Treat the Context Window Like a Budget, Not a Junk Drawer
An open-source context-engineering framework has AI agents draft a context manifest before work begins, disciplining token budgets with decay rules and cross-platform parity checks.
Most AI-assisted coding workflows let the context window fill up by accident—open a few files, paste a stack trace, and before long the chat is bloated and the model starts dropping earlier instructions. The ai-assistant-dot-files project has grown into a full Context Engineering Framework that treats context as an explicit budget rather than a vague pile, defining one canonical set of 24 agents, 53 skills, and 13 inter-agent contracts in a shared/ layer, then projecting them onto six coding tools including Claude Code, Cursor, Windsurf, Copilot, Gemini/Antigravity, and Codex.
At its center sits a context-engineer agent whose only job is to produce a context-manifest.md before any real work begins. The manifest scopes relevant files, surfaces knowledge items and architecture decisions, and estimates token pressure for downstream agents—and it's a governed artifact itself, checked against a formal contract before the pipeline can proceed.
The framework also draws a sharp line between Context, Memory, and Learning: context is the current working set, memory is durable curated knowledge, and learning is the feedback loop that turns repeated delivery evidence into new rules or prompts. A 'context decay' rule keeps artifacts older than two pipeline phases from being loaded verbatim, replacing them with roughly 200-word summaries so later agents aren't stuck rereading stale reasoning.
Because not every AI tool has the same capabilities—Claude Code supports full agent orchestration while Windsurf and Copilot only get persona-style projections—the project maintains parity scripts to stop instructions from drifting across platforms. The practical takeaway for engineers, even without adopting a 24-agent system, is to explicitly separate context/memory/learning, add a pre-flight manifest step, summarize old artifacts by default, and treat the context window as a scarce design surface rather than just a bigger prompt.