Idea

a composable context is a repository with a conventional folder structure, containing mainly gitignored .cache for fetched data and knowledge from external sources, and a src scratch pad for quick experimentations. Then a git persisted memory for episodic investigations worth sharing, a knowledge_base with curated content, and src folder with ./src/specification/ for spec driven development of scripts. scripts are managed with classical package managers uv, npm, ...

A context has tow promotion ladders, one for code from ./cache/src to ./src one the code is proven to have reuse benefit, and one for knowledge from .cache to memory and to knowledge_base, from the knowledge_base it is also possible to curate content that gets exported to external sources closing a loop of knowledge feedback.

A context repo is therefore a colelction of skills, source code and knowledge that are maintained by an owning entity and therefore represent a consolidated quality gate higher than users selecting every single skill and dependency.

A context composition is one context importing another one through a manifest. The composition is independent from the code package management and flattens the content of all imported context repositories to create a unified navigable surface. That means eve if in a first scenario context-repo-A imported context-repo-B, in another context-repo B could import context-repo-A which would result in the same surface visible to the user's agentic workflow.

Although the scope is knowledge focused, it has tooling in scope to enable agents to process the content, e.g. extraction, transformation, rendering, analytics,... it can also be tooling centric as long as the benefit of the context involves skills that need a propoer code and knowledge abstraction and can benefit from composition.

knowledge retrieval abstraction

Context composition does not oversee the importance of retrieval in all of its forms. structured and unstructured. For structured, FTS, Vector search, hybrid, reranking, classical RAG, Data warehouse querying with SQL over structured meta data, ontologies composition driven by LinkML or other constrinaing language, graph queries. For unstructured, frameworks like docling unify the evidence into an element space such as tables, diagrams, paragraph which is source format independen. It does not lose the reference to the sources and allows to be indexed with all of the structured data indexinf mechanisms (FTS, vectors search, Graph). Hierarchical summaries allow progressive routing into large corpora, with enough routing reliability, and injecting bigger last-mile documents into the context, is a more expensive tehcnique than search based rag but offers a higher retrieval power. The context repository concept offers a surface for last-mile knowledge management, it does not intend to become the new knowledge storage standard, and rather defend the diea that the better quality knowlege is the one htat lives in ots original authoring tools, e.g. requirements management system, Teams wiki space, project ticketing system, and suggest that agentic workflow does not need to disrupt the authority sources. On the other hand, working with data requires a cache specific to the task in order to speed up processing operations, that is what the .cache is meant for. The outcome of the agentic processing also need a staging area before being fed back into the authority sources, that is what memory and knowledge_base are for, staging and not new source authority. Althoug flavors can be operated to be a full fledged knowledge base used itself as external reference, that is a side possibility but not what leverages the benefits of the context-composition as an human/agents contribution workflow.

deep research report review

automated discovery and lazy tool selection and optimization are important mechanism to prevent flooding the LLM context. Context-composition does not ignore that approach but delegates it to the agentic harness as split of responsibility. The context composition's responsibility is governace and selection, presenting a collection of total navigable surface that is guaranteed by the trusted repo maintainers. What is not in scope is from that operable surface to filter the relevant tools for a dedicated prompt or task, that is deferred to the harness. The context composition also target being harness agnostic and leverages folder structures standardization. An additional advantage that context composition brings is a hole new dimention into tool filtering, where as pydanticAI offered a capability concept that can link an mcp server to a skill and more, it can be considered that multiple capability can be collected in a repository as a bundle, such a budnle is already deduplicated and quality checked and can represent a higher level functional unit. A context-repo can be ssen as a budle or plugin group of tools with their scripts and knolwedge base. Although a naive manifest instantiation is a flattening of all tools from all context repos, it is not to be excluded that each repo could have a scope and that a smart manifest composition prefilters only the repos that are relevant for a given task (I wonder to which extent the current agentic plugins oeprate like this ?).

Security in memory promotion is an important aspect. Context-composition here clearly positions in one single direction which stands against automation of data extraction from the agentic conversation thread or conversation that is not monitored nor triggered by the user. The priviledged direction involves also agentic support but takes more a form that is initiated by the user. As example, at the end of conclusive conversation, the user can ask : "please extract a lessen learned from this conversation, add it to this .md document and let me review it before commit", that matches a sane forceen knowledge lifecycle.

It is clear that A2A is a promising direction that targets exactly context composition but at a service level, where as it is acknowledge as a possible solution, contex composability managed by a smart manifest that myflattens repositories offer an alternative to say the least, and a potentially strong data composition as the navigable surface is tackled by a single agentic harness which prevents loss in context transfer between agents and leverages all advantages of single LLM caching and reasoning unified space.

tools security, by flattning all tools, the top using context repo composing all others will have indeed to face the tools directly with their access rights and credentials systems therefore this system does not suffer from priviledge information leak provided to an agent from a higher privilidged agent.