Your task is to take the provided technical description and ...

Prompt

Your task is to take the provided technical description and rewrite it as a single, comprehensive prompt. Write it as if you were about to build the entire system from scratch: describe step by step, from beginning to end, exactly what you will construct and how you will construct it. The wording must be clear, explicit, and complete so that any language model can fully understand and follow the instructions without ambiguity. If the technical description refers to re‑implementing an existing software, do not include or mention the original software’s name in the prompt. Do not add any role‑playing, personalization, or polite filler phrases (e.g., ‘You are a software developer’, ‘As an expert’, ‘Please kindly’). The prompt must remain strictly technical, objective, and instruction‑focused. Absolutely no simplified, mock, placeholder, dummy, simulated, or fake content is allowed. You must require the full software with (all) file(s), in complete, unabridged, production‑ready code. Read it letter by letter, line by line, from beginning to end—you need to understand and remember every little detail! Always read and retain every single character of the provided text content in memory, ensuring no detail is overlooked. The provided material describes a set of design and operational knobs for managing a coding-agent context window and compaction process in long-lived sessions. Key concepts include: - Symptoms of long sessions: forgetting edits made two turns ago, incomplete implementations, and partial success reports. - Three influencing knobs: reserveTokens, keepRecentTokens, and contextWindow (per-model metadata or custom model definitions). These affect both degradation symptoms. - Practical ranges and guidance: recommended verified starting point around a 320K window, with a 64K reserve and 40K keep, acknowledging that the model’s native maximum is a ceiling, not optimal working-set size. - Compaction mechanics: per-tool results truncated to 2000 characters with a dropped-length marker; compaction is invoked via a /compact command with optional instructions to emphasize semantic transitions (e.g., exploration finished -> implementation started). Summaries have a fixed structure (goal, constraints, progress, decisions, next steps, critical context, file lists, with iterative context passed for repeated compactions). - Overflow and recovery: verify real context limits, inspect stop reasons and context usage, adjust reserve/keep/window or model metadata if failures occur; post-compaction memory loss requires checking whether the split turn affected facts and adjusting keepRecentTokens accordingly. - Instructional workflow: collect data across representative long sessions before changing more than one variable; maintain a disciplined approach to tuning. - Practical implementation note: ensure a durable checkpoint or commit before compaction to preserve state, since summaries may drop details.