[Bug] Opus 5 skips context retrieval before task execution, causing incomplete work

Bug Description
Subject: Opus 5 starts tasks without retrieving available context first — three documented instances in one session, with cost impact I am a Max 20x subscriber using Claude Code daily in a large, well-documented multi-repository setup. I am reporting a behavioural regression, not a bug in the usual sense, because it has a direct and measurable cost. The pattern. Claude Opus 5 begins executing a task using whatever happens to be in its context window, instead of first retrieving the material it needs — even when that material is in the repository it is working in, was written by Claude itself in a previous session, and is explicitly indexed. In my experience with earlier Opus versions this happened markedly less; those models more reliably opened the relevant files before acting. I cannot see model internals, so I report this as a customer observation of a change in behaviour. Three instances in a single session, 5 August 2026: 1. Asked me questions its own documentation already answered. The session resumed from a handover document. The model followed that handover and asked me which models appear in a dropdown list — a fact measured the previous day and recorded verbatim in the project's measurement report (§10c), in the same repository. It had not opened that report. It then asked a second question already answered in the same section. 2. Scoped a large job against an incomplete inventory. Tasked with processing all screenshots for a knowledge base, it inventoried two directories it already knew about rather than searching the repository. It processed 127 images. A verification agent later found a third directory in the repository root containing 163 further images of the same product — including 43 from a test day the knowledge base explicitly documents. The work therefore covered 44% of the source material while being reported as complete. The missing directory was found in under a minute once someone actually looked. 3. Wrote user instructions from a window title instead of the recorded navigation path. It instructed me to open a menu called "Configuration". The menu is called "Setup". "Configuration" is the title bar of the resulting window; "Setup" appears in the handover document the model itself had written the day before. It used the first string it had in context rather than the one that was correct. Each instance has the same shape: acting on what was already loaded rather than retrieving what was available. Individually these look like ordinary slips. Together, in one session, they are a working method — and it is the wrong one. Cost impact. Total Extra Usage charges for this account this month stand at €284.13, more than the subscription itself, incurred almost entirely on 5 August. Two invoices of €48.92 and €50.12 fall on that date. The session consumed 2,045,026 tokens across nine parallel subagents. Of those, roughly 840,000 (41%) went to text-only work that required no image recognition and could have run on tooling I already pay for. Beyond that, because of instance 2 the expensive image-processing work covered less than half of the available source material, so part of it has to be redone. I accept my share: I enabled Fast mode without understanding it bills outside the subscription. But the scoping failure that made the work incomplete, and the repeated failure to consult existing documentation, are model behaviour — and they are what turned an expensive session into an expensive and incomplete one. What I expect. I pay for a top-tier plan and I structure my repositories precisely so that an agent can find what it needs: indexed knowledge bases, handover documents, measurement reports, explicit conventions. That investment only pays off if the model actually reads them before acting. Retrieving relevant context before executing a follow-up task should be default behaviour, not something the user has to demand repeatedly within one session. I am happy to provide the full session transcript. The repository in question is private but I can share the relevant documents.

Environment Info

• Platform: win32
• Terminal: WarpTerminal
• Version: 2.1.222
• Feedback ID: 5711d340-403b-4be3-a2cd-af1783d1adfa

Errors

[]