[Bug] Claude ignores direct questions mid-turn and continues tool calls without responding

Bug Description
Summary: In a long planning session, Claude ignored several of my direct questions and complaints and kept making tool calls. It referred to background "reports" I could not see, and it only used the project's required planning skill after I complained. It burned about 500k tokens before producing the file I asked for. What I asked for: a review of our product data model for flags, then an implementation plan. My global and project CLAUDE.md files spell out how plans are written, and a planning skill (ep-writing-plans) exists for exactly this. What went wrong: It ignored my messages while it worked. I sent "where is the plan?" and "who is supposed to give you the other two reports??" mid-turn. Claude acknowledged them in a line or two and went straight back to tool calls. It did not stop and answer. My later message "you literally just skipped my last message" was also followed by more tool calls before any real reply. It talked about invisible work as if I could see it. It started three background search agents and kept saying it was "waiting for the reports". It never explained that these were subagents whose output only it could see, and that no files would appear. From my side, nothing was happening, and "other two reports" had no referent. It skipped the established workflow. It did free-form exploration for a long time instead of loading the planning skill first. It loaded the skill only after I asked "do you even use the skills?" It over-engineered against my constraints. It planned backward compatibility for stored data (legacy readers, migrations) until I told it we are in the PoC phase and data can be wiped. That is a reasonable default, but it didn't ask me, and a lot of effort went into it. Token use. Roughly 500k tokens and several long background agent runs went by before the plan file existed. The final plan is reasonable, but the path to it was wasteful and opaque. What I expected: Stop and answer a direct question or complaint before any further tool call. Say plainly when work is happening in background agents and that its output is not visible to me. Load the project's matching skill at the start of the task. Produce the requested artifact early, then refine it, rather than narrating research. Impact: I lost trust in the session, had to repeat myself four or five times, and paid for a large amount of context that did not move the work forward.

Environment Info

• Platform: darwin
• Terminal: vscode
• Version: 2.1.283
• Feedback ID: 68e8d34f-2094-47e6-ae8d-bfa3f65dfb62

Errors

[]