ctxbudgeter v0.2.0 is live — a context engineering toolkit for AI agents

[Kayariyan28]

Body

Hi all,

`ctxbudgeter` is a framework-agnostic **context engineering toolkit** for agentic AI. Hand it the raw materials — system rules, docs, code, memory notes, tool results, the latest user request — and it compiles them into a clean, cheap, **auditable** prompt that fits your token budget, before the model sees it.

Think of it as:
- **Webpack** for agent context
- **pytest** for prompt/context quality
- A **token budget manager** that's deterministic and explainable

## Install

```bash
pip install ctxbudgeter

• Source: https://github.com/Kayariyan28/ctxbudgeter
• PyPI: https://pypi.org/project/ctxbudgeter/
• License: MIT

What it gives you

• Token-budget compiler with full inclusion/exclusion audit trail
• pytest plugin — assert_includes, assert_health_at_least, assert_no_secret_items, GoldenPack snapshots
• Cache-aware adapters — automatic cache_control for Anthropic, deterministic prompt_cache_key for OpenAI, plus LangChain & PydanticAI
• Sensitivity policy — allow | warn | refuse | redact for items tagged secret
• Memory store (Write strategy) + pack.fork() (Isolate strategy)
• JIT references — load context only if it could plausibly fit the budget
• Declarative pack.yaml + CLI — review prompt context in PRs, gate on health in CI
• Multi-modal attachments — images, structured tool schemas
• Async compile + retry-on-overshoot compressor
• Zero LLM calls in the core — local-first, deterministic, fast

Use any framework

LangGraph, CrewAI, OpenAI Agents SDK, PydanticAI, Microsoft Agent Framework, or your own loop. ctxbudgeter sits in front of them and does one thing: make the context you send cleaner, cheaper, and assertable.

A quick read

The [README](https://github.com/Kayariyan28/ctxbudgeter#readme) walks through a 7-day case study: hardening a Claude code-review agent from "context spaghetti" to production-ready. Tokens cut, cache hit rate 0% → 95%, cost per call down ~60%, one production incident class eliminated.

What I'd love to hear

• Use cases — what kind of agent did you wire it into?
• Bugs — first install reports are gold; please open issues
• Roadmap input — what would unblock you for v0.3? (vector store adapter, multi-model packs, OpenTelemetry hooks, web dashboard…)
• Contributions — issues and PRs welcome

Thanks for being here. Ask anything.

— Karan Chandra Dey ([K28 Design Lab](https://k28art.space))

Click **Comment** (or whatever the GitHub button says — "Start discussion") and you're done. This post becomes the pinned-by-default launch announcement on the Discussions tab.