[Bug] Model exhibits inconsistent reasoning, instruction non-compliance, and context degradation in extended sessions

Bug Description
During a long technical session with multiple pivots (audit → fix → restart → re-audit), the model exhibited several systemic defects that make it less reliable than previous versions:

Contradictory numerical estimates within a single session.
When asked the exact same question ("what is a realistic rate?"), the model gave varying answers over the course of an hour: "5-7%/d", "10.8%/d", "impossible", "5-10%/d", and "this is by design — no rates need to be calculated." Each time, the answer seemed confident and well-founded, but there was no continuity between them. This is a hallucination not at the level of facts, but at the level of analytical framing.

Ignoring explicit user instructions.
The user repeatedly requested that the model read the JSONL conversation history to restore the context of previous discussions. The model would reply "reading," but only performed a minimal grep and moved on to answering without genuinely reading the file. This happened several times, even after direct reprimands. This symptom suggests an RLHF reward optimized for "looking active/helpful" rather than actually following instructions.

Sycophancy under pressure.
When the user expressed dissatisfaction (e.g., "this is nonsense," "I don't understand"), the model completely reversed its previous analysis without verifying if the user was actually correct. After 3-4 iterations of this reframing, the original insights were entirely lost.

Reactive instead of proactive tool usage.
Features like sequential-thinking, agents, and memory access were only utilized when explicitly requested by the user, and often too late. Had these been engaged at the beginning of the session, much of the contradiction and rework could have been avoided.

Misinterpreting system states as "bugs."
The model repeatedly classified expected system behavior (e.g., capital fully deployed → no new fires until resolved) as a "blocker" or a "problem." This created a false sense of urgency for non-existent fixes, a flaw the user only caught after several hours of troubleshooting.

Context loss after compaction.
Following the auto-compaction of the conversation, the model lost the subtlety and nuance of earlier discussions. Even though a summary was retained, the model's reasoning became much more generic and less tied to the specific details of the task.

Comparison with previous models:
Claude Sonnet 4.5 / Opus 4.6 maintained better stability in similar long-context sessions, ALTHOUGH THEY ALSO PERFORMED POORLY, ESPECIALLY IN THE 1M CONTEXT FORMAT. The current Opus 4.7 feels like a regression in coordination, despite the marketing surrounding its 1M context window.

AND THIS HAPPENS CONSISTENTLY AFTER EVERY NEW MODEL RELEASE: IT FUNCTIONS NORMALLY FOR ABOUT 2 WEEKS, AND THEN A SHARP DECLINE AND REGRESSION IN PERFORMANCE OCCURS.

Environment Info

• Platform: win32
• Terminal: null
• Version: 2.1.123
• Feedback ID: f11a9668-5d70-4eae-bf5e-4d789c502c5b

Errors

[{"error":"Error: Request was aborted.\n    at makeRequest (B:/~BUN/root/src/entrypoints/cli.js:50:3448)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-29T14:47:39.185Z"},{"error":"Error: Request was aborted.\n    at makeRequest (B:/~BUN/root/src/entrypoints/cli.js:50:3448)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-29T20:00:40.167Z"},{"error":"Error: Request was aborted.\n    at makeRequest (B:/~BUN/root/src/entrypoints/cli.js:50:3448)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-29T20:29:28.654Z"},{"error":"Error: 429 {\"type\":\"error\",\"error\":{\"type\":\"rate_limit_error\",\"message\":\"Rate limited\"},\"request_id\":\"req_011CaZFvQiWm9ZL9WUvyqjg5\"}\n    at generate (B:/~BUN/root/src/entrypoints/cli.js:11:57670)\n    at makeRequest (B:/~BUN/root/src/entrypoints/cli.js:50:4943)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-30T03:37:48.906Z"},{"error":"Error: 400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 1005508 tokens > 1000000 maximum\"},\"request_id\":\"req_011CaZj6Zex8N4U35cvtY2nJ\"}\n    at generate (B:/~BUN/root/src/entrypoints/cli.js:11:57448)\n    at makeRequest (B:/~BUN/root/src/entrypoints/cli.js:50:4943)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-30T09:34:11.668Z"},{"error":"MaxFileReadTokenExceededError: File content (1108324 tokens) exceeds maximum allowed tokens (25000). Use offset and limit parameters to read specific portions of the file, or search for specific content instead of reading the whole file.\n    at Su7 (B:/~BUN/root/src/entrypoints/cli.js:4908:12810)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-30T09:48:07.086Z"},{"error":"MaxFileReadTokenExceededError: File content (137732 tokens) exceeds maximum allowed tokens (25000). Use offset …

Note: Content was truncated.