[Bug] Comprehension/attention regression on claude-fable-5 since ~Jul 25; client prompts unchanged

Bug Description
Since ~2026-07-25, Claude Code's models show a marked comprehension/attention regression for me on both claude-fable-5 and the opus slot: the model no longer fully understands my instructions or the current state and direction of my project, and its attention won't stay anchored on them — it acts on a shallower or stale picture of the project, so outputs drift from what I asked. Same repos, prompts, and workflow as before the onset; not a fallback/routing issue.

To rule out client-side causes, I captured the actual /v1/messages request bodies from three Claude Code versions on the same machine (local dump server via ANTHROPIC_BASE_URL override, identical headless invocation, --model claude-fable-5): v2.1.200 (~early July), v2.1.218 (Jul 22), and v2.1.233 (current). Fable 5's system prompt is essentially identical across all three — 10805 / 11222 / 11220 bytes, same section structure, minor wording edits only. The client-side prompt for Fable 5 was unchanged across the whole window, so this regression cannot come from a client prompt change. On the opus slot, v2.1.219 (Jul 24) switched the default model to claude-opus-5, whose per-model prompt bundle includes the undocumented heron_brook section ("Do not call the AgentTool unless the user requested it…") — present in Opus 5's prompt on v2.1.233, absent for Fable 5 and Opus 4.8 (issue #80988).

Request: please investigate server-side changes affecting claude-fable-5 around Jul 20–25 — model serving, server-delivered dynamic prompt sections / experiment flags, or classifier behavior — since the client side is ruled out for Fable 5. Similar user reports: #83063, #81038, #80873.

Environment Info

• Platform: linux
• Terminal: konsole
• Version: 2.1.233
• Feedback ID: a6fe2bbc-ed71-4ced-b640-b7614ff81946

Errors

[]