Opus 4.6: Model ignores loaded context, requires constant user correction

Environment: Claude Code CLI, macOS, Opus 4.6, Max subscription ($90/month), alwaysThinkingEnabled: false

Summary: Opus 4.6 consistently fails to use context that is loaded into its own conversation. The model states wrong facts, agrees when corrected, but never self-corrects adjacent claims. This pattern makes it unreliable for professional use.

Reproduction (from today, March 13 2026):

A research task ("analyze Obsidian AI extensions") produced 10+ rounds of user corrections on basic factual errors:

1. Ignores infrastructure context. Project CONCEPT.md was loaded via hooks. It states: "systemA: OpenWebUI, systemB: llama-server." The model wrote "OpenWebUI on systemB" in the output, repeatedly, across multiple edits. The correct information was in its context window.

1. Recommends without understanding. Recommended a plugin, then couldn't explain what it does when asked. Had to WebSearch its own recommendation after the user questioned it.

1. No requirements gathering. Wrote a 500-line analysis document, then spent 6 iterations rewriting it after learning the user's actual needs. Requirements were gathered on iteration 4.

1. Passive agreement pattern. Every user correction was met with "you're right, my apologies" — but never triggered a review of other claims. 8 corrections in one session, zero self-initiated re-checks.

1. Marketing language as substance. Described tools using vendor marketing terms ("agent-mode architecture," "structured memory system") without understanding or explaining what they mean. When pressed, the model admitted it didn't know and had to search.

What I expect at this price point:

• Read the context that is loaded. Don't state facts that contradict your own context.
• When corrected, check whether adjacent claims are also wrong.
• Ask what the user needs before writing 500 lines of output.
• If you don't know what a tool does, say so — don't copy marketing language.

1. Does not follow persistent rules, no matter how they are delivered. I have explicit behavioral rules loaded into every session: "verify before stating facts about external software," "write drafts to the drafts folder," "read infrastructure docs before recommending." These rules are in CLAUDE.md (loaded at session start). They are in memory files (loaded by hooks). They are in project-level steering rules. The model reads them — I can confirm they are in context — and then violates them. Not occasionally. Every session. The same violations. For weeks. Despite adding more hooks, more rules, more verification gates, more memories. The model does not learn from corrections. It does not retain guidance across sessions. It does not adhere to simple, unambiguous instructions that are sitting in its own context window.

This is the core problem. If the model cannot reliably follow "check before you claim" and "read the file before you cite it," no amount of user-side infrastructure will fix it.

What I'm doing about it:

• Built 24 hooks, structured steering rules, forced verification gates, mandatory context loading, memory files with explicit behavioral guidance, automated performance tracking with rating capture
• Weekly average: 4/10, trending down

This is not an edge case or a prompting issue. The model has the information, has the rules, and does not use either.

To be honest, I am really, really pissed! This simply is unacceptable for the price point of a Max subscription!