Context Window Budgeter

The number-one silent killer of agent quality is context bloat: system prompts that accrete instructions for months, tool results pasted in full "just in case," entire conversation histories carried forever. The model does not complain — it just gets quietly worse: instructions buried at 60% depth get ignored, costs balloon, and nobody knows which tokens are earning their keep. This skill imposes a budget: every token category gets an allocation, measured against the window, with a compaction strategy for history and a retrieval-instead-of-stuffing policy for knowledge. Practitioners call context management "an underestimated production challenge" — this is the fix.

To give an agent system a real context budget: measure current usage by category, set allocations (instructions, tool schemas, conversation history, retrieved knowledge, working memory), define what happens at each threshold, and specify a compaction/summarization strategy that preserves what matters. The output is a budget document plus the measurement tooling to enforce it — not advice, an allocation model.

• When agent quality degrades on long runs and you suspect the model is "forgetting" instructions (it is not forgetting — they are buried).

• Before extending an agent's capabilities (new tools, longer histories, bigger prompts) — budget the new tokens first.

• When costs scale with conversation length and you need to know where the tokens actually go.

• When designing a multi-turn or long-running agent from scratch — budget on day one, not month six.

• For single-shot prompts with no history — there is no budget problem when the context is one exchange.

• When the model window is genuinely not the constraint (short tasks, small prompts) — do not budget what is not scarce.

• As a substitute for fixing bad instructions — if the system prompt is 4,000 words of contradictions, rewrite it; budgeting a mess just measures the mess.

• When the harness does not expose token counts — the skill needs measurement; without it, you get estimates (it will say so).