DR-009-Infra: AI Agent Context Packaging Tooling Selection

• Date: 2026-07-31

• Issue: #3115

AI Agent Context Packaging Tooling Selection		status: proposed
context: Infrastructure decision: APM is the primary AI agent context packaging tool; Lola is the fallback; OKIT is not recommended version: 1
dec_rec__infra__ai_packaging_tooling		Decision Record

Context / Problem

S-CORE contributors use multiple AI coding assistants (GitHub Copilot, Claude Code, Cursor, etc.). Agent context — skills, instructions, prompts, MCP server references — is today copy-pasted per repository and per tool, unversioned and unaudited.

A packaging layer is needed that:

• Distributes agent context declaratively and reproducibly across all contributors regardless of IDE.

• Provides supply-chain controls (lockfile integrity, SBOM, policy allow-lists) consistent with S-CORE’s existing sbom-tool and safety-oriented process.

• Integrates with CI as a merge gate (audit, drift detection).

Visual Overview

        flowchart LR
    subgraph Sources["Package Sources (git / marketplace)"]
        R1[eclipse-score/mcp-servers]
        R2[eclipse-score/tooling]
        R3[community skills repos]
    end

    subgraph PM["Package Manager (one of three)"]
        APM["APM\napm.yml + apm.lock.yaml\npolicy · SBOM · hashes"]
        Lola["Lola\n.lola-req\nno lockfile hashes"]
        OKIT["OKIT\nflat copy\nno versioning"]
    end

    subgraph Out["Per-Contributor Output"]
        C1[".github/copilot-instructions.md"]
        C2["CLAUDE.md"]
        C3[".cursor/rules"]
        C4["...other IDE configs"]
    end

    Sources --> APM
    Sources --> Lola
    Sources --> OKIT
    APM --> Out
    Lola --> Out
    OKIT --> Out
    

Options Considered

APM — microsoft/apm

Manifest + lockfile package manager for AI agent context (instructions, skills, prompts, hooks, plugins, MCP servers) across 8 coding assistants. Uses apm.yml manifest and apm.lock.yaml lockfile with content hashes.

Pros:

• Policy engine (apm-policy.yml): org-level allow-lists, tighten-only inheritance, CI audit gates.

• Lockfile with content hashes; apm.lock export --format cyclonedx|spdx SBOM export.

• Content-security scanning (hidden-Unicode / prompt-injection detection on install).

• Broadest agent coverage: 8 assistants (Copilot, Claude Code, Cursor, OpenCode, Codex, Gemini, Windsurf, Kiro).

• Drift detection; active Microsoft-org project (3.3k stars, 71 releases).

• Directly complements eclipse-score/mcp-servers and sbom-tool.

Cons:

• Newest governance/policy features are the least battle-tested part of the tool.

• Single-vendor stewardship (Microsoft) even though built on open standards.

• Adds a new onboarding step for contributors.

Lola — LobsterTrap/lola

Lighter, community-governed AI skill/context package manager (Go+Python) with marketplace-based distribution. Declarative .lola-req file (pip-requirements style).

Pros:

• Simpler mental model; lower adoption barrier.

• Vendor-neutral governance (GOVERNANCE.md, OpenSSF Best Practices badge).

• Covers 6 assistants (Claude Code, Copilot CLI, Copilot VS Code, Cursor, Gemini CLI, OpenCode).

Cons:

• No lockfile content-hash integrity, no policy enforcement, no SBOM, no content-security scanning.

• Narrower agent coverage than APM (6 vs. 8).

• Smaller community (109 stars, 8 releases); missing enterprise governance features.

OKIT — Mumme-IT/okit

Minimal, dependency-free (stdlib-only) Python CLI: clones a repo, copies skills/ and agents/ files into whichever tool directories are detected on PATH.

Pros:

• Zero dependencies; trivially auditable codebase.

• Very low learning curve (three commands).

Cons:

• No dependency resolution, versioning, lockfile, policy enforcement, or content-security scanning.

• Single contributor (1 star, 1 fork, 40 commits, no releases); bus factor of one.

• Not defensible as shared infrastructure for a multi-org Eclipse project.

Conclusion

Primary: APM. It is the only candidate with an actual governance and supply-chain-security model (policy enforcement, lockfile integrity hashes, SBOM export, drift detection, content-security scanning) — essential when agent context is executable-in-effect and S-CORE is a safety-relevant open-source project accepting third-party contributions.

Fallback: Lola. Kept on the radar as a lighter-weight, vendor-neutral alternative if APM’s Microsoft stewardship or scope becomes a concern for Eclipse Foundation governance.

Not recommended: OKIT. Pre-production maturity (bus factor 1, no governance). Revisit if the project matures significantly.

Tool	S-CORE Fit	Recommendation
APM	High	USE — primary packaging tool
Lola	Medium	WATCH — fallback only
OKIT	Low	DO NOT USE at this maturity

Note

Status remains proposed until a proof-of-concept pilot validates integration with S-CORE’s Bazel/Sphinx/sphinx-needs infrastructure and CI pipeline.