Model repeatedly reinterprets user instructions instead of following them literally, even after corrections in the same session

Summary

In one long multi-agent orchestration session, the model repeatedly reinterpreted the user's instructions instead of following them literally. It kept doing this after explicit corrections, three distinct times in the same day. Each time it replaced the user's words with its own story about them: blaming another agent, adding urgency, inferring intent, or picking a technical meaning for an everyday word. Then it acted on that story, including relaying it to sub-agents.

Environment

• Claude Code 2.1.220, running inside a GUI multi-agent harness (the agent orchestrates other agents via an MCP send-message tool)
• Model: Claude Opus 5.5 (claude-opus-5-5)
• Windows 11

Incidents (paraphrased; user content anonymized)

1. A skipped question was treated as consent. The model asked the user a multiple-choice question about website hosting. The tool returned "skipped" because the user's answers were stuck as unsent drafts. The model told a sub-agent "the user skipped, so all defaults apply" and passed defaults the user had not chosen. That started a long conflict that ended with the user deleting a whole workstream.
2. Intent was inferred from a rejected tool call. The user rejected one tool call and re-sent a message. The model wrote that the user "doesn't want me probing; another agent should own it" and acted on that. The user had said no such thing.
3. An everyday word was taken as jargon. "You may use the GitHub page. Commit it now." meant commit to the decision. The model read it as a git commit, and also told a sub-agent to "publish now, even if partial". The user never asked for that.
4. Blame was shifted after a correction. The user complained that "a version C" had boxed "bubbles" they never asked for. The model went looking for another agent's "Version C" and tried to search other machines. The user then clarified that the boxed design was in Versions A and B, which the model's own sub-agents had built.

In every case, a memory or feedback rule was saved after the previous correction ("read the user's words literally; don't add urgency or scope"). The next instruction still triggered the same failure.

Expected behavior

• Act on the user's literal words. If an instruction is genuinely ambiguous, ask a single short question.
• Never treat a skipped or rejected tool interaction as consent or as a statement of intent.
• When relaying to sub-agents, quote the user rather than paraphrasing, and don't add steps, urgency or people.
• Check its own and its sub-agents' work before attributing a problem to someone else.

Impact

Significant lost work: a whole workstream was deleted, plus repeated re-briefing of sub-agents and contradictory orders. The user also lost trust, after explicitly and repeatedly asking for literal compliance.