misreads questions in long sessions

Description

In long working sessions (this one was ~4 hours, many tool calls), the model begins misreading straightforward questions. Specific example:

User asked: "do you have access to the IG account?"

Context was clear — we had just discussed using Mima's Instagram photos (@mimasdelicafe) as source material for poster generation. The natural reading is: can we read that account's content?

Model instead checked posting credentials for @dudes.tribe and gave a detailed answer about write access — completely missing the question. Required the user to repeat themselves with frustration before getting a direct answer.

Pattern

• Later in long sessions the model answers the question it expects rather than the one asked
• Stops tracking conversational context accurately
• User has to repeat or rephrase things they already said clearly

Expected behavior

Short, direct answer to the actual question asked, even late in a long session.

Environment

• Claude Sonnet 4.6 via Claude Code VSCode extension
• Long session with many tool calls (SSH, file edits, DB queries, deploys)